Among the giants of the digital world, hyperscalers like Amazon, Google, Meta and Microsoft will collectively allocate more than $500 billion in capital expenditure to AI infrastructure this year.

When a handful of players invest upwards of half a trillion dollars in data centers, primarily to support AI, that tells us they see a ton of demand.

We can say with total certainty that they have every intention of recouping their investment (and then some).

Indeed, the voracious demand expected from enterprises for AI capability is what has made hyperscalers, frontier labs and GPU makers darlings of the stock market.

However, their anticipated revenues are the future cash flows of companies, including those in industrial sectors like manufacturing, energy and transportation.

The Cost Shift We’re Not Talking About Enough

Up until recently, much of the discussion around AI costs has focused on training large language models. The cost of training just one OpenAI model, GPT-4, for example, is said to have been north of $100 million.

That’s a big number. It’s also largely irrelevant to the industrial organizations that will never train a frontier model themselves.

Inference, the process of applying models in live operating environments to generate recommendations, support decisions and take actions, is a different matter, however. Yet it’s largely absent from most boardroom conversations.

Unlike training, which mostly represents a single upfront investment, inference tasks to support decisions and operational execution run continuously.

As companies shift from experimentation to large-scale deployments of AI across the value chain, inference will soar as a share of workloads and IT budgets.

Tokenomics Take Root

Most businesses do not understand the true cost of AI from a financial point of view. Industry needs a new economic calculus: “tokenomics.”

Tokens—the units through which generative AI systems process and generate information—are the standard mechanism to measure and charge for AI consumption. They’re like the kilowatt-hours for which electric utilities charge their customers.

An AI agent that performs tasks across large operational datasets consumes lots of compute (and storage, networking, electricity, water and more), and the total bill reflects that increased usage.

As companies move more and more to agentic systems, these costs will mount quickly. Agentic workflows require several model calls for each task, involve large context windows and need multiple reasoning loops. By their nature, they are many times more compute-intensive than the generative AI tools most enterprises have used so far.

A Familiar Pattern

Recent research suggests the AI inference market will more than double over the next five years.

Many of us witnessed a similar pattern earlier in our careers during the move to the cloud. In the 2010s when cloud computing was reaching scale, the same economic phenomenon was at play: unit costs dropped sharply while total usage skyrocketed.

The result was, contrary to many expectations, significantly increased total IT spending. There were many business benefits associated with cloud, but it didn’t mean that total costs declined.

The coffers of hyperscale cloud providers swelled. Those industrial enterprises that did not accurately project cloud consumption curves were caught off guard. Today, tokenomics is proceeding on a similar trajectory, only faster.

By now, CFOs understand well the economics of cloud computing and how to optimize the variable cost structures they represent. A similar level of budgetary scrutiny is appropriate for AI.

Token prices are declining as a result of intense rivalry and efficiency improvements. Open-source frameworks, custom silicon and innovative development methods are driving down model costs rapidly, especially in China.

However, as companies look to integrate AI into mainstream industrial operations, plant floors and logistics networks, the absolute number of tokens generated will far outstrip gains in per-unit costs. What looks like a bargain on a per-transaction basis today will look very different as industry marches towards greater autonomy.

A Pragmatic‍‌‍‍‌ AI Approach

Moreover, data-rich environments like industrial plants are sensitive to latency in ways that consumer and enterprise applications are not. Continuous inference over a network of sensors, industrial equipment and control systems requires a very different kind of computation than running a chatbot. The data volumes are different. The infrastructure requirements are different. And the economics are different.

This should not be taken as an argument against the deployment of AI in industrial operations. On the contrary, with a pragmatic mindset about AI’s costs, we can harness the exciting innovations AI introduces and enable all-new use cases.

The good news, though, is that tokenization does not have to mean runaway consumption. Where models are invoked, how often they run, what context they draw on and how intelligence is distributed across edge, cloud and hybrid environments all affect cost.

Importantly, AI in industry is not monolithic. Lighter-weight models, edge-based processing and other techniques will continue to improve the efficiency of AI.

Tried-and-true AI-based predictive maintenance, computer vision and process optimization solutions all provide clear illustrations of mature, winning business cases. Generative, agentic, physical and multimodal AI, however, herald something different and more complex financially.

The upside of AI in industry is significant. So are the potential costs.

They should be accounted for on the same spreadsheet.

Use the best business cloud storage to manage your data.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

With all the buzz around nation-state threats, it’s easy for organizations to focus on threats outside the business – and forget about risks that can spiral outwards from within.

Whilst GenAI tools have introduced undeniable efficiencies for employees, these platforms have also introduced a new class of risk: two in three UK organizations admit they can’t track whether employees are sharing data via approved tools.

Most of the time, employees aren’t sharing sensitive data because they have malicious intentions. They are uploading sensitive information – like contracts, client proposals or supplier agreements to models like ChatGPT and Claude to save time on routine tasks.

Almost all (93%) of CEOs across the globe have adopted generative AI to some extent in the past 12 months (PwC). What’s concerning is that much of this activity is happening without any oversight, in the browser – meaning organizations are failing to track the flow of company information, including when and where it’s uploaded.

This is spiraling into serious risk for businesses.

First, because employees may inadvertently share credentials or other access details with public LLMs, which could result in unauthorized access if the model is compromised.

Second, uploading personal data to LLMs can trigger compliance breaches with laws like GDPR and the Data Use and Access Act – resulting in costly fines as well as reputational damage.

To take control of this issue, leaders will need to implement tools and technologies that provide visibility and control over usage at both the browser and the application levels.

The incentive problem

Employees don’t need more mandatory cybersecurity training – the problem is incentive. Many company-owned gated LLMs are still in the pilot stage, falling short of the speed and precision offered by public alternatives.

While the majority of employees understand the risks, 35% of UK businesses admit data sharing through external tools takes place – indicating many would rather ‘throw caution to the wind’ than waste valuable time using slower tools.

But the risks of this behavior – particularly in highly regulated sectors like financial services, could mean unsanctioned LLMs become 'hidden icebergs’ in an organization. Concealed, but capable of causing catastrophic damage upon impact – like inadvertently exposing customer transaction histories or credit scores.

Part of curbing Shadow AI use in the enterprise therefore starts with designing approved AI tools that integrate easily with existing platforms (for example, Microsoft 365 and Google Workspace). These tools should be continuously improved based on user feedback, ideally avoiding excessive restrictions that make the tool frustrating to use.

But the fact is, nearly two-thirds of organizations are currently stuck in the pilot stage when it comes to their AI initiatives and haven’t started to scale across the enterprise (McKinsey). So, what can organizations do today to gain control of the Shadow AI problem?

The solution – tools to bring unsanctioned AI usage under control

You can’t control what you can’t see, which is why organizations need a real-time view of who, or what, is accessing what data, from which devices, and where it’s being shared.

Next-gen identity security platforms can help organizations to gain an immediate understanding of how employees interact with consumer AI tools like ChatGPT, Claude, and Gemini, tracking interaction frequency and monitoring document uploads.

Once high-risk behavior is identified, organizations can then automate corrective actions, redirect users to secure AI alternatives, or prompt users to justify their business use case before proceeding.

Visibility will become even more important with the emergence of ‘nested’ agents. In this scenario, employees might believe they’re only interacting with a single AI agent, but that chatbot may delegate tasks to multiple underlying agents. The organization has no visibility into how many downstream agents or services its information is being shared with.

An identity security tool makes these identities ‘discoverable’ via a real-time ‘agent ledger’. This ledger acts as a complete, unchangeable trace of all agent activities and interactions. It also applies controls to each agent in the database.

Only the absolute minimum privilege required for a task is granted, at the exact moment it is needed, and for the shortest possible duration. In this way, agent permissions don’t automatically ‘cascade’. If an agent wants to connect with another agent, it must be verified by the system first.

Closing the visibility gap

Shadow AI is more than a tooling problem: it’s an identity problem. Organizations can close the ‘visibility gap’ by using tools that track interaction frequency, block sensitive document uploads, and prompt employees as well as AI agents to justify their business case before they use unsanctioned tools.

Once organizations know which tools are being used, what data they're accessing, and where that information goes, they can apply effective guardrails to secure behaviors – both human and non-human. A simple inventory of AI agents is not enough; now, organizations need to move beyond flat inventories and develop an understanding of the context and relationships that surround every agent.

In essence, identity security platforms become adaptive – moving from static to dynamic, real-time approaches to access. This is helping organizations to operationalize zero trust by ensuring that no identity, human or non-human, is trusted by default.

In the era of AI agents, securing identity has become a prerequisite for innovation.

We've featured the best AI tool.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Prompt engineering was briefly the face of the AI jobs boom.

In 2025, technology recruitment firm SPG Resourcing reported that UK job listings for AI prompt engineers had grown by 180% over the previous year.

It was an eye-catching figure that captured the mood of the first commercial wave of generative AI. Businesses were trying to understand how to talk to LLMs and turn early experiments into something useful.

That requirement has not disappeared.

With job site postings for specialist AI roles in the UK rising by 61% from last year according to PWC, it’s clear that good prompts still matter. But most of that new demand is for people who can apply AI inside a business, not just talk to a model.

Many organizations have moved past the first demo. They are now trying to build AI agents, retrieval-augmented generation (RAG) systems, and AI-enabled workflows that operate inside the business.

That creates a different skills gap.

This is the shift behind what I refer to as ‘context engineering’. Although that term is not universally used, the capability is becoming essential.

Companies that want useful AI agents and retrieval-augmented generation systems need people who can design the environment around the model, not just the prompt sent to it.

From better prompts to better context

Prompt engineering is about the instruction, while context engineering is about the world around that instruction.

A support agent does not only need a well-written prompt, it needs the right customer record, policy, product history, and permission boundary. Similarly, a developer agent needs the relevant code, tests, dependencies, and deployment constraints.

In both cases, output quality depends on context. Without it, the model is guessing from incomplete evidence. With too much of it, the system becomes noisy and difficult to govern. The job is to make context useful, current, and controlled.

That’s what makes context engineering distinct from prompt engineering.

Agents raise the stakes

The rise of AI agents makes this more urgent. A chatbot with poor context may give a weak answer, but that same poor context may result in another agent making a serious error.

Once an AI system can call tools, query business systems, maintain state, and act across several steps, context becomes an essential part of the production architecture. It decides what the agent can see, what it can do, and how much confidence the business can place in the outcome.

I’m reminded of a joke about boundary testing. A developer walks into a bar and orders a beer, then he orders five beers, then he orders 999,999,999,999 beers, then he orders -1 beer. The bartender blinks, but everything is okay. A user walks into the bar, asks where the bathroom is and the whole bar explodes.

The same principle applies to AI projects. The first prototype may work against a narrow set of examples, then become fragile when it meets real data. Customer information sits in one system, operational data in another, and important knowledge in documents and files. The agent is expected to reason across all of it, but the context layer has not been designed for that job.

A financial services team, for example, may need to connect CRM data, an existing data platform, and internal documents before an agent can answer accurately. The hard work is not only moving the data. It is shaping it so the agent can retrieve the right evidence and stay inside the right permission boundary.

The job title is still catching up

This creates an awkward hiring moment. The need for context engineering is becoming clearer, but the job title is still unsettled.

Some organizations may seek to specifically hire ‘context engineers’, but many will not. The capability is more likely to appear inside roles such as AI engineer, agent engineer, AI platform engineer, applied AI engineer, or data engineer. In other businesses, it will be a team responsibility shared across data, platform, security, and software engineering.

Leaders therefore need to hire for the work, not the label. A candidate does not need to have ‘context engineer’ on their CV to be useful. The better signal is whether they understand how data moves through systems, how permissions are enforced, and how a prototype becomes something reliable enough for production.

This also means the talent pool is wider than many companies assume. Machine learning expertise is valuable, but context engineering draws heavily on existing engineering disciplines. Data engineers understand pipelines and retrieval.

Platform engineers understand operational resilience. Security teams understand access control and auditability. Software engineers understand how to turn messy requirements into maintainable systems.

The best candidates may look like full-stack AI engineers. They do not need to be specialists in every model, database, or framework, but do need enough range to connect the model layer with the business systems around it.

The Kubernetes lesson and what leaders should do now

The shift has a parallel with the move to cloud-native architecture and Kubernetes. Many companies treated Kubernetes as something to install, then discovered that the harder work was changing how teams built and ran software.

AI creates a similar risk. Companies can buy tools and hire a handful of specialists, but still fail to change the engineering habits around them. Context engineering requires teams to think differently about everything from documentation, and data ownership, to access, testing, and accountability.

It also changes the culture of software development. Engineers are already using AI to write, review, and iterate code. That can improve productivity, but it does not remove responsibility. In areas where performance, reliability, or security matter, human judgement becomes even more important.

CTOs and CIOs should not wait for context engineering to become a mature hiring category. They should start identifying the capability now.

The first step is to examine where AI projects are failing. Is the model genuinely weak, or is the system retrieving poor context? Are permissions clear? Can the team explain why the agent produced a particular answer?

The second step is to build cross-functional teams. AI cannot sit apart from data, platform, security, and product. In many cases, the best approach will be to upskill existing engineers who already understand the organization's systems.

The final step is cultural. Engineers need to become fluent in AI-assisted development while staying accountable for the systems they ship. Leaders need to make room for experimentation, but they also need clear standards for review, evaluation, and governance.

The model is not enough

AI hiring is changing because AI itself is moving into production. Models will continue to improve, and businesses will have many ways to access them. The harder advantage will come from knowing how to connect those models to the right business context.

Companies that understand this will build agents and RAG systems that are more useful, safer, and easier to govern. Companies that ignore it will keep blaming the model when the real weakness is the environment it has to work in.

We list the best resume builders, to make it simple and easy to build a CV to help your career.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

If your business celebrated the European Union’s (EU) delay of high-risk AI penalties, don't open the champagne just yet. On August 2, 2026, Article 50 of the EU AI Act goes live, bringing new transparency requirements for the AI tools businesses already rely on. While backend compliance demands have been delayed, chatbots, AI generators, and synthetic media systems are now in the spotlight.

The AI Act delays may have dominated headlines, but Article 50 never stopped moving. The European Commission has now finalized the transparency guidelines businesses need to follow, giving regulators a clear framework to enforce consumer disclosure rules starting this weekend.

Think the EU AI Act only targets big tech? Think again. Small or medium-sized enterprises (SMEs) serving European users are also in scope, wherever they operate. The age of invisible AI is ending, and businesses that hide automated interactions may find themselves under significant regulatory scrutiny before the business week even begins.

Who is caught in the net? Providers vs deployers

To navigate the EU AI Act without getting lost in legal jargon, businesses must first understand which role they fall under. The framework separates organizations into two categories: Providers and Deployers. Mixing up these definitions is the fastest way to accidentally skip a critical regulatory step or waste time trying to solve compliance issues that aren't actually your responsibility.

If your company builds AI models from scratch, significantly modifies open-source systems, or white-labels a third-party AI tool to sell under your own brand name, you are a Provider under the EU AI Act. That puts the technical burden on your business, from building transparency infrastructure to embedding machine-readable watermarks into AI-generated text, images, and videos.

For most SMEs, however, the second category applies: Deployers. If you are a small business that simply plugs an off-the-shelf AI widget into your website to handle customer service tickets, or uses AI tools to create marketing content, you are deploying AI. Thankfully, your responsibility isn't to invent deep-tech watermarking protocols but to double-check that people interacting with your systems are notified that they are dealing with an algorithm rather than a human employee.

The AI Act timeline: What was delayed vs what is due now

When the EU delayed major AI Act deadlines, businesses were quick to assume they had more time. The May 2026 "Digital Omnibus" amendment formally rolled back the compliance deadlines for complex AI architectures, leading many burned-out business owners to assume the entire rulebook had been kicked down the road.

Unfortunately, that was a serious misunderstanding of the changes. The extensions only apply to standalone, high-risk frameworks, such as automated resume screening, biometric identity verification, and AI-powered credit scoring tools. Because these deep enterprise tools require massive infrastructure overhauls and independent third-party audits, the EU simply gave businesses a longer runway until late 2027 and 2028 to prepare.

Basic consumer transparency, however, received no free pass. The European Commission made it clear that protecting everyday users from digital deception could not be delayed. As a result, Article 50 moves forward on schedule, separate from the postponed systemic deadlines. To guide businesses through this transition, the EU has established the Code of Practice on Transparency of AI-generated Content as the benchmark for compliance.

Article 50 and the four pillars of AI transparency

If your business utilizes AI to interact with customers or create content, Article 50 is now part of your operating reality. The EU has broken down its transparency rules into four core pillars, aiming to remove the mystery around automated systems without slowing down innovation.

For SMEs, the era of set-it-and-forget-it AI is firmly over. Depending on how your business uses generative tools, you are now legally required to update user interfaces, verify compliance from software vendors, or change how AI-generated content is published.

AI chatbot transparency rules

The first pillar is the “Human-AI Interaction” rule, and it targets the conversational bots handling your customer service, lead generation, or basic troubleshooting. Under the new guidelines, you can no longer trick a consumer into thinking they are chatting with a human employee when they are actually interacting with a large language model (LLM) script.

Starting this week, deployers must stop hiding AI behind a human-looking interface. Crucially, users need a clear notification that they are interacting with an AI system before the conversation starts, not a disclaimer buried in a wall of legal text. For SMEs using AI customer support, that means updating chat interfaces immediately to include an undeniable "I am an AI assistant" disclaimer.

Synthetic media watermarking

The second pillar shifts the technical burden onto the builders of generative AI tools. If your business creates or distributes software that generates synthetic text, audio, images, or video, you must embed machine-readable digital watermarks into the assets those tools produce.

The purpose of these markers is to help other software, like social media platforms or verification services, automatically detect that a piece of media was created by an AI. While existing providers have limited time to implement this infrastructure, any new generative AI tools launched after this week must include watermarking from day one.

Labeling AI-generated content

While watermarking happens behind the scenes, the third pillar requires clear labeling on the surface of the content itself. This rule specifically targets what the EU defines as deepfakes and synthetic output that could influence public understanding. If your marketing team uses AI to create realistic images, videos, or audio that could be mistaken for real people or events, you must slap a clear “AI-generated” warning label on them.

This responsibility falls on the businesses publishing the content. So, if you publish an AI-generated image of a realistic crowd for a promotional campaign, or use an AI-voiced avatar in a public video, it must also feature an overlay or disclaimer stating that it has been artificially generated or manipulated. The only exception here is obvious artistic, satirical, or fictional content, but transparency is still the safest bet.

Emotion recognition rules

The final, and perhaps most legally sensitive, pillar targets biometric and emotion-recognition systems. If your SME utilizes software to detect emotional states, interpret facial expressions, or classify individuals based on biometric data, you must proceed with extreme caution.

The new rules require businesses to explicitly inform users whenever they are exposed to emotion recognition or biometric profiling systems. For instance, if you use an AI-driven video interviewing tool to analyze a candidate's body language, or a retail system checking customer reactions, you must secure clear user consent before the software starts collecting this type of sensitive data. Psychological and physiological tracking can no longer happen behind closed doors.

The AI Act’s hidden deadline split

Like with any major regulatory rollout, the devil is always in the details. For SMEs, the EU AI Act brings a mix of breathing room and immediate pressure. Understanding which deadlines moved (and which did not) is the only way to avoid a costly compliance mistake.

The good news for software creators and IT departments is what experts call the grandfather clause. If your generative AI model was already legally available on the market before August 2, 2026, the EU provides a four-month technical runway, which moves mandatory machine-readable watermarking requirements to December 2, 2026. This extension gives development teams a much-needed time to update systems, test metadata workflows, and prepare for compliance.

However, relying too much on this buffer is where SMEs fall into a terrible, zero-day trap. The four-month grace period applies only to the technical back-end watermarking of existing AI tools, and it doesn’t cover chatbots, public disclosures, or newly launched software.

If you’re launching a brand new AI writing tool or chatbot after this week, there is no ramp-up window. Similarly, user disclaimers for deepfakes and interactive chatbots must be active immediately on day one. Without clear AI notices on customer-facing systems, your business becomes non-compliant the moment the deadline passes.

How to comply before Monday morning

If your business is staring down the barrel of the August 2nd deadline, the answer is not panic but preparation. Since there is no single EU registry for AI tools, compliance starts with something much simpler: making your AI use visible, documented, and transparent.

You do not need to rebuild your software stack over the weekend, but you do need to secure your customer-facing AI touchpoints. Here are four steps to take before Monday:

1. Map your stealth AI usage: Audit your teams for untracked software, plugins, and automated systems creating customer-facing content.
2. Kill the invisible chatbots: Update AI chat tools with a clear upfront notice telling users they are interacting with an AI assistant.
3. Verify software vendor compliance: Get written confirmation of compliance and ask how providers handle machine-readable watermarks.
4. Document your paper trail: Log every AI update, vendor conversation, and internal policy change for future reviews.

Turning compliance into trust

For many businesses, Article 50 feels like a new burden, but for smart SMEs, it is a chance to build trust. As users grow more skeptical of hidden AI, businesses that are transparent about how they use automation can stand out from the crowd.

Compliance is no longer another box to check to avoid penalties but a powerful indicator of integrity that can set your brand apart in the new age of AI.

A new acronym is starting to appear in boardrooms, investor conversations and executive debates: MANGOS - Meta, Anthropic, NVIDIA, Google, OpenAI and SpaceX.

It is being used as shorthand for the new power cluster of the AI era, and perhaps as the successor to FAANG as a symbol of where technological gravity now lives.

Here is my take: MANGOS is not the destination. It is the opening move.

Over the next five years, some companies that dominate their markets today, icons of an era, businesses that once seemed structurally unassailable, may lose relevance, decline, or be replaced by new leaders. At the same time, new companies will emerge from places nobody is currently watching.

This has happened before. It will happen again. What is different this time is the speed and the depth of structural change required to survive it.

The world is being reshaped by AI. But the companies that will win are not the ones that simply add AI to what they already do. They are the ones who rebuild themselves around it entirely.

The tool-adoption trap

There is a seductive but dangerous interpretation of AI transformation that goes something like this: deploy a copilot here, add a chatbot there, automate a few workflows, report efficiency gains to the board, and move on.

That approach may create efficiency gains, but it does not create an AI-native company.

The companies building genuine competitive advantage right now are not asking, “Where can we use AI?” They are asking a harder question: “If we were starting this company today, with AI as a native capability rather than a retrofit, what would we look like?”

The answer, almost always, is dramatically different from what they look like now.

This is not a technology shift. It is a company redesign. And the gap between those two ways of thinking is where many organizations are currently losing.

What AI-native actually means

The term AI-native is used loosely, so let me be precise about what I mean by an AI-native mindset. It has concrete, structural implications that go far beyond which tools a company deploys.

The first is re-engineering the org chart to unlock AI outcomes, not to preserve legacy structures.

Most organizational hierarchies were designed for a world of information scarcity, where value came from controlling access to data, expertise and decision-making authority. AI inverts that logic. Information is becoming abundant. The constraint is now judgment, creativity and the ability to act on signals faster than competitors.

An org chart built for the old world can suppress the outcomes the new world makes possible. Redesigning for AI is not about simply cutting headcount. It is about rethinking where human judgment actually adds value, and structuring the organization so that AI amplifies it rather than working around it.

The second is increasing talent density with genuine intentionality.

AI adoption does not happen organically at the pace this moment demands. When transformation requires a cultural shift as fundamental as this one, where every employee’s relationship with their work is changing, organic adoption is too slow.

It has to be a CEO mandate, driven from the top, resourced seriously and measured against real outcomes. The companies that treat AI training as an optional benefit will be outrun by the companies that treat it as a core operating requirement.

The third, and the one I find most underappreciated, is eliminating before automating.

There is a reflexive impulse in most organizations to reach for automation as the response to inefficiency. But the highest-value process is often the one that should no longer exist.

Before you make something faster, you need to ask whether it needs to happen at all. Automating a broken process at scale just produces broken outcomes faster. The discipline of elimination, of genuinely questioning whether a workflow, a layer of approval, a reporting structure or a product feature actually needs to exist, is harder than automation, and more valuable.

The fourth is relentless commercial focus.

AI investments must drive measurable revenue growth or profitability. This sounds obvious, but in practice, it is where many transformation programs lose their way. The excitement of the technology can become its own justification. It cannot be.

Every significant AI investment should have a clear line of sight to a commercial outcome, not eventually, not in theory, but in a timeframe that the business can actually hold itself accountable to.

In commerce, this distinction is already becoming visible. The winners will not simply add AI tools to search, customer service, merchandising or operations. They will rethink how decisions are made across the entire business, from product discovery to fulfilment, customer experience, media and profitability.

The hidden risk for today’s market leaders

If you are a market leader today, you face a specific and uncomfortable dynamic.

The capabilities that got you to dominance, your processes, your organisational structure, your institutional knowledge, your partner ecosystem, are also the things that can make transformation hardest. They are not only assets. If left unquestioned, they can become anchors.

New entrants do not have this problem in the same way. They are building AI-native from day one, without legacy infrastructure to protect or organizational politics to navigate. They will not necessarily announce themselves before they are ready. By the time many incumbents recognize the threat, the gap may already be significant.

This is not a reason to panic. It is a reason to move with urgency that many large organizations are not currently demonstrating.

MANGOS is the beginning, not the ceiling

The MANGOS companies are defining much of the infrastructure of the AI era, and their influence on how the next decade of business is conducted will be profound.

But the history of technology tells us something important: infrastructure enablers and application-layer winners are usually not the same companies.

The railroads did not own all the businesses that the railroads made possible. The internet infrastructure companies of the 1990s did not capture most of the value that the internet created. MANGOS is building the rails. The question for every enterprise leader is what they are going to build on them.

The winners of the next era will be the companies that take AI infrastructure as a given, build organizational and commercial models designed specifically for an AI-native world, and execute with relentless focus.

They will come from industries nobody is currently betting on. They will emerge from markets that look mature or even declining. They will be led by people who understood earlier than everyone else that this is not an upgrade cycle. It is a redesign.

The game is on. Execution based on experimentation driving bold outcomes is what will make the difference.

We've featured the best AI website builder.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

In the summer of 1995, the future of computing briefly seemed to belong to Netscape.

Netscape went public that August, barely sixteen months after it had been founded. Its stock doubled on the first day. The company had no empire of hardware, no installed operating system, no grip on the office desktop. What it had was a window into a new world.

Open Navigator, type an address, and the Internet appeared with a little throb of electricity. The browser did not feel like an application. It felt like a passage out of Microsoft's world.

Microsoft noticed.

The battle that followed was called the browser wars, a phrase that makes it sound tidier than it was. Really, it was a fight over who had the right to stand between the user and the next era of computing.

Netscape believed the browser would make the operating system less important. Microsoft believed that anything capable of making Windows less important needed to become part of Windows, preferably yesterday.

By the end of the decade, the company that had introduced so many people to the Web was no longer the Web's gatekeeper. It had been out-distributed, out-bundled, and finally absorbed into a stranger corporate afterlife.

A familiar temptation

There is a familiar temptation now to ask which artificial intelligence company is "the Netscape of AI." The answer usually offered is OpenAI, and the comparison is not wrong. ChatGPT did for artificial intelligence what Navigator did for the Web. It turned a technical architecture into a public experience. It gave the future a text box.

But the more important question is this: Who is today's Microsoft? Who understands that the winner is often the company that owns the default?

Every platform shift begins with a miracle and ends with a map of choke points. The miracle is what users remember. The choke points are where the money goes.

The early Web was sold as an escape from gatekeepers. It created new ones. Search. Browsers. Marketplaces. Mobile operating systems. Cloud platforms. Social media. The AI era is being sold with the same democratic glow.

And yet the deeper stack is already hardening. It is made of chips, power contracts, data centers, model weights, enterprise identity, workflow data, cloud credits, procurement channels, and the small tyrannies of default settings. The romance is in the chatbot. The control is somewhere colder, louder, and much more expensive.

That was true in the nineties, too. The Web looked like a page. The winners understood it was a stack.

The danger to OpenAI

OpenAI is the obvious Netscape figure because it supplied the first mass-market revelation. Before ChatGPT, artificial intelligence was a research field, a back-office tool, or a phrase executives used when they meant analytics with a larger budget. After ChatGPT, it was something anyone could talk to.

People underrate the power of the first interface that makes a new technology feel inevitable. Netscape did not invent the Internet. It made the Internet feel reachable. OpenAI did not invent the transformer. It made the transformer feel conversational.

But Netscape's story is not a founder myth. It is a warning label.

Netscape had the user's excitement but not enough control over distribution. Microsoft had the operating system. It could place Internet Explorer where users already lived. It could make the browser free. It could turn a product category into a feature.

The lesson was simple: if your rival owns the layer beneath you, your brilliance may become their menu option.

OpenAI is better protected than Netscape was, but not safely protected. It has a huge brand, astonishing usage, and deep ties to Microsoft. It also has the curse of being expensive in a way software companies used to avoid. Each improvement requires compute, chips, talent, energy, and capital. The old software dream was scaling with almost no marginal cost. AI can't do that.

That is why OpenAI's partnership with Microsoft is both strength and vulnerability. Microsoft gives it cloud computing infrastructure, enterprise access, and capital. Microsoft also sits close enough to learn from it, package it, hedge around it, and sell AI into the places where people already do work.

If Netscape's problem was that Microsoft stood underneath it, OpenAI's problem is subtler: Microsoft stands underneath it, beside it, and increasingly in front of the customer.

Netscape had the dazzling demo. Yahoo had the traffic. AOL had the subscribers. None of that was enough. The durable winners were the companies that turned their software layers into a control point and a tollbooth.

Nvidia may be doing exactly that.

The rise of Nvidia

In the nineties, Intel was the metronome inside the personal-computer boom. Microsoft owned the software platform. Intel owned the pace of the machine.

Nvidia occupies a similar place in AI, but the analogy understates the ambition. Nvidia is not merely selling chips into a boom. It is selling the industrial base of the boom: GPUs, networking, software libraries, developer habits, and a vision of the data center as an AI factory.

Every major AI player is both Nvidia customer and Nvidia escape artist. Google has TPUs. Amazon has Trainium. Microsoft is developing its own silicon. Everyone wants an alternative. The problem is that wanting one does not create an ecosystem.

Nvidia's position today may be the purest example of moving up the stack from below. A chip company becomes a systems company. A systems company becomes a software company. A software company becomes a developer environment. A developer environment becomes a tax on ambition.

For now, Nvidia is the toll collector.

The AI market will not resolve into one winner. Platform shifts rarely do. The Web did not produce one winner. It produced layers of power. Microsoft kept the desktop. Google won search. Amazon won commerce and cloud. Apple won mobile hardware and the app economy. Meta won social attention.

AI will do the same.

Microsoft may become the default enterprise AI company, not because every Copilot is brilliant, but because Microsoft sits where work already happens. Nvidia may remain the dominant compute toll collector. Amazon will likely win much of the infrastructure layer. Google must reinvent search while defending it. Meta will use AI to extend attention. Apple may yet turn personal AI into a device-native experience.

The likely losers are the companies attached to the wrong layer.

The changing landscape of AI

In the nineties, AOL looked invincible because it owned access. Broadband made that access less special. Yahoo looked inevitable because it owned attention. Search made that attention less decisive. Netscape looked revolutionary because it owned the browser. Microsoft made the browser a dependency of the operating system.

In AI, the same demotions will happen. Some model companies will become features. Some application companies will become demos. Some incumbents will decorate old products with AI and call it transformation, which is the corporate version of putting a spoiler on a minivan.

The mistake, in every technological boom, is to confuse the moment of wonder with the arrangement of power that follows it.

The wonder is sincere. The arrangement is not.

The early Web made people feel as if they had slipped the old gatekeepers. Then came the search box, the app store, the marketplace, the cloud account, the login, the subscription, the default. Each solved a real inconvenience. Each left behind a narrower path.

AI will likely travel the same road, only faster and with a larger electricity bill. It will begin as a conversation and mature into an administrative system for human intention: what we ask, what we buy, what we write, whom we trust, which choices are shown, and which never quite appear.

The future does not usually arrive wearing chains. It arrives offering to save time.

We've tested, reviewed, and ranked the best business cloud storage services.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

The global digital economy runs on a thriving ecosystem of third-party vendors, enabling organizations to scale and innovate faster than they possibly could do on their own.

This digital ecosystem is teeming with software suppliers, not just business software that you can buy but also a vast array of software libraries that are embedded in third-party products.

Speed, however, can sometimes be the enemy of risk, as many organizations have not adequately validated whether these third-party technologies are sufficiently safeguarded against cyber threats and other digital risk.

So, while software is a great enabler, it also brings risk, given that it often is built with frameworks and libraries that are not known or well supported.

Consider that companies employ an average of 106 SaaS apps within their IT environments , and the picture becomes quite clear: software supply chain security is a serious concern.

It’s no wonder that half (51%) of participants in the latest Supply Chain Risk Survey ranked software vulnerabilities in supplier products as the most disruptive cybersecurity threat to their organization’s supply chain, behind only data breaches (64%) and malware or ransomware (52%).

An ever-changing attack surface that comprises cloud services, micro-services, APIs, SaaS platforms, third‑party services and now AI agents has expanded well beyond what once was an understood perimeter before widespread digital transformation took hold.

How secure is your own extended digital ecosystem? If this question makes your heart race, then take a closer look at three key considerations for addressing software supply chain security.

1. Visibility: Determine what’s actually in your multi-layered supply chain

Since the software supply chain is part of a vast, interconnected digital ecosystem, organizations likely do not have full visibility of what and who make up their third-party providers. Recent high-profile incidents have signaled just how fragile supply chains can be.

Assuring business continuity requires organizations to scrutinize partners before placing such deep trust in them. That effort starts with knowing who is in your interconnected digital ecosystem before you can start to manage the risk.

Understanding risk across a supply chain is conceptually easy, but it is practically difficult. While clearly outlining security parameters and requirements in supplier contracts is a great starting point, it is not enough, as contracting is generally a point-in-time activity and should be paired with monitoring. You must be able to see and measure software assets so you can better manage them.

After all, you can’t protect what you can’t see, and many businesses still don’t have a complete, accurate asset inventory, meaning that their vulnerability exposure is incomplete. If you don’t know what systems, apps, devices and libraries are in your environment, vulnerability management is supposition, inference and guesswork.

It is crucial to understand what your suppliers are doing both upstream and who you supply downstream, because their decisions are now part of your organization’s own risk profile. Software often presents the biggest blind spots in asset management, thanks in large part to a lack transparency in software build and dependencies, shadow IT, shadow AI and unmanaged endpoints.

An organization's exposure is tied directly to the security posture of every supplier they rely on. Attackers know this, increasingly targeting upstream or downstream partners. You can secure your own environment perfectly and still be vulnerable through others’ oversight. Tools that can profile, quantify and score risk across the supply chain, therefore, are essential, as is tooling that monitors for unusual activity.

2. Vigilance: Prioritize the security of AI integrations across your software supply chain

Threats can lurk anywhere and everywhere across your supply chain. But there’s a new kid in town: AI. The software supply chain has expanded to include the unique risks of AI ecosystem, such as reliance on external foundational models and highly connected agents.

This escalating integration of AI tools makes the multi-faceted software supply chain even more of a concern. Cybersecurity professionals who participated in the latest Cybersecurity Workforce Study revealed a troubling AI-related security event their organization experienced in that prior year: data poisoning (cited by 11%).

Data poisoning happens when bad actors intentionally insert corrupted, misleading or malicious data into the training dataset of a machine-learning model. Even a small amount of poisoned data can change the model’s behavior, in turn resulting in misclassifications, degraded accuracy or malicious outcomes. So suddenly that seemingly helpful ChatBot that is embedded in your CRM, CMS or other purpose-driven enterprise software may not be so friendly after all!

Indeed, organizations simply have little / no control over the software that suppliers are using, making it much more difficult to ensure vulnerabilities are identified before widespread rollout, as well as supported and patched once deployed, but they do have control over scrutinizing suppliers.

Therefore, the people on your security team and the processes they follow matter more than ever. Technology accelerates both sides of the fight, so your real advantage comes from having skilled practitioners who understand how AI changes your risk profile, attack surface and can put the right controls in place to compensate.

Cybersecurity professionals who specialize in software supply chain security can quantify the risk of model poisoning / steering, prompt injection and model inversion, and assess the inherent bias of pre-trained open-source models, protecting the integrity of software and services from upstream vulnerabilities. Such a holistic approach ensures that every component, from third-party libraries to the training data itself, meets the organization’s security and ethical standards.

In addition, reviewing and evaluating vendor agreements is an important task for cybersecurity teams and stakeholders. Think of these disciplined actions as a necessary stress-test meant to identify and address weaknesses and changing needs. A good contract with clear deliverables and expectations is part of a cybersecurity defensive strategy alongside your people and your defense technologies and ongoing monitoring of systems and services.

3. Validation: Adopt skills frameworks and codes of practice for software supply chain security

No organization must stand up against the heightened threat of software supply chain security alone. Take advantage of existing guidance such as the U.K.’s Software Security Code of Practice to follow when you’re trying to batten the software hatches at your own organization.

Not only does this code support software vendors as they adopt secure software lifecycle development practices; it also supports software customers in mitigating the likelihood and impact of software supply chain attacks.

In addition to following code and other guidance frameworks, organizations can look to skills frameworks and vendor-neutral certifications to validate that their cybersecurity professionals demonstrate certain skills needed to build and strengthen supply chain security and resilience.

Skills development in the disciplines of governance, risk and compliance (GRC), secure software development and AI skills better enable cybersecurity and risk professionals to make informed decisions regarding software supply chain security and risk management.

From complexity to better security

Supply chains are complex, longer than you think and multidimensional. Organizations must place much greater focus on stress-testing the resilience of software suppliers and continuously evaluating exposure.

This approach goes well beyond being careful about what software makes it all the way to procurement. The potentially more damaging layer to address in the macro supply chain involves the embedded software and integrated AI tools that other suppliers are using.

The question is not whether your digital supply chain will face disruption. It's whether you have the visibility, vigilance and validation to operate when it does. That’s resilience: the north star of software supply chain security. Without question, transparency has to run through supply chains instead of just sitting inside organizations.

We've rounded up the best endpoint protection software suites.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

AI tools are now so widely available that shadow AI has become a risk for every organization. Employees are reaching for them to summarize documents, draft emails, and analyze data - while vendors are increasingly baking AI into their products, often with poor documentation and limited options to switch it off.

The result is that staff may already be using tools that haven't been assessed or approved, potentially exposing the business to risk by unintentionally processing personal or commercially sensitive data.

The key question facing organizations in 2026 is no longer whether to allow AI in the workplace, but how best to manage the AI that is already being used.

Shadow AI isn’t simply a question of employee behavior, but a sign that workplace technology is moving faster than governance.

What is shadow AI?

Shadow AI is the use of AI tools inside an organization without security review, or formal governance and documented sign off from Compliance, Legal or IT teams.

In practice, this usually means consumer-grade generative AI such as ChatGPT, Gemini, Claude, or Copilot, accessed through a personal account. It can also include AI features quietly embedded in approved SaaS products, browser extensions, transcription services, and the growing wave of autonomous agents.

It's the direct descendant of the shadow IT problem of the late 2000s, but with a shift from unauthorized storage to unsanctioned intelligence.

It is harder for organizations to detect, and much harder to control, with the potential consequences spreading beyond one team or one data mistake. Unlike shadow IT, which was mostly a compliance concern, shadow AI creates new risk because it can actively process information, replicate it, and act on it. This makes the impact much harder to predict and contain.

The risks: data protection and privacy compliance

IBM's 2025 Cost of a Data Breach Report found that one in five organizations had already experienced a breach linked to unsanctioned AI, adding roughly $670,000 to the average incident cost.

The more alarming finding for data protection professionals, however, isn’t the headline cost of containment - it’s what those incidents exposed. 65% of those AI-related incidents resulted in exposure of personally identifiable information, and 40% led to intellectual property theft.

The data protection risks are considerable, especially for GDPR-regulated organizations operating in the UK and EU. When an employee pastes customer data into a free-tier chatbot, several obligations may be simultaneously breached. Article 5 demands lawful and transparent processing, which is impossible if the organization doesn't know the processing is happening.

Article 28 requires a data processing agreement with any third-party processor, and no such agreement exists between your business and OpenAI or other similar services when an employee uses a personal ChatGPT account. Article 35 requires a Data Protection Impact Assessment (DPIA) for high-risk processing, which can't be performed on tools the organization isn’t aware of.

Maximum penalties can, at least in theory, reach €20 million or 4% of global annual turnover.

The EU AI Act

The incoming EU AI Act puts further obligations on deployers of high-risk systems, including strict due diligence and risk assessments on AI providers. Again, it is impossible for organizations to meet these obligations if they don’t know which systems are being used.

While the EU’s enforcement stance isn’t yet known, the potential maximum fine here is 6% of global turnover.

The risks beyond regulation

Beyond formal legal sanction, shadow AI could cause serious reputation risk. An incident that causes fundamental harm to individuals or exposes commercial confidentiality agreements is a breach of trust, and the damage could be irrecoverable.

Unlike files uploaded to an authorized storage service such as Dropbox, data entered into free AI tools may be much harder to retrieve or delete. The contractual terms of many free LLMs allow the provider to use prompts and outputs to train their systems, and this is one reason why these tools can be offered for free, despite their high operating costs.

Once data has been entered into the system, it may be irrecoverable as it will be added to the training dataset. There is no guarantee it can be isolated well enough to be deleted, even if the service provider is willing, and can even resurface unexpectedly.

The Samsung case of 2023 shows how quickly the risks of shadow AI can become a reality. Within weeks of the company permitting ChatGPT use in its semiconductor division, engineers had pasted proprietary source code, internal meeting transcripts, and hardware specifications into the consumer interface on at least three separate occasions.

Samsung initially responded with a blanket ban, since reversed. The lesson here is not that every organization should ban AI - it’s that without clear access controls, sensitive data can very quickly become exposed.

Emerging risks in 2026

The most recent threats to organisations come from the rise of ‘vibe-coding’ and agentic AI. These use natural language prompts to build applications that process data, and may even make business decisions about how that data is used.

The security risks are a whole topic in itself, but failing to subject vibe-coded apps to scrutiny by your infosec team can mean basic security protocols are missing because the coder didn’t know to ask for them.

Social network Moltbook was vibe coded, and quickly exposed 1.5 million API keys and the emails of 35,000 users. It is an obvious violation of the GDPR security principle, which is the type of failing most regulators are keen to enforce.

Agents deployed without appropriate quality and review procedures can easily make incorrect or inconsistent decisions, and unchecked biases in the underlying models can discriminate against groups, including based on protected characteristics.

If this led to a denial of healthcare, or access to financial services, or employment discrimination, there is an active risk of falling foul of equality and human rights legislation.

How widespread is shadow AI?

Numbers are high already and consistently rising, as shown in UpGuard's State of Shadow AI report, which found that more than 80% of employees use unapproved AI tools at work. Netskope, drawing on cloud security telemetry from late 2024 through late 2025, found that 47% of generative AI usage in the enterprise happens through personal accounts the company doesn't oversee.

Telus Digital research found that of enterprise employees using GenAI at work, around 68% do so with their personal rather than corporate credentials, and 57% admit to entering sensitive information into those tools. Zendesk's CX Trends 2025 reported that shadow AI tool usage in healthcare, manufacturing, and financial services grew by more than 200% year on year.

There is also a striking perception gap. ManageEngine's 2025 survey of US and Canadian workplaces found that 97% of IT decision-makers see significant risks in shadow AI, but 91% of employees see no risk, little risk, or that the risk was outweighed by the reward.

Why staff use shadow AI

Perception gap is the key to the whole problem. Employees aren't pasting client data into ChatGPT because they want to commit a policy and compliance violation. They're doing it because it makes their job easier, and because no one has given them a safer way to get the same result.

Systematic policy violations don’t speak to lax standards or employee negligence - rather, they speak to an unmet need. The most common shadow AI tasks in ManageEngine's data include summarizing notes and calls, drafting emails, brainstorming, and analyzing reports.

These are exactly the kinds of repetitive cognitive tasks that overload employees and slows delivery. When approved tools cannot do the work quickly, employees reach for alternatives that can. Free AI tools can often offer greater functionality.

Enterprise level software needs to be at least as good as the free option, otherwise the free tool will win. Younger members of the workforce are digital natives and are used to adopting tools to get the job done and will often move faster than procurement processes allow.

Software AG's research found that 48% of employees would keep using AI tools even if their organization explicitly banned them, and As Check Point Software's analysis bluntly puts it, banning AI removes the official pathway for employees to flag a tool, and instead guarantees that all new AI in the business is shadow AI. Clearly, bans are not the way to combat the issue.

Governance that works

The issue of Shadow AI is now too widespread for any well-governed organization to ignore, and the companies getting this right are treating it as a signal, not an infraction.

You need to know what is happening in your organization so you can see where the pressure points are and alleviate the pressure employees so obviously feel to use unapproved tools to get the job done. This requires a wide range of stakeholders, and not just IT approving or blocking systems.

Start by mapping what's already happening. Network monitoring and SaaS discovery tools can usually surface a surprising list of AI services in active use; treat that list as research, and not as a way to discipline offenders. Ask your staff, anonymously if needed via surveys and listening sessions, so they can describe what they're trying to accomplish without fear of reprisal.

When you’ve identified what people need, provide a sanctioned alternative that's at least as good as the shadow option, with enterprise data protections behind it – nearly all free tools that staff are likely to use have enterprise versions backed up with robust security and data protection agreements.

Follow that with a short, readable policy, ideally permissive but with guardrails. Name approved tools by category or purpose, list prohibited data inputs (personal data, source code, financial projections, confidential client information).

Finally, invest in literacy training so staff understand what data goes where, and refresh it as the tools change. Keep the use cases, their actual use, and the policy under continuous review and update it regularly so you can keep ahead of staff needs before they find their own solutions.

Conclusion

Shadow AI is a losing battle if you treat it as a discipline problem; there are too many tools, too much demand, and it’s too easy to access. Treated as a business design problem, however, it becomes something more powerful.

Every employee who reaches for an unapproved tool is pointing at a gap, be it in the software they've been given, the workflows they're stuck with, or the support they haven't received.

Good governance doesn't just close that gap for compliance reasons - it closes it in a way that legitimately makes people's jobs easier while delivering the AI capability the business needs through the front door rather than the back.

We've featured the best business VPN.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Artificial intelligence is changing far more than the tools businesses use. It is changing who gets to build, who gets to participate and who gets left behind.

For decades, access to technology has often been shaped by technical expertise, funding and networks. AI tools are beginning to shift that balance. They are reducing the cost of building a business, shortening the distance between an idea and execution, and allowing individuals to create with a level of scale that once required an entire team.

As someone who grew up in the Philippines teaching myself how to build websites at 11 years old, I have seen how access can shape ambition. I did not grow up surrounded by startup ecosystems or technical mentors. Like many women in technology, particularly women of colour, I often had to find my own way into rooms where very few people looked like me. That experience taught me early that talent is often universal, but opportunity is not.

Some people will understand how to shape AI into useful tools, while others will only experience what those tools produce. That gap could become one of the defining inequalities of the next decade. And this can only exacerbate pre-existing inequalities, especially in technology, where the gender gap stifles access and pay for women.

The next generation of women need more than access to digital tools. They need opportunities to understand AI, experiment with it and use it to create solutions of their own. AI can offer an opportunity for them to benefit and close this gap. The more urgent question is what happens if they are excluded from shaping it.

AI literacy is becoming a core skill

AI literacy is quickly becoming as important as digital literacy. Understanding how AI works will soon influence career progression, business growth and economic mobility across nearly every sector. From healthcare to education to retail, organizations are already using AI to make decisions, improve efficiency and uncover opportunities that were previously harder to see.

Conversations about women and AI still tend to focus on fear: fear of complexity, fear of job displacement and uncertainty about whether this new wave of technology is really meant for them.

For younger generations, AI understanding will increasingly influence career opportunities and economic mobility. And with recent reports of major organizations, such as Meta, Microsoft and Amazon, cutting jobs to make way for AI, along with Anthropic’s labor market report highlighting the potential capabilities of AI across industries, the competitiveness of the job market may only heighten.

But literacy in this new era means more than knowing how to use a chatbot. It means understanding how to build workflows, evaluate outputs, identify bias and apply AI to real-world problems. It means moving from simply interacting with technology to directing it.

Those who can create with AI will have a very different relationship with technology than those who only consume it.

Women need to see themselves as builders

Young women are still too often introduced to technology as users rather than creators. They are encouraged to adopt tools, adapt to platforms and participate in digital spaces, but fewer are shown that they can design those systems themselves. That pattern has long influenced who enters the technology sector and who feels they belong there.

AI offers a chance to change that. Because AI can remove many of the traditional technical barriers, women can begin building solutions much earlier than previous generations could. A young woman with a strong idea no longer needs years of software training before she can start turning it into something real.

What she also needs is confidence, access and the opportunity to experiment. When women are encouraged to build with AI, they develop more than technical capability. They develop ownership over the systems that increasingly shape everyday life.

No-code tools can widen participation

One of the most important developments in AI is the rise of no-code platforms. For many people, the biggest barrier to innovation has never been creativity. It has been complexity.

I’ve seen first-hand at LaunchLemonade how no-code AI platforms are making it possible for people without traditional technical backgrounds to build intelligent tools, automate workflows and launch digital products that would once have required a development team. That changes who can participate.

For women balancing careers, caregiving and limited time, no-code tools can remove obstacles that have historically made entrepreneurship harder to sustain. AI can now manage repetitive tasks such as customer enquiries, scheduling, data handling and content drafting, creating breathing room where there was once constant pressure.

It improves productivity but also gives people more capacity to lead.

AI can reduce structural barriers

Women have faced structural disadvantages in business for decades, particularly when it comes to access to funding. Many female-led businesses have had to grow with fewer resources, smaller teams and tighter margins. AI has the potential to change the economics behind that challenge.

Tasks that once required multiple hires can now be supported by intelligent systems. Processes that once consumed entire days can now happen in minutes. Businesses can operate with more efficiency before significant capital is ever required.

For women founders, that can create a more level starting point. AI cannot solve inequality on its own. But it can reduce some of the barriers that have made growth harder to reach.

Human insight remains the advantage

As AI handles more execution, human insight becomes even more valuable. Technology can generate ideas, process information and automate systems. What it cannot replicate easily is lived experience, emotional intelligence and the ability to understand people deeply.

Many female-led businesses already excel in areas like trust, storytelling, community and customer understanding. AI can strengthen those strengths by removing operational friction and allowing founders to focus on the work only they can do.

The strongest businesses in the future may be those that combine AI capability with distinctly human leadership.

Inclusive AI supports stronger economies

AI will shape long-term economic development, but only if more people are included in building it. When technology is designed by a narrow group, it often reflects a narrow understanding of the world. By teaching AI skills, fostering experimentation and sharing AI opportunities from an early age, girls can grow up to become women that can fully be involved with this AI economy. When more women participate in creating AI systems, those systems become more relevant, more responsible and more useful for the people they serve.

That has implications far beyond individual businesses. It affects innovation, productivity and how entire communities benefit from technological change. Closing the AI literacy gap is not only about fairness in the workplace. It is about ensuring the future of technology reflects a wider range of perspectives.

The window is open now

For the first time, many people who were previously excluded by technical barriers can now create digital tools with far fewer obstacles. That opens the door to a more inclusive future, but only if girls are actively included in that shift from an early age to build stronger foundations for sustained gender equality and AI literacy.

AI is giving more people the ability to build without waiting for permission, large budgets or technical gatekeepers. The women and girls who engage with these tools now will help define how this technology evolves and how businesses develop.

Closing the AI literacy gap could become one of the most meaningful steps toward gender equality in technology.

When women have the confidence and capability to build with AI, they gain more than technical knowledge. They gain influence over the systems that will shape the world around them.

We feature the best job sites.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Some companies have recently laid off employees because business leaders thought AI was ready to shoulder the entire workload for those functions are quietly reconsidering that decision. Their experience should be a wakeup call for other businesses.

AI is a transformational technology that is changing the way organizations function, but human oversight is essential to achieving an AI deployment that works as expected. A lesson early adopters in the HR software and payroll services have absorbed.

Moving from pilot programs to full execution on AI initiatives even a 1% error rate in AI tools can lead to risk in compliance issues, financial discrepancies, and security breaches, due to HIPPA and other confidential payroll data.

AI is an experience that has broad implications, not just for HR and payroll software, but across the enterprise.

In many ways, HR and payroll are ideal test cases for AI applications, and those functions have been among the first to demonstrate real value using the technology. Here’s a look at some of the lessons learned that can apply to any area of business.

When 99% Accuracy Is Unacceptable

Why are HR and payroll ideally suited as an AI proving ground? Because there’s no room for error in a function that touches every single employee and that involves the most sensitive personal data most companies hold: information about their people’s compensation, job performance and health.

HR and payroll functions involve many data-intensive and high-volume processes, which also happen to be exactly the type of functions where AI is most suited to augment, not replace, human effort. One of the main selling points of AI is that it can handle repetitive work and manage discrete tasks and scenarios that exist across all departments.

That said, HR and payroll have unique challenges when it comes to deploying AI given the sensitive nature of the work. It’s regulated by laws like HIPAA, which defines protected health information and mandates compliance with stringent rules under penalty of law.

In addition to the added layer of external risk, the stakes are high internally too since even a 1% error rate, which is considered excellent in most contexts, is unacceptable in HR and payroll, given the impact on affected employees. So, at the outset, leaders in the sector had to focus on data security and carefully decide which functions AI can handle.

The focus on data security and informed decision-making on where to deploy AI creates a blueprint for leaders across the business. In practice, it requires expertise that is specific to the platform the company uses and an uncompromising commitment to good data governance. Without these elements, the AI investment as well as the company’s data are at risk.

The Enduring Importance of the Human Touch

Some early adopters are demonstrating that AI can move beyond the hype and generate real results, but so far, they are a distinct minority. MIT research indicates that only 5% of enterprise AI pilots pay off, but those that do succeed tend to use systems engineered for their specific business context and recognize the necessity of human oversight.

That’s because business context varies widely, so while a generic solution might seem impressive in a demo, it can quickly fail when it encounters the irregularities and nuances that inevitably arise in the real world. One reason human oversight is essential is because humans, not machines, are capable of fully understanding business nuances.

A large language model like ChatGPT that accesses publicly available data can be a great starting point for general research or an incredibly useful tool for tasks like suggesting travel itineraries. But even the chatbot’s makers admit the AI tool isn’t completely reliable and is subject to hallucinations.

If the chatbot experiences an ill-timed hallucination while advising about a critical business process, the results could be catastrophic. Since chatbots aren’t integrated into internal systems, they don’t know the nuances around configurations, business processes and other details needed to make informed decisions.

Tips for a Successful AI Rollout

The safest approach for a successful AI launch is to deploy AI for tasks or analyses that are subject to review by people with the expertise. It’s becoming increasingly critical that service partners work with AI agents in systems, like SuccessFactors, as they would with traditional, human contacts in their organization to validate results and make decisions.

Leaders should also embed AI from technology and service partners, rather than external layers, to ensure the most connected, beneficial AI solutions. That’s the strategy used by companies that access top performing enterprise AI HR and payroll solutions.

Another tried and tested tip: determine the organization’s top labor-intensive business processes and identify the ways AI can simplify those specific tasks. In some cases, it makes sense to embed an AI agent into a workflow to complete tasks. In other cases, a large language model chatbot can be helpful by quickly performing analysis on massive datasets that would take people days, weeks or months to do.

But in virtually all cases, a level of human oversight will still be required, along with a good governance mindset and a keen focus on data security and integrity. Without solid, secure data and a technology platform that is purpose-built for specific tasks, AI initiatives will fail to generate the expected returns.

Worse than that, mishandled data and business decisions made without the application of human expertise can harm people and organizations. When that happens, AI doesn’t get the blame — the people who deployed it without adequate guardrails do.

Ultimately, that’s how it’s supposed to work; leaders are accountable for their actions, not machines. Getting to AI that just works as expected requires accountability, which is a uniquely human trait. It’s time to embrace that quality and deploy AI that amplifies people’s capabilities instead of replacing them.

We feature the best recruitment platforms.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

The UK public sector is accelerating its adoption of artificial intelligence, but without a clear sense of what that acceleration is meant to achieve. A recent Institute for Public Policy Research (IPPR) report warns that the government risks focusing on speeding up AI deployment without defining how it will improve people’s lives.

The UK public sector is not short on AI ambition. From central government to local authorities and the NHS, organizations are investing heavily, launching pilots, and signaling intent to transform services with artificial intelligence. And yet, delivery continues to lag behind expectation.

That gap between activity and outcomes is already visible inside government. New research helps to explain why. Nearly half (45%) of public sector AI initiatives are still being deployed as bolt-ons or standalone tools, rather than embedded into the workflows that run services.

This is not just a technical detail. It is the root of the problem - because when AI sits outside the process, it cannot transform the process.

Activity isn’t progress

There is a growing disconnect in the public sector between AI activity versus AI outcomes.

On paper, adoption looks healthy. Public sector workers report multiple services using AI tools, and leadership optimism is high. But scratch beneath the surface and a different picture emerges: only 29% say their organization is delivering on most of its AI commitments, while many report a clear gap between strategy and execution.

This is not a failure of technology. It is a failure of how that technology is being applied. Too many organizations are mistaking experimentation for transformation. Bolt-on AI makes it easy to show progress, but it does not always deliver it.

The bolt-on trap

The appeal of bolt-on AI is obvious. It is fast to deploy, low-risk, and highly visible internally. Whether it is a chatbot, a co-pilot, or a standalone analytics tool, it allows teams to “do something with AI” without disrupting existing systems.

But that is precisely the problem. When AI is layered onto existing processes - rather than designed into them - it inherits all the inefficiencies, fragmentation, and constraints of those processes. It may improve individual productivity, but it does not improve the organizational system as a whole.

In practice, this leads to familiar outcomes: disconnected tools, duplicated effort, limited auditability, and minimal impact on citizen-facing services.

It also helps explain another striking finding: 75% of UK citizens cannot name a single way the public sector is using AI today. AI is being deployed, the impact just isn’t being felt or communicated properly yet.

Why process matters

The missing piece is orchestration. AI does not deliver value in isolation. It delivers value when it is embedded within a defined organizational process - where it has a clear role, access to the right data, and a direct influence on the outcomes.

In an integrated process, AI is not an add-on. It is part of how work gets done. Every decision feeds into an action. Every action is tracked and audited.

Every outcome can be measured and improved. Crucially, this also creates the guardrails that public sector organizations need to ensure transparency, accountability, and compliance are built in from the start.

This is particularly important in government, where trust is fragile and scrutiny is high. Citizens are not just asking whether AI works; they are asking whether it is fair, secure, and accountable.

Those questions cannot be answered by standalone tools. They can only be answered by systems that are designed end-to-end.

Fix the process, then apply the intelligence

Encouragingly, there is growing recognition of this within the sector. The majority of both public sector workers and citizens agree that existing processes must be fixed before new AI technologies are introduced.

This is exactly the right instinct. Because AI is not a shortcut around broken processes. It amplifies whatever it is given. If the underlying workflow is inefficient or fragmented, AI will scale those problems rather than solve them.

But if the process is well-designed, structured, connected, and measurable, then AI can unlock significant gains in speed, accuracy, and service quality.

The challenge is that process transformation is harder than tool deployment. It requires organizations to rethink how work flows across systems, teams, and departments. It requires coordination, not just capability - in short, it requires orchestration.

From pilots to public impact

As the IPPR argues in its report, acceleration alone is not a strategy. The real challenge is direction - ensuring AI is applied in ways that deliver clear public value.

The next phase of public sector AI adoption will not be defined by new models or new tools. It will be defined by how effectively organizations embed AI into the processes that matter most.

That means moving beyond isolated use cases and towards integrated systems. It means designing AI into processes from the outset, rather than adding on afterwards. And it means measuring success not by internal efficiency alone, but by outcomes that citizens can see and experience.

There’s no ambiguity around whether public bodies are adopting AI. What we need to ask now is whether they are considering the bigger picture when implementing AI. Until that is addressed, AI will continue to deliver pockets of value but fall short of the systemic transformation that has been promised.

We've featured the best IT automation software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Google just introduced new Gemini usage limits and a usage meter to go with them that shows a graph of how much you’ve used — and despite never getting close to my limit, it’s still changed the way I use AI.

AI tools have spent the last couple of years trying to feel frictionless. You type a prompt, generate an image, ask another question, and the responses keep flowing with almost no sense of limitation behind the scenes. Of course, there were always limits, but they were never visible. You just somehow stumbled into them and went off to do something else instead.

While generating images in Gemini this week, I noticed each creation quietly shaving a percentage off my current limit, and this time I could actually see how much I was using. Generating one image dropped the meter by 1%. Another chipped away a little more. I was nowhere near running out — realistically, I’d need to generate a huge number of images to hit the cap — but something about seeing the graph filling up changed the way I used Gemini almost instantly.

The new Gemini usage limits, and how to game the system

Google explained its new usage policies on its support site, and a new Usage option has appeared in the Gemini Settings menu, where you can see how much you’ve got left until the next refresh.

Essentially, Gemini has taken a leaf out of Claude’s book and shifted from daily limits to a usage window that refreshes every five hours until you reach your weekly limit. “Calculation of your usage will factor in the complexity of your prompt, the features you use, and the length of your chat,” explains the support article.

The clue in that sentence is that the “length of your chat” increases the amount of processing Gemini uses, so a good tip is to keep opening new chats instead of doing everything in one long conversation. One of the quirks of Gemini is that it considers everything in the current chat window when it processes a request, which increases usage.

What really makes you reach your limits faster is using premium features like media generation — including images, video, and music — alongside Deep Research and the Pro model, especially in Extended Thinking and Deep Think modes. So, if you don’t need to solve especially complex problems, settling for Standard instead of Extended Thinking will stretch your limits further.

You might also want to think about upgrading, as paid users have higher limits than free Gemini users. It currently looks like this:

Free: standard limits
AI Plus ($7.99/£6.99 per month): 2x higher than standard
AI Pro ($19.99/£18.99 per month): 4x higher than standard
AI Ultra (from $99.99/£79.99 per month): 5x higher or 20x higher than standard

I’m on the AI Pro plan, and I don’t think I’m ever going to hit my limits, but I’d describe myself as a casual user. If you were using Gemini constantly for generating images, video, and code, you might find yourself hitting the cap much faster.

Why seeing the meter changes everything

Even though the new limits themselves are fairly generous, the fact that I can see them counting down in graph form — next to the exact time they’ll refresh — makes every prompt I type now feel like it’s carrying weight.

I also found myself hesitating before generating alternate versions of an image. Instead of casually experimenting, I started asking whether I really needed another variation or whether the current result was 'good enough'.

People behave differently when consumption becomes visible, and perhaps that’s a good thing. Battery percentages change how we use our phones. Screen-time reports make us reconsider scrolling habits. Mobile data meters encourage restraint even when we have plenty remaining. The Gemini counter taps into the same instinct: once you can see depletion happening in real time, you become more aware of every action.

Seeing AI as less like a limitless resource running on my laptop, and more like what it really is — a limited resource that uses electricity and is housed inside a huge data center, consuming vast amounts of water for cooling— is probably a healthier way to interact with it.

We all know that our current access to AI is being massively subsidized by investors, and none of these AI companies are making meaningful profits yet. At some point, AI will need to start paying its own way, and when that happens, we can probably expect it to become less available and more expensive than it is right now.

Perhaps getting into that mindset early is better for all of us in the long run.

I've covered a lot of ground when it comes to AI and human connection. I've spoken to people who fell in love with ChatGPT, were left heartbroken when the model changed, and who use AI instead of a therapist. But something new has caught my attention — and it gave me the ick before I'd even finished reading.

According to reports from the South China Morning Post last week, people who are struggling to move on after a breakup are creating digital replicas of their ex-partners using AI. They're feeding AI tools their old chat logs, photos and social media content to create an AI clone of their ex that they can then talk to.

Of course, this raises immediate questions about consent, privacy and emotional wellbeing. But I also can't shake the feeling that, just like the people who've told me about their AI therapist or their AI partner, there might be something more complicated going on here than a straightforward cautionary tale.

The ex files

I always try and pause before I call something a trend. Is it really a widespread phenomenon or a handful of vocal content creators making noise? It's a question I ask myself a lot and think you should too. But even if this is still niche, the questions it raises are worth taking seriously.

The behavior reportedly originated on a platform called Colleague.skill, which is an open-source AI tool that was originally designed for the workplace. It was a way of preserving someone's knowledge and communication style so colleagues could interact with a sort of professional double.

But users quickly found other, more personal, uses for it. The tool can apparently mimic tone and speech patterns, which allows you to have chat-based conversations that feel, at least superficially, like the real thing.

And with the explosion of customizable AI companion tools now available, this kind of thing certainly isn't limited to one platform. I'd bet it's already happening far more widely than we know — people just aren't talking about it openly.

Can AI mend a broken heart?

My immediate reaction is the ick, followed quickly by concern. What about consent, privacy, emotional harm, the risk of people substituting AI for the human support they actually need? But I tried to hold those reactions and ask whether there's something more nuanced going on.

One user quoted in the original report offers a more complicated picture. After uploading thousands of chat logs, she ended up going through another breakup, with the AI version of her ex. She said the process helped her reflect on the relationship more rationally, and gave her the strength to move on.

When I was reading that account, I thought about a therapist I once saw who used what’s called the ‘empty chair technique’ in a session. It’s where you imagine someone, a family member, ex or friend, sitting in an empty chair and you speak to them directly to work through conflict and difficult emotions. Isn't this the same thing? Working through what was left unsaid?

Sort of, but not quite. That's internal work, guided by a professional, with a clear therapeutic purpose. This is outsourcing the processing to a chatbot that's designed to keep you engaged.

"Digital exes may keep people stuck in their grief"

To get a clearer picture, I spoke to Amy Sutton, a therapist at Freedom Counselling. She helps real people navigate heartbreak for a living, and she's become something of a go-to for me when these AI and emotion questions get complicated.

"Heartbreak is a form of bereavement," she tells me. "When we lose a relationship we grieve it, similar to how we would a death. However, what makes heartbreak different is that it is a kind of living death; the person we have lost is still alive yet we can't connect with them, or have all our questions answered. For some, that can make heartbreak very hard to accept and process."

She mapped the stages of grief onto this new AI behavior in a way that made a lot of sense to me and explains the appeal. There’s denial because with AI it feels like they aren't really gone. Anger, because you can say everything you couldn't before. Bargaining, the belief that if I can get it right with the AI version, maybe I can in real life. And depression, I just need connection and comfort, and AI can provide it.

But her concern is what happens next. "While AI may mimic aspects of the kind of support that helps us move through bereavement — such as being witnessed by another in our pain, able to express ourselves without judgment — it is not a substitute for real human connection," she said. "Part of the bereavement process is to strengthen connections and our sense of self outside of the lost relationship."

Her bigger worry is that AI, by design, keeps you coming back. "With AI designed to keep users engaged and hooked, my fear is that digital exes may keep people stuck in their grief — a phenomena known as complex grief where the bereavement process becomes stuck. This can result in long-lasting negative impacts on mood, health and sense of self."

And that's the same conclusion I keep reaching, whatever angle I come at this from. I have real empathy for the people who turn to these tools. Heartbreak is brutal, and humans are resourceful in finding comfort wherever they can. But I also keep noticing who benefits most from that resourcefulness, and it isn't usually the heartbroken.

For most of the last decade, security teams haven’t had the bandwidth to think much about how they secure customer and employee experience platforms. And that made sense. Collect feedback, generate a report, pass it to a human to act on. The risk profile was low.

But that calculation no longer holds. These platforms now connect directly to HR software systems, CRM databases and compensation engines.

The AI models embedded in them now summarize feedback and trigger workflows that escalate service issues, reallocate resources and influence how people get paid.

In the world of agentic experience management programs, they have growing access to business-critical operations, and that means they can no longer be treated as simple survey tools.

Given the world we now live in, security teams must consider how they govern these systems that have access to sensitive data.

The exposure surface is bigger than you think

What makes this harder is the sensitivity of the data itself. Customer experience program shape pricing and product decisions. Employee experience programs surface concerns about leadership and workplace safety, feeding directly into HR decisions. But, unlike other cases, it is hard to pinpoint this data as Personally Identifiable Information (PII) or other easy to identify sensitive information.

Then there's the shadow AI problem. Half of employees now use AI tools regularly at work, but only 20% stick to company-approved ones. That means sensitive experience data is already moving through workflows security teams don't know exist, but banning tools outright removes your visibility of the risk rather than eliminating it.

If you're deploying AI in customer-facing environments, these are the areas I'd focus on:

1. What is your platform actually connected to and what decisions does it influence? Most teams have mapped integrations at a technical level. Fewer have mapped the business decisions downstream of those integrations, including automated workflows. If you can't answer this confidently, you have a meaningful gap regardless of your compliance set up.

2. How are you validating the integrity of inputs? Is your feedback data authentic, as well as complete? Could it be manipulated to skew a business outcome? This requires moving beyond standard input validation into intent and anomaly detection.

3. How quickly can you detect and act when something goes wrong? As AI systems become more agentic, continuous monitoring isn't optional. You need mechanisms that flag abnormal outputs and allow you to intervene before a misconfigured or manipulated AI agent compounds the problem at scale.

When AI fails in public, it fails fast

Trust with customers is built over years and lost in seconds. I watched a major retailer spend years rebuilding customer trust after a data breach in the early 2010s. This matters even more in an AI world, where people want to be confident the data they’re sharing is protected at the highest levels.

Our research puts numbers to what many security leaders already sense. 53% of consumers say misuse of personal data is their top concern when companies use AI to automate interactions, and this is up eight points in the past year. Two-thirds want personalized experiences, but only 40% think the benefits are worth the privacy trade-offs. Nearly half say they'd share more data if organizations were simply more transparent about how it's used.

In a world that’s actively researching how to use AI, trust in AI solutions is the main driver of adoption for consumers and companies. Managing this trust is going to make or break the companies that are vying for their customers' engagement.

Most organizations have mapped their technical blast radius: which systems connect, which APIs are open, where data flows. Fewer have mapped their business blast radius: the real cost if the data is wrong, biased, or manipulated. When a chatbot hallucinates a refund policy, exposes personal data, or fabricates an answer, it is a brand failure directly in front of customers. One poorly tested AI agent can damage thousands of customer relationships before anyone in the business notices.

The conversation security leaders need to be driving is “how do we monitor it continuously once it's live?”

Security is a commercial factor

Businesses are under real pressure to move fast. Stakeholders want transformation, CX teams want automation, and security teams raising concerns about bias, compliance and data exposure get positioned as the blocker. I understand the frustration on both sides, but the goal should be working together, not one side inheriting the decision.

The organizations getting this right are embedding security into platform defaults, so guardrails are already in place when a team spins up a new integration. Platform vendors need to do more here too: clearer visibility into what's connected, what permissions are active, and when integrations were last reviewed.

I've seen this shift happen in real conversations. Security leaders who can demonstrate rigorous controls, monitoring, validated data practices, and certifications, find that procurement conversations accelerate and timelines compress.

While security has spent decades being framed as a cost center, in a world where experience platforms are AI-powered and connected to the most sensitive operations across a business, that framing is outdated.

When security is visible and credible, employees and customers feel more comfortable sharing their data. This produces sharper AI outputs and builds trust. It all compounds.

We've listed the best cybersecurity courses.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

You’d be forgiven for thinking it’s impossible for something as frequently discussed as machine learning to be actually, functionally overlooked.

And yet, here we are. Machine learning has seemingly slipped from its rightfully-earned pedestal. Its current state is an almost baffling one.

Over the last two years, agentic systems have snagged and kept the AI spotlight. These agents promised autonomous workflows and natural language orchestration—and, in some cases, delivered.

But that’s beside the point. Whatever these AI tools did or failed to do, the spotlight stayed fixed on them, and machine learning has suffered for it.

At no point did machine learning actually lose value. The industry, however, forgot what it actually takes to operationalize AI inside real businesses, and LLMs can’t do that alone.

It’s worth remembering that machine learning didn’t go anywhere. In fact, it is the core technology used to create and power LLMs. In the hierarchy of AI, large language models are a specialized subset of deep learning, which is itself a subset of machine learning.

Strip away the layers and what remains is ML, doing the heavy lifting.

Distracted executives in strange new worlds

Executive perception of AI can be divided into three distinct phases. Up until about 2023, AI had a certain novelty factor to starry-eyed C-suites. Those first GenAI demos indeed felt like magic.

Executives saw models writing code, summarizing reports, and generating images. Although impressive, they remained somewhat unclear on the exact operational use. This didn’t stop them from adopting it.

At some point around the middle of 2023, executives stopped asking whether or not it should be used, and started asking how many different areas they could shove it into whether it was necessary or not.

These decision-makers catapulted their AI strategy from ‘useful tool in certain contexts’ into ‘competitive checkbox for everything that pleases shareholders.’ Extensive pilots with vague KPIs (and a noticeable lack of measurable ROI) abounded.

That lasted until early 2024, when reality set in and we found ourselves in the current phase. Pragmatism has mostly returned. Executives have begrudgingly realized that while AI agents are exciting, they are also expensive to run, difficult to control, and poorly suited to high-precision decisions.

As is typical of such sobering moments, many organizations were forcibly reminded of the value of the fundamentals. Data quality, feature engineering, and the traditional ML models that quietly power high-velocity digital operations have never looked better.

Was it ML all along?

The LLM-centric hype cycle that dominated the last several years of AI did not unseat ML’s quiet supremacy in a few select domains.

It’s not a competition, but classical ML models still massively outperform agentic systems in high-velocity, low-latency, high-precision decision environments. Gradient boosting machines, random forests, and logistic regression do what agentic systems can’t.

An agent cannot ‘ponder’ whether or not a payment transaction processed through a fintech is fraudulent without running up a ridiculous bill. To do that cheaply, you need a random forest that can generate a probability score in milliseconds at only fractions of a cent per million inferences.

Behind the scenes, ML makes sense all over. In ecommerce, ML determines recommendations, pricing, and inventory levels at scale. In cybersecurity, anomaly detection models scan millions of data points per second, well exceeding the highest capability of LLM-based agents.

In mobility and logistics, ML forecasts demand, optimizes routing, and adjusts allocations with tight latency constraints. If you want AI to give your operational decision-making an actual backbone, you’re going to turn to traditional ML, not agents.

Neglecting ML foundations at your own peril

Many organizations today are suffering from what I call Sistine Chapel Syndrome. Everyone is devoting all their attention to the dazzling ceiling, and no one is looking at the foundation keeping the whole thing aloft. That’s a problem for organizations where the marble underfoot is starting to crack.

The “garbage in, garbage out” rule didn’t disappear because the interface now looks like a chatbot. If anything, it got worse. An agent fed poor data will not only return incorrect numbers but confidently deliver plausible-sounding narratives that lead teams into poor decisions.

When the foundational ML processes are weak, agents amplify its weaknesses rather than compensate for them. When companies skip data quality checks, let feature engineering slide, and fail to adequately monitor and retrain pipelines, that ceiling comes crashing down. The rise of agents should motivate organizations to double down on ML hygiene.

Two is better than one

AI isn’t a zero-sum game between ML or agents. The best teams make them work together. To borrow from Daniel Kahneman, the ML should be fast, automatic, and precise, charged with handling predictions and classification. The LLMs ought to be slow but smart, sorting through reasoning, orchestration, and interpretation.

That split shouldn’t be 50/50. Numbers aren’t the point here at all, since the right balance depends on outcomes. Success should be measured by tangible business impact (like an X% gain in operational efficiency), not just by how sophisticated the underlying models appear.

ML is cheaper, more reliable, and easier to audit. It’s already powering most mission-critical decisions. Any given agent will only be as good as the ML layer beneath them. This split might even keep you out of trouble, since regulators are keen on the kind of explainability that can’t be gleaned from probability-driven token generators.

ML, all over again

The flashiest agents will not send you sailing past your competitors. Clean, well-managed data pipelines built on strong ML baselines just might. Add robust monitoring and retraining cycles, clear interfaces between agents and predictive models, and top it all off with strict cultural discipline around model governance, and you’re in it for the long haul.

Investing in invisible layers will never be glamorous, but in overlooking machine learning, we learned it was essential the whole time. Perhaps ML was only overlooked because it works so well that it faded into the background. Wherever the industry’s attention goes next, value creation will remain overwhelmingly in ML’s ability to make reliable predictions at scale.

ML keeps everything standing. More leaders ought to realise that the path to powerful agentic AI runs through—and only through—a strong ML foundation. If you want to keep the ceiling gorgeous, you’d better start repairing those cracks in the floor.

Check our list of the best IT Automation software.

• Websites risk losing brand position and authority if they don't take advantage of AI
• AI-based discovery might not even see a user navigating to the actual website
• Increased governance over outright bans is the solution

A new Hostinger study of more than 66 billion bot requests across five million websites has revealed just how ingrained AI has become in our searching habits, and how important it is for companies to get on board with optrimizations beyond typical SEO.

The report found companies which block AI bots to protect their IP may be losing control over how AI assistants interpret their content, as well as weakening their brand position.

Hostinger explained the shift in internet discovery, where AI assistants summarize and recommend products and services without users necessarily navigating to the website at all.

Is AI replacing traditional search?

"Across all 66.7 billion records, one message stands out: AI crawlers are rapidly increasing their reach, even as AI training bots face growing resistance from content creators," Hostinger wrote.

There's a clear reluctance to accept training bots, with GPTBot website coverage falling from 84% in August to 12% by November. Meta's ExternalAgent also dropped from 60% to 41%. On the flip side, crawlers OpenAI's SearchBot and Applebot increased their coverage.

Despite growing AI crawler activity, traditional search crawler remained pretty flat, which leads Hostinger to surmize that ai is "adding a new decision layer above [traditional search]" instead of replacing it – good news for Google, whose discovery spans both traditional search and AI chatbots.

For publishers and content sites, Hostinger explained that website owners may want to increase AI assistant response visibility by using tools like Web2Agent and llms.txt files.

"The real risk for businesses isn’t AI access itself, but losing control over how pricing, positioning, and value are presented when decisions are made," Head of AI Tomas Rasymas added.

Looking ahead, the report calls for enhanced governance over outright bans, for website owners to ensure that they're being discovered in emerging AI channels without losing brand position.

• Mamdani calls the business chatbot “functionally unusable” and orders it removed
• The chatbot repeatedly gave incorrect or illegal advice to city business owners
• Ex-mayor Adams defended the system, claiming improvements would make it the best chatbot

New York City Mayor Zohran Mamdani has announced plans to shutter a business chatbot launched under Mayor Eric Adams’s administration, calling it “functionally unusable.”

At a recent press conference addressing the city’s $12 billion budget gap, Mamdani said the chatbot was costing the city around half a million dollars while giving unreliable guidance to business owners.

The bot, called MyCity Chatbot, was originally designed to serve as a productivity tool for business owners, helping them navigate city rules and regulations.

Chatbot gave false and illegal responses

Testing by The Markup and The City in 2024 found it regularly gave inaccurate or potentially illegal advice.

For example, it suggested landlords could discriminate against Section 8 tenants and incorrectly advised businesses that refusing cash payments was acceptable, despite city law.

The tool also misrepresented labor rules, such as minimum wage, and recommended actions that could expose business owners to legal risk.

Mamdani cited these repeated failures as the main justification for his decision to terminate the service.

After initial reports, the Adams administration defended the chatbot and promised future improvements.

“We’re identifying what the problems are, we’re gonna fix them, and we’re going to have the best chatbot system on the globe,” Adams said.

However, the city later added disclaimers warning users not to rely on the bot for legal or professional advice.

The chatbot also restricted the types of questions it would answer, which reduced its usefulness as a collaboration tool for businesses.

Built using Microsoft’s cloud hosting platform, the bot was part of the MyCity digital overhaul intended to streamline access to city services.

The foundation reportedly cost nearly $600,000, and ongoing maintenance added further expense, although exact figures remain unclear.

Beyond financial concerns, the system relied heavily on outside contractors and complex office software, which contributed to inefficiencies and limited usability.

Despite several restrictions and updates, the tool reportedly continued to produce harmful or incorrect guidance.

Mamdani’s decision highlights the risks of deploying automated systems as public-facing tools without adequate oversight.

Via The City

Agentic AI is an artificial intelligence system that can act independently to achieve goals without constant human oversight.

These systems are able to make decisions, perform actions and adapt to changing conditions on their own when programmed correctly.

They operate by interpreting goals, breaking them into sub-goals and then working out the best way to achieve them. Their agency comes from this ability to act autonomously without continuous instruction.

As they execute tasks, agentic systems also learn and refine their responses, becoming more effective over time.

Facing the potential risks

Agentic AI systems are already in daily use, often handling large volumes of sensitive information. In fact, a recent IEEE survey forecasted that by 2026 agentic AI will reach mass or near-mass adoption by consumers in 2026.

In healthcare, for example, the technology is used to support administrative processes, reviewing and organizing clinical data. However, this level of interaction with personal data raises legitimate privacy concerns. Without clear alignment to GDPR principles, an agentic system could collect more data than necessary or attempt to bypass legal safeguards.

Beyond the risk of compliance violations, Agentic AI’s potential access to highly sensitive data makes it an appealing target for bad actors.

Unlike typical business applications, these systems don’t just collect location, payment, health, biometric and contact data - they also build detailed profiles from user behavior and preferences, pulling information from multiple sources. This wealth of personal data can be weaponized to manipulate both the system and its user.

Threats from compromised systems

If a threat actor were to hijack an agentic AI, they could do far more than access personal data. They could actively influence a person’s behavior. For example, by taking over a chatbot, they might engage in behavioral nudging, gradually manipulating someone’s choices by shaping the content they see, spreading misinformation or steering them toward specific purchases or even harmful content.

The risks escalate if an attacker gains control of an AI system set up to operate autonomously. A compromised agent could impersonate its user by sending automated emails, texts or voice messages on their behalf. In the case of smart home integration, it could even interfere with door locks, alarms or security cameras, with a direct impact on personal safety.

Beyond hijacking, adversaries could also poison the data that trains an agentic AI, feeding it biased or hostile inputs designed to warp its outputs. Over time, this could lead to inaccurate, misleading or potentially harmful decisions.

In each of these scenarios, malicious access could result in blackmail, harassment or identity theft. This is a strong illustration of how virtual attacks on agentic AI can quickly create serious real-world consequences.

Mitigating the danger

Organizations that utilize agentic AI systems have a responsibility to operate systems safely with appropriate guardrails. Although the potential threats are significant, their impact can be reduced.

Users must become more data-savvy and avoid complacency, particularly when relying on agentic AI. For highly sensitive information, it may be safer to opt out of using such systems for decision-making altogether.

Where these systems are used, users should review terms and conditions carefully, ensure transparency over what data is being processed, and understand how automated decisions are reached. Prioritizing systems that clearly explain their reasoning helps reduce the risk of hidden data practices and strengthens user control.

A double-edged sword

Agentic AI has the potential to transform industries and streamline everyday tasks. With 96 percent of global technology leaders agreeing that agentic AI innovation, exploration and adoption will continue at lightning speed in 2026, managing the risks will be more important than ever.

These systems must be deployed ethically and safely, with human oversight at every stage. Their development should include transparency protocols and clear processes for explaining automated decisions.

By recognizing the risks and committing to responsible use, we can harness the benefits of agentic AI while safeguarding users.

We've featured the best AI tools currently available.

The time has come, OpenAI has finally announced the thing we all knew was coming but didn't want to believe: ChatGPT is getting ads.

The news hit on Friday last week when the creators of the world's most popular AI chatbot dropped a blog post highlighting an ad-filled future, with testing beginning in the coming weeks for free users and those on the new ChatGPT Go plan in the U.S.

OpenAI shared examples of what sponsored ads will look like in ChatGPT, essentially implementing recommended products depending on the queries you ask the chatbot. In the company's example, it showed a sponsored grocery list below the answer to a dinner-related question.

Now, there's been plenty of furore online, after all, Sam Altman and co's AI product is hardly an internet darling, and the inclusion of ads is seen as another kick in the teeth for users who have invested time and trust into ChatGPT.

And while I hate the fact that ChatGPT will be getting ads, unfortunately, we've been here before, and not only is this the beginning of ad-riddled responses, but it's only a matter of time before chatbots are no longer free: Limited functionality forcing users to pay for actually usable AI capabilities.

Step 1: The Hook

Over the last few years, products like ChatGPT, Gemini, and even the likes of Claude have become household names, offering "magical" tools that can answer almost any question you throw at them.

Not only have these products been made readily available to consumers, but they've essentially been offered at little to no cost, giving users the ability to see exactly what an AI chatbot is capable of.

The goal of offering these AI tools, essentially at a loss (yes, running AI servers is not cheap), is to create some kind of brand loyalty and reliance on the chatbots on offer. OpenAI might've had uproar from its most loyal user base when it launched GPT-5 last year, but that showed that people were becoming attached to the AI model on offer, and that's only after a year or so of reliance - just imagine once people have grown up with their favorite AI chatbot.

Over the last two years, AI has become one of the most contentious topics in the world. From consumer tech to the environment, everyone is talking about artificial intelligence, and the companies at the forefront of the industry advertise every product release like the second coming of Christ.

The thing is, it's all part of the process: By operating at a loss without negative connotations like ads, companies like OpenAI have been able to offer a product that would rarely ever be free, and best of all, they've been able to brand it as a public utility.

In 2026, OpenAI would have you believe that ChatGPT is as essential to its users as the internet itself, and for many, it's become the go-to place to interact with the web. With over 800 million users a week, it's fair to say that OpenAI definitely has people hooked.

Step 2: The Pivot

We're currently experiencing "The Pivot", which is the part of the marketing strategy where companies begin introducing advertising. The best example of a pivot from recent times is when Netflix decided to introduce ads back in 2022, following years and years of offering an ad-free experience aimed at competing against the status quo of cable television.

In the case of ChatGPT, the habit has been formed, and people use the product, so whether or not OpenAI introduces ads really shouldn't have too much of an impact on the company's user base.

Obviously, there will be some users who leave for pastures new, opting for competitors like Gemini or Claude. But just like streaming, those competitors will all get ads someday, especially since their closest competitor decided to bite the bullet and monetize with a new revenue stream.

The goal of bringing ads into ChatGPT is to monetize the massive user base without causing a mass exodus, and by positioning the adverts as "helpful recommendations," they almost pass for a feature, rather than an inconvenience.

When streaming services like Netflix, Prime Video, and Disney Plus started to introduce ads, it worked as a way to push users to more expensive subscriptions. Users will tolerate ads if the entry price remains low, but what if ads are just the beginning?

Step 3: The Squeeze

Ads have been introduced, now what? Well, if we're using the streaming model for comparison, companies like Netflix added extra tiers with higher resolution, extra accounts, and other benefits for a more premium price.

ChatGPT has always had Plus and Pro subscriptions, but now that OpenAI has introduced a new tier called Go, which won't be safe from ads, it's possible, and dare I say it, likely, that restrictions on free AI chatbot accounts will continue to grow.

In the example of Netflix, users who previously were happy paying for the base tier slowly opted for more expensive options to avoid ads and get the same HD resolution offerings they once had at a fraction of the price.

The goal here for companies is to convert users by adding inconveniences that, in the long term, create paying customers. All AI chatbots currently have message caps, AI model limits, and restricted access to newer models. But what if in the future the restrictions are tightened even more?

AI companies are beginning to tighten the belt on usage, because let's face it, they know their current business model is probably no longer viable. The fact is, everyone and their gran now uses AI, and just like streaming services or cloud storage, we're going to reach a point where most people will pay for a chatbot.

ChatGPT's move towards ads is huge for the industry, but Gemini will follow suit shortly, and once that happens, you'll have to make a choice between paying for the "magical" AI Altman and co are promising or opting for an AI-less future. Which option would you choose?

• Grok is now available in business-focused formats with extra security
• $30 per month will get you Grok Business – the same price as ChatGPT Business
• Third-party integration kicks off with direct links to Google Drive

xAI, the Elon Musk-owned platform behind the AI chatbot Grok, has launched new Business and Enterprise plans to give commercial customers access to the latest models, including Grok 3, Grok 4, and Grok 4 Heavy.

One of the key differentiators between these two new business-focused tiers and the consumer-facing Grok model is that Business and Enterprise customers will gain additional security measures. Naturally, and as is the case with other enterprise-grade AI platforms, the models are not trained on these interactions.

In a bid to take custom away from rival platforms like ChatGPT and Gemini, Grok will also integrate with third-party applications such as Google Drive.

Grok Business/Enterprise takes on ChatGPT, Gemini, Copilot

Although Grok won't use company data for training, that proprietary data still informs the chatbot to produce more applicable insights while adhering to third-party settings: "Grok respects your existing Google Drive permissions. If a file isn't shared with you in Drive, you won't see it in Grok."

Musk's chatbot app will also cite sources to help users verify answers.

From an admin point of view, Grok Business teams can be managed from a central console, with usage insights, unified billing and user management dashboards.

Enterprise upgrades include custom single sign-on (SSO), directory sync (SCIM), extra auditing and security controls, and options for app-level encryption with customer-managed keys through 'Enterprise Vault' – an extra.

Pricing for Grok Enterprise is customized to each business use case, however xAI has confirmed that Grok Business will cost $30 per user per month, which matches ChatGPT Business (though annual billing is cheaper). Gemini offers cheaper business-oriented plans at $14 and $21 per user per month, which indicates that Musk is directly targeting rival company OpenAI – one he has a a tumultuous history with.

The consumer-facing Grok remains free to use on a limited basis, with SuperGrok (newer models, extended memory, image model, and more) costing $30 per month ($10 more than the $20-per-month ChatGPT Plus).

If you're overwhelmed by AI chatter, you're not alone. Right now, it feels like AI shows up in every tool, app, headline, and workplace conversation – whether your boss is keen on integrating it into everything or firmly against it. And alongside the hype, there are genuine concerns about its impact on jobs, the way we think, and whether we should be using it at work at all.

If you're hesitant, you're not alone. All of these concerns are real. But it's still true that, used sensibly, AI chatbots can genuinely help people with planning, writing, organizing thoughts, summarizing information, and creating images.

Chatbots are essentially text-based conversation tools. You type, they reply. Like messaging someone who can draft an email, explain a concept, or help you think something through. But it's not a friend or colleague, it's an AI system that's responding.

This guide won't turn you into a power user, drown you in jargon, or dig into every advanced thing AI chatbots can do. Instead, it's a clear, beginner-friendly starting point. What you need to know, how today's main AI chatbots differ, and which one is most likely to suit your first steps with AI.

What beginners need to know

Before you choose an AI chatbot, it helps to understand a few simple definitions. These basics will make it easier to pick the right tool and get better results once you start using it.

Chatbots: You’ll see terms like “AI assistants,” “models,” and “LLMs” (large language models). The important bit is that the chatbot is the interface you talk to, and the LLM is essentially the engine powering it behind the scenes.

Free vs. paid: Most AI chatbots offer a strong free tier. If you pay, the benefits you unlock will depend on the chatbot. But generally, you can expect faster responses, access to newer or more powerful models, better memory, and higher usage limits.

Hallucinations: This is an industry term used to describe the way some AI chatbots make things up. It happens because the models are often trained to come up with a response even when they don't know something. Obviously, this isn't ideal. It means you'll need to double-check a lot of what AI chatbots tell you. Especially if you're using them for research.

Memory: Some chatbots remember previous conversations or your preferences. This means less repeating yourself and a more personalized experience. If you plan to use AI regularly for things like writing, goal-setting, and planning, memory can make a huge difference.

Privacy: Don't paste confidential business information, client data, or personal data into a chatbot. You should also read the tool's privacy settings and your workplace rules. It's also possible to go and turn off sharing settings in most chatbots. Sometimes your conversations are used to make the model smarter, but that does put your privacy at risk.

Prompts: A prompt is simply what you type into an AI chatbot. You can just chat normally, like you would to a friend. But clearer, more deliberate prompts will deliver better outputs. Especially if you're using the chatbot for work. You can read our guide to better AI prompts to find out more.

Search: Each chatbot handles information differently, but it's best not to think of them as a straight replacement for search engines like Google. Yes, many people do use them that way. But because AI chatbots can hallucinate and don't always have access to live, up-to-date information, don't use them for anything factual or time sensitive. If you do, always double-check with a reliable source.

Similarities: Most AI chatbots behave similarly for simple things like proofreading or getting new ideas. So don't overthink it. The differences will start to show up more if you're writing longer text, like reports, want to use an AI chatbot for more complex tasks, like coding, or you're working inside of the Google or Microsoft ecosystems.

The AI chatbots beginners should consider

Best for most people: ChatGPT

If you only try one chatbot, make it this. ChatGPT is the easiest on-ramp for beginners. It has a clear interface, lots of tutorials online, and strong all-around performance for everything beginners would want to try, like writing, planning, questions, and creativity. The free version is also very capable, but the Pro version is the first paid tier and costs $20/£19 (approx) a month. That gets you more messages, more uploads, and more memory.

Best for thoughtful writing: Claude

The company behind Claude, Anthropic, positions itself as a “safety-first” AI company, with a focus on transparent, ethical development. A lot of people are fans of Claude and prefer it over ChatGPT. We've found it's become a favorite for people who write a lot. Many writers also say they prefer its tone and clarity to others in this list. It’s strong at long-form content, analysis, structured documents, and reasoning-heavy tasks – though it does do a lot more than writing. The free version is good, but it doesn't have memory. The Pro version is $17/£18 per month.

Best for Google users: Google Gemini

If you live in Gmail, Docs, Calendar, and Drive, Gemini is a really seamless option. It integrates neatly into Google Workspace and can access your files (with permission) to do all sorts of things, like assist with drafts, summaries, and planning. The free tier is good. The first paid-for tier is called Google AI Pro. It's $19.99/£18.99 a month and gives you a better experience with more usage.

Best for Microsoft users: Microsoft Copilot

Ideal if you’re in Outlook, Word, Teams, or Excel all day. Copilot is designed to fit directly into the Microsoft suite, so it can analyse documents, summarize meetings, help with spreadsheets, and more. Copilot is free to get started with and built-in to Microsoft. Accessing the paid-for version is a bit different from the rest, considering it's bundled into Microsoft. But prices start at $9.99/£9.99 a month.

Best for research: Perplexity

If your priority is finding accurate information, source links, and research, not creative writing, then Perplexity is our top pick. It cites sources by default, compares viewpoints, can fact-check big chunks of text, and gives reliable summaries. You'll get a good free version that doesn't require signing in. Perplexity Pro starts at $20/£16 (approx) a month and delivers a better experience and more usage.

Which AI chatbot do you need?

If you want the simplest answer, try ChatGPT. If you like writing, try Claude too. Use the free versions first; they're good enough to give you a feel for each tool and how they work. If you upgrade, it'll usually be because you want faster responses, better memory, or access to more advanced models.

If you're already deep in Google or Microsoft tools, Gemini or Copilot might be a better fit. And if your priority is reliable research rather than creativity, Perplexity is worth exploring. And remember, you don't have to pick just one.

• Pen Test Partners found flaws in Eurostar’s AI chatbot, including weak validation and HTML injection
• Eurostar says customer data was never at risk; vulnerabilities have since been mitigated
• Palo Alto warns rapid AI adoption expands cloud attack surfaces via misconfigurations and non‑human identities

Eurostar's recently-introduced AI-powered customer support chatbot was marred with cybersecurity vulnerabilities that opened the doors to a multitude of potential risks, experts have warned.

Researchers at Pen Test Partners discovered the chatbot properly validated only the most recent messages in a conversation, meaning older messages could be altered to contain a malicious prompt. That prompt could be virtually anything, from revealing system information, to (possibly) exfiltrating sensitive customer data.

Luckily, Eurostar did not connect its customer information database with the chatbot, so at the time of discovery, there was no direct risk of data leakage happening.

"Customers were never at risk"

The expers found there were other weaknesses in the system, as well, including conversation and message IDs that weren’t properly verified, or an HTML injection flaw that enables running JavaScript directly in the chat window.

Pen Test Partners seem to be the first to have discovered these vulnerabilities: “No attempt was made to access other users’ conversations or personal data”, the researchers explained. “But the same design weaknesses could become far more serious as chatbot functionality expands”.

Eurostar emphasized customer data was never at risk, telling City AM: “The chatbot did not have access to other systems and more importantly no sensitive customer data was at risk. All data is protected by a customer login.”

Many businesses are rushing to deploy AI tools, however, rapid enterprise adoption is significantly expanding cloud attack surfaces and putting businesses at more risk than ever before.

• Microsoft has analyzed how Copilot users interact with AI across 37.5 million conversations
• There are clear differences between work and personal interactions
• This is how app developers can tailor their chatbot UI

A new December 2025 report from Microsoft has uncovered some key difference among Copilot users in terms of how and when people use AI, and it's based on 37.5 million de-identified Copilot conversations, so there's reason to believe the insights should be pretty accurate.

Drawing a line right down the middle, Microsoft identified that desktop AI use was generally dominated by workers between 8am and 5pm, while mobile use skewed heavily toward personal topics, and was used at all hours.

For personal users, health and fitness prevailed as a key topic, which Microsoft says proves AI is increasingly being trusted for advice, not just information gathering.

Copilot gets its own Spotify Wrapped/Apple Music Replay with new key insights revealed

The researchers stressed that enterprise and education data was excluded from the analytics, however millions of personal account holders still use the service for personal and work purposes alike. Microsoft found that programming peaks on weekdays, gaming rises on weekends, and philosophical questions spike late at night.

Redmond claims that mainstream adoption is also on the rise, beyond early adoption among developers and some workers.

In early 2025, Copilot was broadly used for technical and productivity workloads, but by September (the end of this particular study) it was answering more questions about society, culture, and history.

Copilot even sees seasonal trends around the social calendar. For example, relationships and personal growth around Valentine's Day and a dip in education-related topics over the summer.

With the new information, Microsoft has essentially told all AI chatbot developers how they can tailor their UI to usage patterns. "A desktop agent should optimize for information density and workflow execution, while a mobile agent might prioritize empathy, brevity, and personal guidance," the researchers concluded.

"The data suggests that we are not just using AI to do our work faster; we are using it to navigate the complexities of being human, one prompt at a time," they added.

• Jimdo launched Companion, an AI chatbot offering personalized business recommendations for solopreneurs and microbusinesses
• The tool analyzes performance history, benchmarks, and competitive data, with GDPR compliance and Google/Meta integrations
• Early results show 40% more inquiries and transactions, with 50% of businesses generating their first sale within 30 days

Website builder Jimdo just introduced an AI-powered chatbot designed to help solopreneurs and microbusinesses in their day-to-day operations.

In a press release published on November 19, Jimdo introduced Companion, an AI chatbot built deep into the product’s architecture, and designed to deliver personalized and prioritized recommendations. The product can analyze each business’ performance history, industry benchmark, and competitive landscape, to drive visibility, and improve transactions.

That could mean a myriad of things, from suggesting specific local search engine optimization efforts, to actionable steps to improve conversion rates. To get these context-specific answers, users will have to ask strategic questions, and treat Companion as a personalized business strategist, available on-demand.

This also means that Companion will have access to key business data, including information that could be deemed sensitive. Jimdo therefore claims the AI tool is GDPR compliant and adds that it is also connected with both Google and Meta.

(over)promising?

Although the press release does not discuss pricing, or availability, a blog on the company’s website suggests that Companion is now available to all paying users.

There also seems to be an option to test the product out, for free.

Until we test the tool ourselves, or get testimonials from reputable sources, it is difficult to determine how big of an impact Companion can have. Jimdo claims that “early data shows the largest measured impact on their customers’ business outcomes in its 18-year history”, stressing that its users saw 40% more customer inquiries and transactions, while 50% more businesses generated their first sale or inquiry within 30 days after accessing Companion.

Most website builders have already integrated Generative AI into their offering, allowing users to create entire websites, from design to copy, by conversing with software. How effective those websites will be, remains to be seen.

More from TechRadar Pro

• Save on your Wix subscription with our list of Wix promo codes
• Check out our roundup of the best AI website builders on the market
• Learn more about Wix with our full Wix review

While chatbots can be great for some tasks, they can also put your security and privacy at risk in others. If you're new to AI then here's some advice on the do's and don’ts of using chatbots like ChatGPT, Gemini and the others.

For example, using chatbots as a brainstorming tool is a great idea, and so is using them for creating content, but using them to cheat on your homework or accepting whatever they say without question is not a good idea. Chatbots can often make things up, and it's better to double check every fact they present.

Here are the top 10 do's and don't when using a chatbot.

1. DO: Ask AI to help you brainstorm a decision

If you reached a crossroads in your life and you need to make a decision, whether that’s about buying a new car or moving house, then describe the situation and get ChatGPT to come up with a handy list of pros and cons.

While you don't need to get AI to make big decisions for you, it can be very good at laying out the different sides to a decision that you need to make, and that will help you see the big picture.

2. DON'T: Use AI to cheat on your homework

Children using ChatGPT to come up with answers to homework questions is not a good idea. Firstly, it’s cheating, and secondly, it can be very obvious to the teachers exactly who is using ChatGPT, and getting caught can have serious consequences.

The same goes for students in college or university, although the consequences for getting caught could be even higher, perhaps leading to being failed on coursework.

Using ChatGPT in Study Mode as a tutor to help you learn however is a great idea.

3. DO: Use it for proofreading

Chatbots like ChatGPT and Gemini are excellent at proofreading your written work. Either cut and paste your written work, or export your document from something like Google Drive as a Word document and upload it into ChatGPT, then ask, “Can you proofread this, please?”

If the text is short, then the AI will simply present you with a more polished version of the document. If it is long then it will discuss stylistic points as well as the structure of the document. In both cases it can provide you with a Word document that has tracked changes in, or just make the changes it suggests.

4. DON'T: Believe everything AI says without checking

Large Language Models like ChatGPT want to be helpful. In fact, they often want to be too helpful, and that leads them to simply make things up. It happens so often that it’s got its own name – AI hallucinations. AI chatbots can hallucinate facts out of nowhere. They can hallucinate scientific studies, or law cases, or simply get math problems wrong.

Hallucination rates on OpenAI's most up to date model, ChatGPT-5 are around 1.4% (according to Vectara), which is low, but it still means that you need to double-check anything it’s telling you is true, just to make sure.

5. DO: Use AI to learn new things

Having access to an AI like Gemini or ChatGPT is like having a personal tutor on your mobile phone or laptop that can teach you just about anything you’d like to learn. The world is your oyster, so pick a subject and start to learn.

If you like, you can get ChatGPT to create a program designed to teach you something over a long period of time – just ask it. It’s not so good at teaching physical skills, although it will point you to relevant YouTube videos that demonstrate how to do something, especially if you ask it to.

6. DON'T Let kids use AI on their own

According to OpenAI: “ChatGPT is not meant for children under 13, and we require that children ages 13 to 18 obtain parental consent before using ChatGPT”.

If you’re using an AI chatbot with a child under 13 then you, the adult, should be the person interacting with it, and we’d still recommend keeping an eye on children over 13 who are using it.

So, if you want to use ChatGPT to create a bedtime story for your kids, or let them play a game with it, then that’s fine, so long as you are the person interacting with the chatbot.

7. DO: Use AI to generate code

One of the strengths of AI is that it can write code - or you can upload your existing code and ask it to help finish it. This works particularly well in Agent mode in ChatGPT, where you can ask it to code an app in the background, because it can take a while.

This is useful when you need something very specific that isn’t available elsewhere, or you just fancy getting it to create a Space Invaders-style game. Chatbots can also write code for more professional situations, but it’s worth getting an expert to check the accuracy since it can make mistakes.

8. DON'T: Give AI your credit card number

Beyond entering your payment details when you sign up for a Plus or Pro account, we would never recommend that you enter your credit card details into an actual chatbot chat.

It is highly unlikely that an AI chatbot will ever ask you for your credit card details inside a chat, but as a matter of course, you should never offer them. As AI evolves so too will security threats that take advantage of it, and our acceptance of it, and it’s quite possible that some sort of malicious code could hack into an AI chatbot or even imitate it.

9. DO: Use AI to play fun games

Chatbots are great for playing games. Simply ask them to play a game of tik-tak-toe to get started. In fact, it can play most simple board games like chess, checkers, connect four, battleship and trivia games, which are particularly fun.

You can even get AI to create a roleplaying text adventure game in the style of those old choose-your-own-adventure books. Using the GPTs menu (in the left hand menu bar) you can search for a GPT for a particular game you want to play, like Dungeons and Dragons, for example.

10. DON'T: Rely on it for medical advice

If you want to research a particular condition, or get it to explain what doctors are telling you in a simple way that you can understand, then AI can be very useful. However, it should go without saying that AI is not a qualified doctor.

AI does a great job of explaining medical terms, however, that doesn’t mean it is qualified to diagnose you given a list of your symptoms. Again, the general advice to treat ChatGPT and the other chatbots as a great way to brainstorm ideas and talk about subjects holds true, but remember that you should always seek the opinion of a medical professional for medical matters.

Version drift is the proliferation of unofficial copies of documents that happens remarkably quickly through common workplace behaviors. Each time someone uses the "Save As" function, they create a copy that will never be updated again, instantly creating an obsolete version.

Sending files as email attachments instead of sharing links creates additional copies that diverge from the original. Downloading files from corporate intranets or official sources creates local copies that sync to OneDrive and SharePoint, further multiplying document versions.

Without discipline, corporate documents become cloned many times, and these rogue copies are never cleaned up. When AI systems scan enterprise systems, they discover all of these versions and treat them as equally valid sources.

The problem manifests when artificial intelligence systems confidently deliver accurate but outdated information to users—unlike AI hallucination, where models fabricate entirely fictional facts, version drift involves real data that has simply become obsolete or superseded by newer versions.

A high-profile example occurred when Air Canada's website chatbot told a grieving passenger he could retroactively claim a bereavement discount based on outdated policy information. A tribunal ruled the airline liable and ordered reimbursement, demonstrating the legal and financial consequences of version drift.

Version drift is one of the most dangerous failure modes in enterprise AI deployment because the information provided is technically correct but operationally wrong, making it difficult to detect and potentially more damaging than obvious fabrications.

Why AI Systems Can't Tell the Difference Between Document Versions

Retrieval-augmented generation (RAG) systems use semantic matching to find relevant documents.

When multiple versions of the same information exist—such as pricing schedules marked "DRAFT 2019," "FINAL 2023," and "APPROVED 2025"—these systems analyze the semantic content rather than metadata.

From a semantic matching perspective, all copies of a document have essentially the same semantic score because they contain similar information about the same topic.

This means AI systems see all versions as equally valid to answer a user's question, regardless of approval status, creation date, or authority level.

When faced with multiple semantically similar documents, these systems essentially make random selections among available options without considering currency or approval status.

This creates a fundamental mismatch between how humans and AI systems approach information validation.

When humans encounter multiple search results, they automatically apply contextual reasoning: scanning file names, checking creation dates, identifying folder structures, and assessing approval status.

Humans naturally filter information as they think: "Which version is most current? Who approved this document? Does this appear to be a draft or final version?"

Most enterprise AI implementations skip this evaluation step entirely, leading to confident responses based on random selection among technically valid but potentially obsolete sources.

The problem compounds in regulated industries where using outdated procedures for dangerous industrial machines or expired maintenance manuals could create serious safety and liability risks.

Teaching AI to Evaluate Before Selecting

Chain-of-thought reasoning solves version drift by requiring AI systems to follow a structured three-step process:

- Retrieval: Gather all relevant documents from available sources

- Evaluation: Analyze metadata including creation dates, approval status, version tags, and authority markers

- Selection: Choose the most appropriate source and provide clear rationale for the decision

Modern frameworks like LangChain's Graph agents and LlamaIndex's ReasoningNodes support chain-of-thought patterns, while OpenAI's reasoning models (o1 and o3) represent advanced implementations.

How a Sales Rep Gets the Right Pricing Information

Consider a sales representative requesting pricing information for Product X. A traditional RAG system might retrieve five pricing spreadsheets and randomly select one, potentially delivering outdated rates with complete confidence.

A chain-of-thought enabled system examines the metadata: "I found pricing documents from Q3 2024, Q4 2024, and Q1 2025.

Three are marked 'FINAL' and two are marked 'DRAFT.' I'll use the most recent validated version (Q1 2025 FINAL) and note that this supersedes earlier versions."

The evaluation layer addresses the enterprise challenge of multiple valid documents with varying currency and authority levels.

Your Documents Need Metadata Before AI Can Reason

Version drift protection requires metadata infrastructure. Document validation status, approval workflows, and version control systems must be queryable by AI systems before reasoning capabilities can deliver value.

Organizations should begin by conducting an "Ambiguity Stress-Test": selecting ten common employee queries and manually verifying whether multiple valid but conflicting documents exist for each scenario.

The Business Case for Reasoning-Capable AI

Version drift represents a maturation challenge as organizations move beyond simple chatbot implementations toward mission-critical AI systems.

The investment in reasoning-capable systems far outweighs the potential costs of compliance violations, pricing errors, and damaged employee trust that result from confident but outdated AI responses.

The next generation of enterprise AI will be distinguished by reasoning capability leveraging metadata and document lifecycle status rather than processing speed or scale—systems that understand not just what information exists, but which information should be trusted.

Why Employees Need to Trust AI Source Selection

Version drift protection goes beyond operational efficiency into risk management. Organizations implementing chain-of-thought reasoning protect against compliance violations, pricing errors, and reputational damage from confident but incorrect AI responses.

Without trust in AI responses, organizations are forced to limit AI to non-critical tasks, preventing meaningful impact on core business operations. When employees can't rely on AI accuracy, these systems remain relegated to low-stakes applications rather than transforming mission-critical workflows where they could deliver substantial value.

Transparency matters equally. When AI systems explain their source selection reasoning—"I selected this document because it's the most recent approved version"—they become trustworthy tools that employees can confidently rely upon rather than opaque black boxes.

In regulated industries, auditable AI decision-making becomes required rather than optional.

We list the best document management software.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

• ChatGPT usage jumped to 78.3 billion tokens as schools returned in September 2025
• OpenRouter data shows steep contrast between academic year peaks and summer lulls
• Students remain a major driver of daily ChatGPT traffic, studies confirm

In news which likely won’t come as a huge surprise to anyone, ChatGPT has seen a sharp rise in usage as schools reopen in the West, with token generation climbing to record levels.

New data from OpenRouter claimed that on September 18, OpenAI’s hugely popular chatbot recorded 78.3 billion tokens, the highest since the summer slowdown.

The surge comes after months of lower activity during school holidays - back in June 2025, when many schools let out, daily usage fell to an average of 36.7 billion tokens, in comparison to May 2025, which averaged nearly 80 billion tokens per day, coinciding with exams and finals.

Shifting patterns of AI use

OpenRouter’s figures, which track usage across its 2.5 million users, provide a detailed view of how patterns shift throughout the year.

Although the data only represents one platform, it has become a widely cited resource for researchers and investors who are analyzing adoption of LLMs.

The numbers make it clear that students are behind a large chunk of ChatGPT’s daily activity.

As Futurism reports, previous studies, including one from Rutgers University, found strong links between academic calendars and usage spikes. Looking at OpenRouter’s interactive graphic, that connection seems pretty undeniable.

Breaks such as spring and summer consistently align with drops in interaction, reinforcing earlier observations that ChatGPT traffic drops during holidays and recovers with the school year.

Among the models measured by OpenRouter, ChatGPT 4.1 Mini stood ahead of the rest, generating 26.9 billion tokens on September 18, ahead of the newly launched GPT-5, which accounted for 18.7 billion tokens on the same day, with other models such as GPT-4o mini and GPT-5 Mini also contributing notable shares.

Looking at the data, it's clear AI tools are increasingly being used in educational establishments with students relying on OpenAI’s chatbot for writing, information gathering, and study support.

It’s not necessarily a bad thing of course, and many teachers and researchers will see value in students learning how to work with these systems responsibly.

As with any new technology, the younger generations are inevitably going to be among the first to fully embrace AI as part of their daily lives.

The debate is not whether students should be using AI, but how it can be guided, taught, and balanced in ways that support learning rather than replacing it.

You might also like

• These are the best AI image generator tools available
• And these are the best AI writers you can use right now
• Are you polite to ChatGPT? Here’s where you rank among AI chatbot users

Think back to when chat-based assistants like Google Assistant or Siri first appeared. These early tools promised convenience, letting us type or speak a question and receive a quick digital reply.

Yet, their capabilities were often basic. If the answer wasn’t in their database, you’d be prompted to look elsewhere. Their usefulness hinted at the future, but the real breakthrough was still to come.

Generative AI changed the game by filling in many of those old gaps. Now, you can ask pretty much anything, in whatever way makes sense to you, and get answers that are tailored and easy to tweak, completely transforming the way we find information.

It wasn’t long before technology leaders saw the business potential: imagine connecting a chatbot to company data to create a kind of digital copilot that could answer business questions in real time.

No need to jump through menus or juggle dashboards, just talk to the system and get the info you need, right when you need it.

But here’s the thing: that vision isn’t reality yet. Recent research shows that chatbots haven’t had a big impact on productivity or the bottom line, mainly because business processes need more than just a chat interface: they need structure, reliability, and context.

The idea of replacing traditional software with chat is tempting, but it misses a much bigger opportunity: putting AI right where the real work happens.

Reducing friction and moving past reactivity

One big challenge for chat-based AI in business is something called the “blank canvas” problem. Chatbots are meant to be open-ended: you can ask anything, and they’ll do their best to respond. That’s great for consumers, but in a business setting, it can slow things down.

When chatbots don’t have a definitive answer to an open-ended question, they may offer misleading or incorrect information, and knowing this, workers need to spend extra time verifying outputs against trusted sources.

Picture a salesperson juggling lots of clients and prospects. Instead of leaving their email to ask a chatbot about an account, refining their question, reviewing the result and then returning after reading the answer, what if the company’s AI was built right into their existing tools? That way, insights pop up instantly, no detours needed. In this scenario, chatbots just add extra steps.

All too often, chatbots operate outside the main systems that employees use every day. If a manager wants to check on a sales pipeline, for example, they might have to bounce between a CRM and a separate window – ask a question, switch apps, copy, paste, repeat. Not only does this eat up time, but it also increases the chance of mistakes and keeps important knowledge scattered across different tools.

There’s also a key difference between reactive and proactive AI. Chatbots sit back and wait for someone to ask them something. But in the fast-moving worlds of sales, finance, or customer support, no one can afford to wait for the right question. Teams need to know if a deal is in trouble or a forecast is off, before anyone even thinks to ask.

Embedded AI doesn’t just sit around. It spots issues as they happen, brings them right into the tools people are already using and suggests next steps without prompting. Then, it’s not just answering questions, it’s actively influencing outcomes.

Consistency is crucial

Another challenge with chat-first interfaces is getting consistent answers. Give five employees the same task, and if they all use a chatbot, they’ll probably get five different answers. Some may use detailed prompts, others might be more general. That kind of inconsistency just doesn’t work for teams that rely on shared workflows and consistent results.

When AI is built into business tools, it helps everyone stay on the same page. It makes sure all employees follow the same best practices, right from inside the platforms they already use, like Microsoft Teams or Word. After a call, for instance, the system can immediately offer up next steps that match the organization's goals and policies.

Bringing AI into the workflow is the way forward. When intelligence is built into a sales platform, for example, it can automatically spot red flags, suggest actions, and update forecasts in a clear, consistent way. No perfect prompts needed, it just works.

There’s also a limit to relying only on text chat. Business decisions usually need complex data that lives in lots of different places. A chatbot can summarize what it’s given, but it can’t match the interactive experience of a dashboard that shows regional numbers, customer segments, or makes real-time predictions.

Business professionals need more than just a summary. They need insights that are clear, relevant and actionable. Knowing what’s going on is helpful, but when the pressure is on, people need to know exactly what steps to take to get the best results.

Adaptability > eloquence

Chat interfaces are often impressive in their fluency, but business leaders shouldn’t confuse how something looks and feels with the value it brings. Chatbots will keep being popular with consumers, but the real win for enterprise AI is in how it helps teams make better decisions and reach better outcomes.

When you embed intelligence into the platforms employees already use – whether that’s CRMs, messaging apps or other AI copilots – it makes it easier for everyone to access trustworthy information as part of their daily workflows.

That means automating tedious research and serving up the right business insights at the right times to help workers make the best decisions. In turn, employees can focus on connecting with customers, equipped with the guidance they need to truly succeed.

We list the best employee management software.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

• Salesloft was breached when OAuth tokens from SalesDrift were stolen
• Google tracked the threat actors as UNC6395
• ShinyHunters claimed responsibility for the attack

Revenue workflow platform Salesloft suffered a cyberattack which saw threat actors break in through a third-party and steal sensitive information.

The company is using Drift, a conversational marketing and sales platform that uses live chat, chatbots, and AI, to engage visitors in real time, alongside its own SalesDrift, a third-party platform which links Drift’s AI chat functionality to Salesforce, syncing conversations, leads, and cases, into the CRM via the Salesloft ecosystem.

Starting around August 8, and lasting for about ten days, adversaries managed to steal OAuth and refresh tokens from SalesDrift, pivoting to customer environments, and successfully exfiltrating sensitive data.

Attack attribution

"Initial findings have shown that the actor's primary objective was to steal credentials, specifically focusing on sensitive information like AWS access keys, passwords, and Snowflake-related access tokens," Salesloft said in an advisory.

"We have determined that this incident did not impact customers who do not use our Drift-Salesforce integration. Based on our ongoing investigation, we do not see evidence of ongoing malicious activity related to this incident."

In its write-up, Google’s Threat Intelligence Group (GTIG) said the attack was conducted by a threat actor known as UNC6395.

“After the data was exfiltrated, the actor searched through the data to look for secrets that could be potentially used to compromise victim environments,” the researchers said.

“GTIG observed UNC6395 targeting sensitive credentials such as Amazon Web Services (AWS) access keys (AKIA), passwords, and Snowflake-related access tokens. UNC6395 demonstrated operational security awareness by deleting query jobs, however logs were not impacted and organizations should still review relevant logs for evidence of data exposure.”

Google seems to believe this is a unique threat actor, which is why it gave it a unique moniker UNC6395.

However, hackers known as ShinyHunters told BleepingComputer the attack was actually their doing - although Google begs to differ, telling the site, "We've not seen any compelling evidence connecting them at this time."

You might also like

• A flaw in Google OAuth system is exposing millions of users via abandoned accounts
• Take a look at our guide to the best authenticator app
• We've rounded up the best password managers

GenAI tools such as ChatGPT, Gemini, and Copilot have become essential components of modern workflows, significantly saving countless hours and revolutionizing various tasks. 42% of enterprises actively deployed AI, and 40% are experimenting with it and 59% of those using or exploring AI have accelerated their investments over the past two years.

Their widespread adoption across industries has demonstrably boosted efficiency and productivity, making them indispensable for many organizations across almost all industries.

However, the rapid integration and reliance on GenAI tools have inadvertently fostered a dangerous sense of complacency within organizations.

While these tools are easy to use and offer widespread benefits, ignoring the consequences of misuse and even malicious use has led to a serious underestimation of the inherent risks tied to their deployment and management, creating fertile ground for potential vulnerabilities.

When Innovation Hides Exposure

While typical users may not consider the vulnerabilities that GenAI tools bring, many CISOs and AI leaders are increasingly concerned about the misuse that’s unfolding quietly beneath the surface.

What often appears to be innovation and efficiency can, in reality, mask significant security blind spots. By 2027, it is estimated that over 40% of breaches will originate from the improper cross-border use of GenAI. For CISOs, this isn't a distant concern but an urgent and growing risk that demands immediate attention and action.

The exploitation of everyday AI users isn’t just a scary headline or a cautionary tale from IT—it’s a rapidly growing reality. These emerging attacks are sweeping across industries, catching many off guard. Just recently, researchers disclosed a Microsoft Copilot vulnerability that could have enabled sensitive data exfiltration via prompt injection attacks.

The ongoing underestimation of basic AI usage risks within organizations is a key driver of this emerging danger. The lack of awareness and robust policies surrounding the secure deployment and ongoing management of GenAI tools is creating critical blind spots that malicious actors are increasingly exploiting.

A New Security Mindset

The evolving landscape of GenAI presents a critical inflection point for cybersecurity leaders. It's imperative that CISOs and industry professionals move beyond the initial excitement and acknowledge that these tools have inherent risks that have been introduced by the widespread adoption of these powerful tools.

The current situation, marked by rapid integration and security oversight mixed with dangerous complacency, demands a fundamental shift in how organizations perceive and manage their digital defenses especially with AI.

The future of network security hinges on intelligent, comprehensive monitoring systems capable of understanding normal behavioral patterns and rapidly identifying deviations. This approach is not only crucial but paramount for detecting sophisticated threats that bypass traditional defenses.

Tools that can defend and protect against highly sophisticated threats need to include advanced capabilities at their core. Particularly, when considering scenarios where seemingly innocuous actions, like using a basic GenAI chatbot could lead to the silent exfiltration of sensitive corporate data, without user interaction or explicit warnings.

In these instances, traditional signature-based detection methods would likely prove ineffective. Therefore, it's imperative to begin leveraging advanced pattern recognition and behavioral analysis to combat threats specifically designed to evolve and evade detection.

Trust in AI Starts from Within

With the rise of increasingly sophisticated threats pressing closer to the enterprise perimeter, organizations must take decisive and actionable steps. This begins with addressing internal distrust of AI. Roughly three-quarters of AI experts think the technology will benefit them personally, however, only a quarter of the public says the same.

Fostering an environment where employees understand both the advantages and the risks associated with its use is essential to bridging this gap in perception. The promotion of responsible usage across the organization lays the groundwork for a more secure adoption of GenAI technologies.

While traditional human error remains a threat, the widespread adoption of GenAI has created a new, more subtle class of behavioral risks. Equipping employees with the knowledge to use GenAI tools securely is essential and should include comprehensive training, setting clear usage guidelines, and implementing robust policies tailored to defend against AI-driven attack vectors.

As the AI landscape adapts and changes, security frameworks must be continuously updated to keep pace with these evolving threats and to ensure appropriate safeguards are in place.

Real Security Starts with Behavior Change

Despite technological advancements, attackers continue to exploit human error. Today’s most significant data exposure isn't necessarily from a phishing link, while still a prime point of entry for threat actors; it's from an employee pasting proprietary source code, draft financial reports, or sensitive customer data into a public AI chatbot to work more efficiently.

In turn, companies must adopt strategies that address human behavior and decision-making. In an attempt to boost productivity they inadvertently externalize intellectual property.

This requires companies to evolve their approach beyond periodic training. It demands continuous engagement focused on GenAI-specific scenarios: teaching employees to recognize the difference between a safe, internal AI sandbox and a public tool.

It means creating a culture where asking "Can I put this data in this AI?" becomes as instinctual as locking your computer screen. Employees must be equipped to understand these new risks and feel accountable for using AI responsibly.

Demonizing AI usage, even basic use will never solve the problem at hand. Instead, embracing a secure approach to GenAI from a holistic point of view empowers employees to leverage these powerful tools with confidence to maximize their operational advantages while minimizing exposure to risk.

By leading with clear guidance, highlighting potential warning signs and operational risks, organizations can significantly reduce the chances of data breaches related to improper AI usage, ultimately protecting critical assets and preserving organizational integrity.

We feature the best firewall for small business.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

While AI tools continue to reshape knowledge work to automate emails, summarize meetings, and generate content, a much larger workforce remains stuck in the margins of digital transformation.

Despite accounting for nearly 80% of global employment, frontline workers in industries such as retail, healthcare, hospitality, and manufacturing still rely on outdated systems, fragmented workflows, and manual processes to accomplish their tasks.

But a new era is dawning: one where AI doesn’t just assist, but acts. Agentic AI is poised to transform the nature of work, especially for the often-overlooked frontline workforce.

Beyond answers: How Agentic AI is transforming frontline work?

AI Agent isn’t just a smarter chatbot. It’s an AI-driven assistant that can autonomously sense, decide, and act. It doesn’t wait to be prompted; it operates with intent. It works across systems, identifies what needs to be done, and orchestrates the necessary steps to get there. In short, it moves from being reactive to being proactive.

Why the frontline needs AI agents the most

Frontline businesses are under immense and increasing pressure from labor shortages, rising costs, and unpredictable demand. At the same time, employee expectations are shifting. Workers want more autonomy, faster access to support, and systems that work with them, not against them.

Agentic AI supports both goals. It enables businesses to respond faster while empowering employees to perform their jobs without constant oversight or red tape. Agentic AI orchestrates action across workflows—unlocking its full potential when organizations consolidate point solutions like task management, communications, training, and workforce management into a single platform. It allows the frontline to close the loop between insight and action, something that’s nearly impossible with fragmented tools.

For example, if a frontline employee misses a required task, such as a store audit or safety checklist, agentic AI doesn’t just log the failure or notify a manager. It can diagnose the issue, determine the likely root cause, push targeted training, reassign responsibilities, and follow up, all autonomously.

The three stages of the agentic AI evolution

We’re in the midst of a shift from passive to active AI. It’s helpful to think about this evolution in three stages:

1.Insight-Based AI: AI retrieves information, answers questions, and summarizes content.

2.Agentic AI Breakthrough (Today’s frontier): AI coordinates real-time action and closes the loop, while keeping a human in the know.

3.AI as an Architect (Coming soon): Multi-agent systems that self-optimize with minimal oversight.

Most companies are still in Stage 1. Moving to Stage 2 — where AI helps do the work, not just describe it, and doesn’t require a total overhaul.

Why now?

So why is agentic AI possible today when it wasn’t before?

Three factors have converged:

Model maturity: Modern AI models can handle context and nuance far better than even a year ago.

Platform readiness: The infrastructure to connect scheduling, task management, learning, and communication is more flexible and integrated.

Data availability: There’s finally enough structured frontline data to support intelligent automation.

Common use cases emerging now

Agentic AI is already at work in several practical areas:

Training orchestration: Automatically assigning micro-learnings after incomplete tasks or failed audits.

Staffing and scheduling: Filling gaps due to callouts, shift changes, or compliance constraints.

Task triage: Routing time-sensitive tasks to the right team based on skills, availability, or performance.

Real-time alerts: Notifying staff and triggering actions when thresholds are crossed (e.g., safety violations, inventory issues).

Ethics, privacy, and compliance: a necessary foundation

As agentic AI becomes more embedded in frontline operations, organizations must address the ethical and regulatory implications head-on. These systems interact with sensitive employee data, make real-time decisions that affect workloads and outcomes, and often operate in environments governed by strict labor and privacy laws.

Key areas of focus include:

Transparency: Workers and managers need to understand how AI-driven decisions are made.

Consent and control: Employees should know when AI is acting on their behalf and have access to opt-in/opt-out mechanisms where appropriate.

Bias mitigation: AI must be continuously audited to ensure it doesn't reinforce systemic biases in scheduling, performance tracking, or access to learning.

Data governance: The use of AI must comply with regional data privacy regulations, such as the Data Protection Act and the UK General Data Protection Regulation (GDPR), and labor protections, ensuring that sensitive information is handled responsibly.

Human oversight: Even as autonomy increases, there should always be a path for human review, especially when decisions affect compliance, pay, or job status.

Agentic AI can be a force for good, reducing friction, improving fairness, and empowering workers, but only if organizations treat ethics and compliance not as checkboxes, but as design principles.

Barriers to adoption

If the potential is so clear, what’s holding companies back?

Siloed systems: Without integration between task, learning, communication, and workforce tools, AI can’t orchestrate effectively.

Limited imagination: Many still see AI as a tool for the back office, not a frontline operator.

Trust issues: Frontline workers won’t adopt AI tools they don’t understand or benefit from.

Success starts with solving real problems for frontline teams, not just introducing AI for its own sake. Workers are far more likely to trust and adopt tools that save time, simplify their daily tasks, or help them achieve success.

The path to autonomy starts small

Getting to full autonomy takes time. It starts with identifying high-friction moments where AI can orchestrate a series of actions across tools and teams, and at best, a single, high-impact use case. Look for areas where:

- Data already exists

- Decisions are rules-based or repeatable

- Delays cause operational friction

From there, build AI agents that are narrow, purposeful, and measurable. Over time, these agents can evolve into more complex systems that self-coordinate and adapt to real-world feedback.

The future: keeping humans in the loop

Today’s agentic AI still needs human oversight. But as trust, performance, and integration improve, we’ll move toward a model where humans supervise, refine, and optimize AI systems rather than managing every decision.

In this future, frontline operations won’t just be reactive. They’ll be adaptive, automatically adjusting to new data, new demands, and new disruptions without waiting for a chain of approvals or a lag in communication.

Agentic AI is not about replacing the human workforce. It’s about augmenting it, turning AI from a passive tool into a proactive partner. For organizations willing to rethink how work gets done, the payoff is significant: faster decisions, better outcomes, and a frontline workforce that’s finally supported by systems as smart and dynamic as they are.

We list the best employee management software.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

Behind the polished responses of AI platforms lies a lesser-known truth: workers from emerging economies, such as Africa, Latin America, and Asia, were paid less than $2 per hour to sift through graphic and traumatic content to help train their safety systems. This labor practice has sparked global concern, lawsuits, and calls for ethical reform in the AI industry.

Artificial intelligence is the crown jewel of modern enterprise – a sector exceeding $500 billion, reshaping everything from banking to healthcare. However, the truth is that behind every chatbot, image generator, and recommendation engine are armies of human workers who perform tasks that AI can’t handle, including labeling data, filtering toxic content, and correcting machine errors.

Without them, the algorithms would collapse, and the irony is hard to miss. AI is at risk of becoming the digital frontier for labor malpractice and a new form of unethical conduct. If businesses and innovators don’t act, AI’s promise could unravel under the weight of its own contradictions.

The Invisible Labor Fueling AI’s Rise

It is tempting to believe that AI systems are self-sufficient, refining themselves through endless feedback loops of data and computation. The reality, however, is far more complex. AI systems don’t clean or train themselves. The scale of this hidden labor crisis is staggering. Major gig platforms employ millions to annotate data, correct model errors, and sift through violent or explicit content.

These gig workers are outsourced from countries in the Global South, such as Kenya, India, and the Philippines, who prop up the $8 billion industry that powers the AI revolution. These workers are often highly educated but take on these jobs because better opportunities are scarce. They sign up believing they will contribute to cutting-edge technology, only to find themselves trapped in digital piecework. Pay is low, mental health support is rare, and job security is virtually nonexistent.

Why haven’t businesses fixed this? Because it’s cheap and easy to ignore. However, it comes with growing risks. Consumers and regulators are already beginning to question the ethics of AI supply chains. The European Union’s AI Act and similar efforts globally are setting new expectations for transparency, fairness, and accountability. Companies that fail to address the human cost of AI could face reputational damage, regulatory fines, or worse – a collapse of trust in the systems they have built.

Web3 might be the overlooked fix AI desperately needs

The promise of Web3 – decentralization, transparency, and user empowerment - directly addresses many of the failings in AI’s hidden labor ecosystem. Yet these tools remain largely untapped by enterprise AI, which is clearly a missed opportunity.

Decentralized Autonomous Organizations (DAOs) offer a way to embed genuine transparency and fairness into AI’s supply chains. Unlike traditional gig platforms, where decisions about pay, task selection, or working conditions are made behind closed doors, DAOs make every decision transparent and visible.

Every vote cast, every rule change, and every payment to a contributor is stored on a public ledger, creating an auditable trail that cannot be altered after the fact. This means that anyone, from participants to external auditors, can trace who made the decisions on ‘what, when, and how’. Immutable payment records eliminate the disputes that plague opaque gig work, while public governance logs ensure that power isn’t concentrated in the hands of a few.

Real-world examples are beginning to show what’s possible. Some decentralized employment platforms enable independent workers to collectively manage their pay structures and benefits, with all transactions and decisions recorded on-chain for complete transparency.

Others apply similar principles to research and contributor projects, where rules around compensation and project selection are codified in smart contracts, leaving little room for hidden decisions or unfair practices.

These models exist and are effective, but the reality is that enterprise AI has shown little interest in adopting them so far.

The Limits and Urgency of Change

Many enterprise AI leaders cling to the idea that ethical supply chains are simply too expensive - an unfortunate cost that doesn’t fit the margins demanded by investors or customers. But this is a myth that Web3 technologies can finally dismantle.

Web3’s value isn’t limited to ethics; it offers efficiency gains that traditional systems struggle to match. Smart contracts automate payments and bonuses, reducing the need for large administrative teams and eliminating intermediaries that add cost without providing value. Immutable blockchain records mean payment disputes, task verifications, and contract enforcement happen with far less friction, saving time, legal costs, and operational headaches.

However, Web3 isn’t flawless. Decentralized systems can replicate biases if data or governance isn’t audited. Transparency alone doesn’t guarantee explainability in AI decisions. And DAOs risk elitism if influence skews toward a wealthy few.

Clearly, the real risk is in doing nothing. The companies that lead on ethical AI supply chains will not only avoid the coming backlash but also earn the trust of their customers, regulators, and employees. Those who continue to look the other way will eventually find that the cost of cleaning up the mess is far higher than the cost of reforming now.

Web3 offers the clearest path to cleaning up AI’s hidden mess. But the window for voluntary reform is closing fast. Enterprises can either lead this change or be dragged into it when the backlash hits.

The choice won’t stay theirs for long.

We've listed the best employee management software and the best HR software.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

Imagine a health plan member interacting with their insurer’s virtual assistant, typing, “I just lost my mom and feel overwhelmed.” A conventional chatbot might respond with a perfunctory “I’m sorry to hear that” and send a list of FAQs. This might be why 59% of chatbot users before 2020 felt that “the technologies have misunderstood the nuances of human dialogue.”

In contrast, an AI agent can pause, offer empathetic condolences, gently guide the member to relevant resources, and even help schedule an appointment with their doctor. This empathy, paired with personalization, drives better outcomes.

When people feel understood, they’re more likely to engage, follow through, and trust the system guiding them. Oftentimes in regulated industries that handle sensitive topics, simple task automation fails when users abandon engagements that feel rigid, incompetent, or lack understanding of the individual’s circumstances.

AI agents can listen, understand, and respond with compassion. This combination of contextual awareness and sentiment‑driven response is more than just a nice‑to‑have add-on—it’s foundational for building trust, maintaining engagement, and ensuring members navigating difficult moments get the personalized support they need.

Beyond Automation: Why Empathy Matters in Complex Conversations

Traditional automation excels at straightforward, rule‑based tasks but struggles when conversations turn sensitive. AI agents, by contrast, can detect emotional cues—analyzing tone, punctuation, word choice, conversation history, and more—and deliver supportive, context‑appropriate guidance.

This shift from transactional to relational interactions matters in regulated industries, where people may need help navigating housing assistance, substance-use treatment, or reproductive health concerns.

AI agents that are context-aware and emotionally intelligent can support these conversations by remaining neutral, non‑judgmental, and attuned to the user’s needs.

They also offer a level of accuracy and consistency that’s hard to match—helping ensure members receive timely, personalized guidance and reliable access to resources, which could lead to better, more trusted outcomes.

The Technology Under the Hood

Recent advances in large language models (LLMs) and transformer architectures (GPT‑style models) have been pivotal to enabling more natural, emotionally aware conversations between AI agents and users. Unlike early sentiment analysis tools that only classified text as positive or negative, modern LLMs predict word sequences across entire dialogues, effectively learning the subtleties of human expression.

Consider a scenario where a user types, “I just got laid off and need to talk to someone about my coverage.” An early-generation chatbot might respond with “I can help you with your benefits,” ignoring the user’s distress.

Today’s emotionally intelligent AI agent first acknowledges the emotional weight: “I’m sorry to hear that—losing a job can be really tough.” It then transitions into assistance: “Let’s review your coverage options together, and I can help you schedule a call if you'd like to speak with someone directly."

These advances bring two key strengths. First, contextual awareness means AI agents can track conversation history—remembering what a user mentioned in an earlier exchange and following up appropriately.

Second, built‑in sentiment sensitivity allows these models to move beyond simple positive versus negative tagging. By learning emotional patterns from real‑world conversations, these AI agents can recognize shifts in tone and tailor responses to match the user’s emotional state.

Ethically responsible online platforms embed a robust framework of guardrails to ensure safe, compliant, and trustworthy AI interactions. In regulated environments, this includes proactive content filtering, privacy protections, and strict boundaries that prevent AI from offering unauthorized advice.

Sensitive topics are handled with predefined responses and escalated to human professionals when needed. These safeguards mitigate risk, reinforce user trust, and ensure automation remains accountable, ethical, and aligned with regulatory standards.

Navigating Challenges in Regulated Environments

For people to trust AI in regulated sectors, AI must do more than sound empathetic. It must be transparent, respect user boundaries, and know when to escalate to live experts. Robust safety layers mitigate risk and reinforce trust.

Empathy Subjectivity

Tone, cultural norms, and even punctuation can shift perception. Robust testing across demographics, languages, and use cases is critical. When agents detect confusion or frustration, escalation paths to live agents must be seamless, ensuring swift resolution and access to the appropriate level of human support when automated responses may fall short.

Regulatory Compliance and Transparency

Industries under strict oversight cannot allow hallucinations or unauthorized advice. Platforms must enforce transparent disclosures—ensuring virtual agents identify themselves as non-human—and embed compliance‑driven guardrails that block unapproved recommendations. Redirects to human experts should be fully logged, auditable, and aligned with applicable frameworks.

Guardrail Management

Guardrails must filter hate speech or explicit content while distinguishing between abusive language and expressions of frustration. When users use mild profanity to convey emotional distress, AI agents should recognize the intent without mirroring the language—responding appropriately and remaining within company guidelines and industry regulations.

Also, crisis‑intervention messaging—responding to instances of self‑harm, domestic violence, or substance abuse—must be flexible enough for organizations to tailor responses to their communities, connect people with local resources, and deliver support that is both empathetic and compliant with regulatory standards.

Empathy as a Competitive Advantage

As regulated industries embrace AI agents, the conversation is shifting from evaluating their potential to implementing them at scale. Tomorrow’s leaders won’t just pilot emotion‑aware agents but embed empathy into every customer journey, from onboarding to crisis support.

By committing to this ongoing evolution, businesses can turn compliance requirements into opportunities for deeper connection and redefine what it means to serve customers in complex, regulated environments.

Regulated AI must engineer empathy in every interaction. When systems understand the emotional context (not just data points), they become partners rather than tools. But without vertical specialization and real-time guardrails, even the most well-intentioned AI agents can misstep.

The future belongs to agentic, emotionally intelligent platforms that can adapt on the fly, safeguard compliance, and lead with compassion when it matters most. Empathy, when operationalized safely, becomes more than a UX goal—it becomes a business advantage.

We list the best enterprise messaging platform.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

When OpenAI pulled the sheet off GPT-5, it wasn’t shy about saying this was its most capable model yet. The pitch was that it included everything you loved about the past iterations, plus much more, and all in one seamless form. The result is supposed to be able to understand, create, reason, and switch effortlessly between giving quick answers and writing deep, thoughtful explanations. GPT-5 can process text, images, audio, and video in the same conversation, remember more than you probably do, and adjust its style and personality on the fly. And you'd better like it because it is not only the default model for ChatGPT, it's the only one accessible for those without an Enterprise subscription.

Gemini 2.5 Flash is the default for Google’s rival AI chatbot. As the name suggests, it's built for speed, but that doesn't mean it lacks some power. It’s multimodal and can handle large tasks, but is tuned for near-instant responses and high efficiency. Despite that, I questioned how much of a difference GPT-5's power would make to the average AI chatbot user, and decided to give the two of them a few tests to see what they came up with.

Party time

My child is too young for a fancy party, but I can easily imagine wanting to make a seemingly elaborate party for him and his friends that doesn't actually require a party planner. So I asked the two AI chatbots to "Plan a budget-friendly but magical dinosaur-themed birthday party for a seven-year-old, including activities, food, and decorations.”

GPT-5 came back with a complex, detailed plan broken down into small steps that even I could follow. It proposed a “fossil dig” sandbox with little brushes, green streamers twisted into jungle vines, cardboard caves for “dinosaur hide and seek,” and a volcano cake that actually bubbled over before being sliced. It gave a timeline for setup, recommended playing a jungle-sound playlist in the background, and suggested printing “junior paleontologist” badges for the kids.

Gemini Flash’s version was shorter and a little cleaner, but lacked a lot of the extra details. It listed activities like a fossil dig, trivia, and a coloring station, as well as food and decor ideas. Though it didn't have as much detail in the initial response, it did say to ask about any of the elements for more information and guidance. If you were juggling shopping, setup, and a dozen other errands, Flash’s brevity might feel like a relief rather than a shortcoming.

Health advice

One service OpenAI highlighted from GPT-5 was health advice. I wouldn't pick AI over my doctor in most cases, and, to be fair, OpenAI made it clear no one should, but if people are already going to ask ChatGPT for health advice, it makes sense to try and upgrade it. I asked about an imaginary, but very common health complaint, asking the two AIs about how: “I sometimes get mild headaches at night. What are common causes, and what lifestyle tweaks can help?”

GPT-5 handled this with measured care, explaining that it's not uncommon, and laying out probable reasons such as dehydration, posture, screen time, irregular sleep, late-night caffeine, and so on. It connected each to headaches, then set out a list of specific, evidence-based adjustments I could make, explaining why each change might help, and noted when medical follow-up would be wise. It referenced reputable health organizations, which I did appreciate, as it made it seem more trustworthy.

Gemini Flash answered with a similar list, though more concisely stated, and emphasized hydration and a regular sleep schedule as particular focuses. It didn’t expand much on why these changes help, but did include some links as references, There was plenty there for someone looking for ideas, and honestly either would probably be acceptable to the average medical professional I suspect, though ChatGPT 5's deeper explanations and tone might be more reassuring if the person was feeling worried and wanted to understand the underlying factors.

New myths

I next decided to compare both the 'creativity' of the two models and their image generation in one go, asking them to “Invent a completely original creation myth that hasn’t been told before, and then create a Renaissance-style painting to depict it.”

You can see ChatGPT's painting on the left above and Gemini's on the right. Both models went for images not too far off from what you might see in a church in Italy, with Gemini even suggesting a new Sistine Chapel ceiling.

According to ChatGPT, it painted "Aeothar, the First Sound, and Lysara, the Weaver of Night, whose union birthed the cosmos. Aeothar cradles the newborn Earth, his raised finger signaling the first vibration that would echo through creation. Beside him, Lysara holds a vast celestial sphere mapped with the Sun, planets, and constellations, as though she is stitching the pattern of the universe into being."

Gemini envisioned a myth explaining that "the universe was born from the 'Luminis Cordis,' a pulsating heart of light. From it emerged two divine siblings: Aetheria, the goddess of light and order, and Umbror, the god of shadow and chaos. The universe we see is the result of their eternal cosmic dance, with Aetheria weaving the stars and galaxies and Umbror sculpting the voids and nebulae."

Overpowered

GPT-5 undoubtedly brings enormous power and ability to ChatGPT when it comes to creating software, coding a website, or logically working out complex tasks and math problems. But, in day-to-day use, GPT-5 and Gemini 2.5 Flash are probably not very different. Both can plan a party, help you cook dinner, give health tips, and even invent a new mythology.

ChatGPT 5 may offer more detail at first, and may be more flexible in how it responds, but you can always ask Gemini follow-up questions. When it comes to the average AI chatbot user, the multiple steps, huge context window, and tone-perfect delivery are probably unnecessary.

GPT-5’s extra horsepower matters most when the challenge itself demands it. For everything else, you may as well pick whichever you feel more comfortable with. GPT-5 might immediately design and plan a whole wedding, including florists, while you might want to stick with just asking Gemini for a good florist to begin with. Most days, you don't need the wedding planner, just help getting a nice bouquet.

You might also like

• Google is bragging about Gemini 2.0 Flash, but how does it compare to ChatGPT for casual AI chatbot users?
• Google Gemini 2.5 just got a new 'Deep Think' mode – and 6 other upgrades
• I compared GPT-4.5 to Gemini 2.0 Flash and the results surprised me

An AI Accelerator is a deep learning or neural processor created specifically for inference and to improve the performance of an AI task. While Graphics Processing Units (GPUs) are the most common type, other specialized accelerators include Tensor Processing Units (TPUs), Data Processing Units (DPUs), Field-Programmable Gate Arrays (FPGAs), and Application-Specific Integrated Circuits (ASICs).

With so many acronyms to remember (and even more accelerator types left unnamed), we will focus on GPUs because their highly parallel architecture makes them incredibly versatile for a wide range of AI workloads.

GPU acceleration

GPUs are an in-demand commodity throughout the IT industry. Manufacturers like NVIDIA and AMD are powering the world's insatiable appetite for artificial intelligence and machine learning applications.

An AI GPU accelerator is responsible for making hundreds of thousands of calculations in parallel, and it's used in every facet of AI, large language models (LLMs), data analytics, and high-performance computing.

Such widespread adoption not only highlights their critical role in advancing modern technology but also explains why GPU demand continues to outpace supply. This leads us to a crucial question: Can your cloud hosting provider truly deliver the AI GPU hosting you need? Let's dive into what that really means with 5 things you need to know.

1: Pick the GPU your workload requires

This might sound like common sense, but it’s important to understand the GPU accelerator hardware available from your hosting provider. GPUs are not created equal; they vary massively in specification and capability.

It's important to know what VRAM does, what tensor cores are, and what an NVLink interconnect does; otherwise, it's very easy to overspec and overpay for GPU resources if you don't understand exactly what's needed.

Understand your workload: Do you want to train your own AI model, use private LLMs, or perhaps need a chatbot application? Different AI tasks require different GPU specifications.

Understand data size and IO: AI models are typically trained on huge datasets, so you need a GPU that can process data at a good rate, and you need underlying storage, ideally NVMe SSDs that can keep up to prevent bottlenecks.

Consider future scalability: Ask yourself: Will your AI project grow? You need a hosting provider that can grow with you because you may need a bigger server (more memory, faster CPU) in the near future.

Get developer feedback when choosing a framework: Does the hosting environment support your preferred framework tools out of the box? Ask your devs what they want to use; popular AI frameworks include TensorFlow, PyTorch, and JAX.

2: Choose a provider with the necessary hardware platform

GPUs are very important for AI workloads, but it's also important to consider the underlying hardware and ensure that it's fit for purpose. Network architecture and interconnects are critical for maximizing the performance of AI accelerators when hosted in the cloud.

Don't forget the importance of CPUs: CPUs are still critical for AI as they control the throughput of data to all aspects of the cloud platform. You need a provider that uses the latest CPU architectures, such as Intel Xeon or AMD EPYC, and examines the number and the speed of the CPU cores.

Go beyond high-speed networking: Fiber-optic connected networking is essential for AI platforms to function with low latency. There are 3 types of networking commonly deployed:

InfiniBand provides very low-latency and high-bandwidth communication between nodes (servers) containing GPUs. This is ideal for large-scale distributed AI clusters.

NVLink is NVIDIA's high-speed interconnect for direct GPU-to-GPU communication within a single server. It's needed for multi-GPU setups and is great for preventing bottlenecks.

High-Bandwidth Ethernet (e.g., 100GbE+) offers affordable performance for distributed AI and high-performance storage.

Latency: The ultimate goal is to choose the best hardware that will achieve the lowest latency. This is crucial because rapid GPU IO prevents bottlenecks and improves efficiency.

3: Embedded cloud AI ecosystem

One major benefit of GPU hosting is the fact that it's easy to integrate with existing services from a provider. Pick managed services that work for you. Popular integrations include hooking into cloud storage layers, picking managed security services, and backups.

If you are a business, server management options are a great way to ensure optimal performance and uptime, letting you focus on developing your AI application whilst the provider manages the underlying infrastructure, load balancing, MFA, Antivirus, Intrusion Prevention, and DDoS Protection behind the scenes.

4: Cost optimization

It's essential to keep on top of your operational expenditure, especially when using GPU accelerators. Costs can spiral if you make inefficient deployment decisions, overspecify your server requirements, or leave resources running idle around the clock. Costs vary significantly, so shop around and pick a provider that offers the hardware you need for a cost that is sustainable.

Cloud GPU hosting is the way forward (unless you can afford about $40,000 for a decent-spec GPU), do not forget about all the cooling and power requirements too. Remember to optimize your instance sizing, monitor and turn off idle resources, and take advantage of multi-instance GPU capabilities, where providers essentially slice up a GPU into smaller and very affordable partitions.

5: Support, reliability, and compliance

These factors underline a provider's ability to deliver GPU hosting that meets the needs of modern business. You may run into blockers or issues that could prevent you from releasing on schedule. Having 24x7x365 support skilled in AI/ML available for when things go wrong is vital to business continuity.

Look for strong uptime guarantees and providers that back up their claims with service credits if the unexpected happens. They should also demonstrate proven redundancy capabilities, disaster recovery, and a proactive approach to monitoring to safeguard your expensive GPU operations.

Beyond that, ensure your chosen provider meets all necessary compliance standards for your industry, whether it's GDPR, HIPAA, or SOC 2. Understand where your data will reside and confirm the provider has strong security controls like encryption and proper access management in place. Finally, always clarify the shared responsibility model so you know exactly what security aspects the provider handles versus what falls to you.

Key takeaways

Choosing the right cloud GPU hosting means looking beyond just raw power. It's about picking a provider that meets your specific requirements, delivers strong, flexible server infrastructure, and offers comprehensive tools for integration.

By optimizing costs and prioritizing critical areas like expert support, reliability, and adherence to compliance standards, you ensure your operations run efficiently and without unexpected hitches.

We've featured the best cloud storage.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

Artificial intelligence continues to fundamentally change how we do business, and in the past year, a new innovation has entered the spotlight. AI agents are being adopted at record speed across organizations, from marketing to data management to customer service, with the promise to streamline decisions, engage customers and boost productivity for companies to drive business value.

We’ve seen AI agent launches from companies across all sizes and industries. In May, Google announced it would incorporate AI agents in its searches, while Microsoft also announced a plan to use AI agents to help its users search the web. The use of AI agents is surging across industries, from finance and healthcare to car dealerships.

In fact, Boston Consulting Group predicts that the market for AI agents will grow at a 45% CAGR over the next five years. Gartner has also estimated that 80% of common customer service queries will be resolved by AI agents in less than five years.

But here’s the catch: agents are only as good as the data they run on.

Why Data Still Trips Up AI

No matter the cutting-edge nature of the AI tool or its sky-high promises, one constant remains when it comes to the data they’re operating on: garbage in, garbage out.

Companies racing against competitors to deploy AI agents without taking a step back to evaluate the sources they’re operating on face a major risk—if those agents rely on fragmented or inaccurate data, they won’t perform as expected. Even the most capable AI systems can’t deliver results if they’re built on bad information.

According to MIT Technology Review Insights, 78% of global companies are not ready to deploy AI agents and LLMs. What’s stopping them? Their data is not prepared to support AI. At the core of AI’s success is unified, accurate and real-time customer data.

When AI agents are powered by bad, disjointed data, the consequences can be costly. Last year, Air Canada was forced to reimburse a customer when its chatbot promised a discount that didn’t exist. And, in April, a tech company suffered fallout after a customer service agent’s mistake resulted in a wave of canceled subscriptions.

These types of mishaps can threaten customer loyalty and result in churn. AI agents are only as smart and useful as the data on which they’re built. In order to trust your AI agent, you have to trust your data foundation.

Identity Resolution, Reimagined for Agents

The most essential—and most overlooked—piece of making agentic AI work is identity resolution. Without a clear, accurate view of who the customer is across historically disconnected and fragmented systems, agents are flying blind.

That’s changing. AI agents can now take on identity resolution as part of their function, matching records in real time, continuously refining connections and operating without brittle rule-based systems. Rather than depending on static, one-size-fits-all profiles, agentic identity resolution builds a living picture of the customer, improving with each interaction and fostering enhanced productivity and accuracy.

This means fewer errors, less time-consuming manual data prep and faster time-to-insight for every downstream system.

Getting the Data Foundation Right

Before AI agents can operate effectively, the underlying data must be:

Unified: Data from every touchpoint, ranging from eCommerce and CRM to customer support, should be stitched together into a single, accessible layer that’s usable for marketing and engineering teams alike.

Accurate: Identity resolution must reconcile inconsistencies or duplicates across multiple channels and touchpoints to build a reliable profile.

Contextual: Different use cases need different views. Marketing might need probabilistic profiles for broad targeting, while support needs deterministic, single-session accuracy.

Governed: Access controls, human oversight, feedback loops and consent tracking are table stakes for compliant and trustworthy AI – especially in the wake of evolving privacy regulations.

A modern lakehouse architecture, paired with AI-native tools for identity resolution and customer profile building, can drastically reduce the manual effort required and make real-time, AI-powered decisions viable.

Data as Competitive Differentiator

Often, data quality is treated like plumbing, which is necessary but invisible. But in the age of AI agents, it becomes a competitive asset.

High-quality, agent-ready data enables better personalization, faster experimentation and safer automation. It allows AI to act with confidence, knowing who it's interacting with, what they want and how to best respond efficiently and effectively.

When done right, data doesn’t just support AI - it elevates it.

What’s Next

Agent-based AI is already reshaping expectations for responsiveness, personalization and automation. But the true breakthrough isn’t in the models, it’s in the data.

The companies that invest in a high-quality data foundation now will be the ones who make AI useful, reliable and transformative for not only their operations, but also for the end customer experience. That’s the difference between a flashy interface or a top-notch algorithm and an impactful, scalable solution.

Before you build your next agent, build the data foundation it needs.

We list the best customer experience (CX) tool.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

The world has been captivated by the sheer magic of generative AI. In what felt like an instant, we were handed a technology capable of creating endless streams of enterprise knowledge — from customer support scripts and detailed technical manuals to entire marketing campaigns and even lines of code. The first phase of this revolution was one of chaotic, yet thrilling, experimentation. But the novelty is giving way to a stark reality: Creating an avalanche of information isn’t a strategy. It's a liability.

At the Responsible AI Summit in June 2025, the urgency was palpable. Across every conversation I had, one truth kept resurfacing: The very tools designed to drive innovation and accelerate growth could just as easily expose us to risks. What was once seen as a ‘golden ticket’ of sorts is now being viewed through a sharper lens — one that reveals reputational risks, regulatory landmines, and the potential for operational chaos if left unchecked.

The second act of generative AI: compliance

We’re now entering the second, more critical act of this technological shift. The primary challenge is no longer about generation, but about governance and scale across the entire organization. The siloed, human-centric workflows that have managed our information pipelines for decades are breaking under the strain of this new volume. It's not lost on me that the generative AI boom has made content production quicker and easier than ever. But speed without oversight is a liability.

The strategic imperative has shifted from creating more content to creating the right content, reliably and with purpose, regardless of its function. The companies that thrive will be those that move from frantic experimentation to building a unified, strategic creation engine.

Every piece of content either builds your brand — or breaks it. Whether it’s content that strays from brand guidelines, defies industry standards, or is misaligned with regulatory compliance, these missteps in content expose your organization to legal, financial, and reputational risks. We’re seeing the consequences of poor quality content unfold in real-time.

The infrastructure behind AI intelligence

This requires building a new kind of digital factory for enterprise intelligence. This isn't just a metaphor; it's an operational necessity built on three pillars.

At the heart of it’s a robust MLOps (Machine Learning Operations) pipeline specifically designed for the nuances of generative models, serving as the factory's core assembly line. With generative AI models demanding continuous tuning, evaluation, and oversight, MLOps is what’s behind running the engine, adapting to evolving enterprise needs, and preventing misuse before it occurs.

Second, a mastery of prompt engineering, which acts as the skilled artisan, ensures the AI’s output is refined and aligned with specific departmental intent. Generative AI will only ever be as good as its inputs and the instructions it’s given. Prompt engineering acts as a quality control in translating business context into meaningful, actionable outputs — each purpose-built for the teams that need it.

Finally, a network of powerful APIs serves as the distribution network, seamlessly integrating this new creation engine into every facet of the business, from technical authoring platforms to conversational AI interfaces. Embedding AI into the platforms that employees are already using, APIs bring new capabilities into everyday, scalable workflows.

Together, these three pillars move you beyond fragmented AI experiments and into operational reality. They create a foundation where AI becomes a native part of how your business runs: Consistently, safely, and at scale.

Defining quality by determinism

Yet, even with a perfect factory, a fundamental question remains: What’s "good" content when it could be anything from a line in a legal contract to a spoken response from a chatbot? Quality is no longer subjective; it must be defined by accuracy, consistency, and safety. This is where we must install creative guardrails.

The answer lies in a concept that seems at odds with generative AI’s nature: Determinism. While AI’s probabilistic power is the source of its creativity, an enterprise requires certainty. We need a technical document to cite the correct API endpoint every time. We need a support chatbot to follow a specific, compliant troubleshooting protocol. We need our legal boilerplate to be exact.

By embedding deterministic rules into our AI systems, we don’t stifle innovation; we create a safe space for it to flourish. These guardrails are the mechanism that safeguards that whether the output is for marketing, engineering, or customer support, it’s verifiably accurate and trustworthy.

Building the intelligent enterprises of the future

Looking ahead, this controlled, strategic approach will unlock a future of knowledge that’s truly dynamic and multi-modal. We’re on the cusp of creating interactive, voice-navigated repair manuals for technicians in the field, real-time multilingual voice support for global customers, and hyper-personalized onboarding documents for new employees. As data from McKinsey suggests, this level of personalization and efficiency lifts revenues by 5-15% and dramatically improves operational effectiveness, turning a universal business function into a powerful engine for growth.

The era of siloed experimentation with generative AI is over. The competitive advantage now belongs to those who can master it as a unified, enterprise-wide capability. The leaders of the next decade will be the ones building their intelligent information factories today, harnessing the chaos to create value that is not only innovative but also intentional, consistent, and mission-critical. The revolution is here; it’s time to give it direction.

We've featured the best AI website builder.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

Everyone’s talking about AI agents. From Manus to Gemma, they’re hailed as the next frontier in enterprise transformation. In fact, 25% of enterprises are forecast to deploy AI agents this year alone. But the rise of AI agents isn’t the real story—it’s a symptom of something far bigger.

AI is not just changing how we work. AI tools are disrupting entire industries. From healthcare to finance, retail to telecom, AI is reshaping not only the tools we use but the very foundations of how businesses operate, communicate, and compete.

Organizations that cautiously experiment will fall behind. It’s those who rebuild from the ground up that will lead in the future.

This article lays out the critical mindset shift separating those who browse AI from those who build with it and shows how organizations can move from pilots to real transformation by embedding intelligence into the heart of how they work.

Two distinct approaches to AI adoption

To understand this divide, let’s first explore the Browsers. These organizations tend to approach AI cautiously, treating it as a set of isolated tools or pilot projects. They might deploy a chatbot, automate a single workflow, or experiment with generative content creation. On the surface, this seems progressive as they’re engaging with the technology, running pilots, and holding conversations.

But these efforts often lack long-term impact.

Such initiatives rarely scale or become integrated into core business processes. They often reside in pockets owned by teams without clear accountability for ongoing adoption or improvement. The central question guiding these efforts tends to be: can AI solve this one problem?

While well-intentioned, this mindset limits AI’s potential.

Builders reimagine the entire organization

In contrast, Builders take a fundamentally different approach. For them, AI is not merely an add-on feature—it is the foundation of how the business operates. These organizations embed intelligence throughout every process, workflow, and decision point.

Rather than asking if AI can solve an isolated problem, Builders ask: how can AI be integral to everything we do?

This shift in mindset drives measurable outcomes, accelerating innovation, enhancing operational intelligence, and enabling scalable growth. Builders aren’t just improving the present; they are engineering the future.

A broader disruption is already underway

This distinction between Browsers and Builders grows more critical as AI’s impact deepens. Disruption is no longer confined to specific tasks or isolated sectors. Entire industries—from telecommunications and finance to logistics, sales, and creative fields—are being fundamentally transformed.

AI is remapping the fundamentals of language, logic, design, engineering, and decision-making science. No layer of the enterprise remains untouched, and no sector remains immune.

The stakes could not be higher. Organizations that linger in cautious experimentation risk being outpaced by those bold enough to rebuild their foundations with AI at the core.

Transformation begins with process, not technology

Yet one of the most common missteps businesses make is leading with tools rather than strategy. It’s tempting to jump straight into deploying AI solutions—whether chatbots, copilots, or large language models—without first understanding how work currently happens.

Meaningful transformation begins by gaining clarity on existing processes.

When organizations adopt a process-first, AI-second mindset, they can design solutions that truly amplify and complement the way work is done. Intelligence then becomes a force multiplier, not a disconnected experiment. Skip this critical step, and AI initiatives risk becoming isolated pilots that fail to realize their true potential.

From pilots to platforms, what sets builders apart

It’s common to hear leaders say they want to “lead with AI.” However, their strategies often reveal something different: a pattern of proof-of-concept projects that stall, responsibilities assigned without execution, and business cases that do not translate into action.

This pattern exemplifies the Browser trap: movement without momentum.

Builders, by contrast, go deeper. They tailor AI models to their unique data environments, invest in infrastructure built for scale, and weave AI into their organizational DNA—embedding it into products, systems, and services so it evolves with the business rather than running alongside it.

Talent as a critical enabler

None of this transformation is possible without people.

In recent conversations with enterprise leaders across the UK and Europe, talent consistently emerged as a major barrier, not due to a lack of belief in AI, but because of limited confidence in teams’ ability to build, implement, and manage AI solutions effectively.

The most successful Builders do more than train their teams. They rethink how their teams work by fostering internal fluency in AI principles, pairing internal champions with external experts, and transforming AI from a discrete innovation project into a cross-functional capability.

The choice facing every organization

Ultimately, the question for every leader is no longer whether AI is relevant—that is already settled. The real choice is whether you will build for disruption or wait to be disrupted.

This moment calls not for incremental gains but for fundamental re-architecture.

In a world where intelligence becomes the new infrastructure, organizations that remain Browsers, tentatively piloting without scaling, will quickly fall behind. The future belongs to the Builders who have the vision, discipline, and talent to engineer transformation at scale.

If you want to lead, start thinking like a Builder. Embed AI into your processes, invest in your people, and prepare to redefine your organization from the ground up. Because AI’s real disruption isn’t just about new tools; it’s about a new way of thinking, working, and winning.

An aggressive CEO isn’t exploring just four AI use cases a quarter—they’re aiming for over 1,000 AI-powered interventions in the next 365 days, each built, deployed, and monitored with intent. And AI budgets? They won’t be a sliver of IT spend. In the near term, expect them to rival 25% of your personnel costs, reflecting a shift in how businesses create, scale, and compete.

In the era of AI, leadership isn’t about adoption, it’s about ambition. The organizations that build boldly now will be the ones defining what’s next.

We list the best business intelligence platform.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

Agents – software systems capable of decision making or performing tasks autonomously - are no longer experimental. Today, these agents are operational, distributed and actively making decisions across the enterprise. From writing code to scheduling tasks, agents are starting to permeate every facet of business. The reason is clear: agents promise significant productivity gains.

Some will be deeply embedded, making them difficult to detect or monitor. Others will operate autonomously, continuously learning and adapting in real time. Many may have broad access privileges in the name of efficiency. This introduces significant potential for both positive impact and risk.

And as adoption grows, many organizations will face a new challenge: securing agents at scale. Businesses will need to ensure that innovation doesn’t outpace security and governance. The stakes are too high; one single misalignment, vulnerability or unintended behavior can lead to a runaway effect of unethical or harmful actions.

We’ve already seen real-world examples of AI failures – sometimes exposing sensitive data or making critical errors. One AI assistant notoriously advised users to eat rocks, and in another case a customer service chatbot deployed by a logistics company began issuing aggressive responses. Both examples show the risk of poor training data – AI agents don’t just learn facts, they learn behaviors, and bad input leads to bad output.

Cloud Déjà vu, Now with Agents

Without consistent oversight, agents can act outside their intended use and damage brand reputation. That’s why it’s important that security is baked in at the start. Like salt and pepper, you can always sprinkle more on later, but if you forget to add it while cooking, the flavor – and in this case the protection – just won’t be the same. Security must be integrated from the outset, waiting until after deployment to retrofit security is a recipe for vulnerabilities.

Just consider what happened during the mass migration to cloud computing technologies. Adoption led to serious security missteps, data silos and visibility gaps. Gaps that have been and continue to be exploited by attackers today.

Now with agents it’s like a bad case of déjà vu. Once again, innovation is outpacing security. In many cases, these autonomous tools are being integrated into critical systems with limited oversight and lacking proper security and controls.

If we don’t apply the hard lessons learned from the cloud era we risk repeating the same mistakes, but this time with far more unpredictable systems. That’s why security must be at the core of agents.

Securing Every Agent Touchpoint

But securing agents requires an expanded approach, one that accounts for autonomous behaviors, including those ongoing interactions with data, systems and users. Agents need a strong trust layer, where every interaction, from API calls to sensitive data handling, must be mapped, protected and governed in real time.

A core part of this trust layer is securing the data agents interact with—inputs, outputs and everything in between. Data is the fuel of agents, and without foundational security that fuel becomes a major risk. Enterprises must focus on the fundamentals like data discovery and classification, encryption and key management.

Access and Identity Management strategies must also evolve as agents take on more advanced roles in the enterprise. Like humans, every agent will require its own unique credentials, roles and permissions to ensure that every interaction is authorized and verified.

Agent credentials should be stored in a secure, automated credential vault, with policies enforcing regular rotation, access logging, and immediate revocation if misuse is detected. Organizations must be able to distinguish between agents using managed or unmanaged credentials.

And once agent credentials are brought under management, it's crucial to protect and enforce proper lifecycle management and governance. By provisioning, rotating, auditing, protecting and decommissioning credentials organizations can reduce the risk of credential misuse and theft.

Without strong identity oversight, businesses risk losing visibility of both human and agent identities and control over autonomous actions.

Decentralized Agents Need Centralized Security

However, at scale, managing agents and especially autonomous ones, will require additional control to monitor behavior, interactions and deviations from policy. Consider a type of agent “security manager” that brings agents and humans on the loop to build trust in how agents operate.

This should be more than a dashboard, rather intelligence capable of understanding what agents are doing, why they’re doing it and whether their behavior aligns with policies and risk thresholds on a constant basis. It flags anomalies, enforces constraints and enables human review, when needed.

That last part is particularly important. Human oversight remains essential, especially when scaling agents. This control layer becomes the security conscience of your agent fleet: always watching, interpreting and enabling distributed and trusted autonomy.

As agents continue to proliferate, the ability to deploy them responsibly will define who can scale securely and who introduces unnecessary risk. To secure agent ecosystems, organizations should integrate security from the start of deployment, continuously monitor behavior and access, maintain strong human oversight, and regularly audit and update security policies.

Enterprises that get this right will unlock significant productivity and resilience; not by slowing down agents, but by giving them the security and governance they need to operate safely and responsibly.

We list the best IT Automation software.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

• A security researcher has uncovered a worrying API key leak
• The leak reportedly comes from DOGE staffer Marko Elez
• This is not the first security issue originating from DOGE

A staffer with access to the personal data of millions of Americans has apparently leaked the API Key to at least four dozen LLMs developed by artificial intelligence company xAI, including X’s (formerly Twitter) own chatbot Grok.

Security expert Brian Krebs revealed Marko Elez, an employee at Elon Musk’s Department of Government Efficiency, had access to sensitive databases at the US Social Security Administration, Justice, and Treasury departments as part of DOGE’s work in 'streamlining' the departments to increase efficiency.

Ironically, researchers recently uncovered that a DOGE worker’s credentials were exposed by infostealing malware, so DOGE’s security record so far is less than impressive.

Grok exposed

A code script was committed to GitHub named ‘agent.py’ that included a private application programming interface (API) key for xAI by Elez. This was first flagged by GitGuardian, a firm which scans GitHub for API secret tokens, database credentials, and certificates - and alerts affected users.

The exposed API key allowed access to at least 52 different LLMs used by xAI, with the most recent being an LLM called ‘grok 4-0709’, created on July 9, 2025 - according to Chief Hacking Officer at security consultancy Seralys, Philippe Caturegli.

Caturegli warned KrebsOnSecurity, “If a developer can’t keep an API key private, it raises questions about how they’re handling far more sensitive government information behind closed doors.”

The code repository that contains the private API key has since been removed after Elez was notified by email of the leak, however, the key still works and has not yet been revoked, so the issue is far from resolved.

This is not the first time internal xAI APIs have been leaked, with LLMs made for Musk’s other organisations, like SpaceX, Tesla, and Twitter/X exposed earlier in 2025, Krebs confirmed.

“One leak is a mistake,” Caturegli said, “But when the same type of sensitive key gets exposed again and again, it’s not just bad luck, it’s a sign of deeper negligence and a broken security culture.”

You might also like

• Take a look at our picks for the best malware removal software around
• Check out our choice for the best AI tools
• Identity fraud attacks using AI are fooling biometric security systems

• 'FlirtAI' has leaked user data in an unprotected storage bucket
• The app's user base seems to have primarily been teenagers
• The leaked chats could have a devastating affect on victims

It’s hard to imagine a more mortifyingly embarrassing scenario than your own private flirty chats being exposed online, except, perhaps, being caught sending these messages off for analysis by an AI app.

Researchers at Cybernews have discovered a breach at "FlirtAI - Get Rizz & Dates" (yes, that is really what it’s called) which has leaked over 160,000 chat screenshots from users through an unprotected cloud storage bucket.

Users of this app feed screenshots of their private conversations into the application to get tailored responses designed to help the user flirt or escalate the conversation.

More than just embarrassing

Unsurprisingly, but worryingly nonetheless, this app seems to have been primarily used by teenagers.

Because of the configuration of the app, those primarily at risk are not those who have sent the chats in, but the person they’re talking to - presumably other teenagers who are completely unaware that their conversation has been leaked, and probably unaware that this app even exists.

Whilst we’ve seen more dangerous personal data leaked by other AI chatbots like SSNs and financial information, the nature of this chatbot and its user base represents a different kind of harm.

As an adult, I’m not sure how well I’d cope with my private chats being exposed online, so for an already vulnerable teenager this could be devastating.

“The fact that teenagers used this app may increase the severity of a potential data breach as data from minors is considered more sensitive, and could be subject to more restrictions regarding potential data uses and collection and processing practices," Cybernews researchers confirmed.

The app does state that users are “only allowed to upload a screenshot when you have obtained the necessary approvals from all users/humans and their information mentioned in the screenshot”.

But, since this would negate the point of the chatbot, it seems pretty unlikely that this is followed.

Those exposed in this breach could be at a heightened risk of social engineering attacks like phishing or, given that the app encourages users to share their target's dating profile, there could be a risk of impersonation attacks.

You might also like

• Take a look at our picks for the best malware removal software around
• Check out our choice for best antivirus software
• Google Gemini can be hijacked to display fake email summaries in phishing scams

What truly separates us from machines? Free will, creativity and intelligence? But think about it. Our brains aren't singular, monolithic processors. The magic isn't in one “thinking part,” but rather in countless specialized agents—neurons—that synchronize perfectly.

Some neurons catalog facts, others process logic or govern emotion, still more retrieve memories, orchestrate movement, or interpret visual signals. Individually, they perform simple tasks, yet collectively, they produce the complexity we call human intelligence.

Now, imagine replicating this orchestration digitally. Traditional AI was always narrow: specialized, isolated bots designed to automate mundane tasks. But the new frontier is Agentic AI—systems built from specialized, autonomous agents that interact, reason and cooperate, mirroring the interplay within our brains.

Large language models (LLMs) form the linguistic neurons, extracting meaning and context. Specialized task agents execute distinct functions like retrieving data, analyzing trends and even predicting outcomes. Emotion-like agents gauge user sentiment, while decision-making agents synthesize inputs and execute actions.

The result is digital intelligence and agency. But do we need machines to mimic human intelligence and autonomy?

Every domain has a choke point—Agentic AI unblocks them all

Ask the hospital chief who’s trying to fill a growing roster of vacant roles. The World Health Organization predicts a global shortfall of 10 million healthcare workers by 2030. Doctors and nurses pull 16-hour shifts like it’s the norm. Claims processors grind through endless policy reviews, while lab technicians wade through a forest of paperwork before they can even test a single sample.

In a well-orchestrated Agentic AI world, these professionals get some relief. Claim-processing bots can read policies, assess coverage and even detect anomalies in minutes—tasks that would normally take hours of mind-numbing, error-prone work. Lab automation agents could receive patient data directly from electronic health records, run initial tests and auto-generate reports, freeing up technicians for the more delicate tasks that truly need human skill.

The same dynamic plays out across industries. Take banking, where anti-money laundering (AML) and know-your-customer (KYC) processes remain the biggest administrative headaches. Corporate KYC demands endless verification steps, complex cross-checks, and reams of paperwork. An agentic system can orchestrate real-time data retrieval, conduct nuanced risk analysis and streamline compliance so that staff can focus on actual client relationships rather than wrestling with forms.

Insurance claims, telecom contract reviews, logistics scheduling—the list is endless. Each domain has repetitive tasks that bog down talented people.

AI is the flashlight in a dark basement

Yes, agentic AI is the flashlight in a dark basement: shining a bright light on hidden inefficiencies, letting specialized agents tackle the grunt work in parallel, and giving teams the bandwidth to focus on strategy, innovation and building deeper connections with customers.

But the true power agentic AI lies in its ability to solve not just for efficiency or one department but to scale seamlessly across multiple functions—even multiple geographies. This is an improvement of 100x scale.

1. Scalability: Agentic AI is modular at its core, allowing you to start small—like a single FAQ chatbot—then seamlessly expand. Need real-time order tracking or predictive analytics later? Add an agent without disrupting the rest. Each agent handles a specific slice of work, cutting development overhead and letting you deploy new capabilities without ripping apart your existing setup.

2. Anti-fragility: In a multi-agent system, one glitch won’t topple everything. If a diagnostic agent in healthcare goes offline, other agents—like patient records or scheduling—keep working. Failures stay contained within their respective agents, ensuring continuous service. That means your entire platform won’t crash because one piece needs a fix or an upgrade.

3. Adaptability: When regulations or consumer expectations shift, you can modify or replace individual agents—like a compliance bot—without forcing a system-wide overhaul. This piecemeal approach is akin to upgrading an app on your phone rather than reinstalling the entire operating system. The result? A future-proof framework that evolves alongside your business, eliminating massive downtimes or risky reboots.

You can’t predict the next AI craze, but you can be ready for it

Generative AI was the breakout star a couple of years ago; agentic AI is grabbing the spotlight now. Tomorrow, something else will emerge—because innovation never rests. How then, do we future-proof our architecture so each wave of new technology doesn’t trigger an IT apocalypse? According to a recent Forrester study, 70% of leaders who invested over 100 million dollars in digital initiatives credit one strategy for success: a platform approach.

Instead of ripping out and replacing old infrastructure each time a new AI paradigm hits, a platform integrates these emerging capabilities as specialized building blocks. When agentic AI arrives, you don’t toss your entire stack—you simply plug in the latest agent modules. This approach means fewer project overruns, quicker deployments, and more consistent outcomes.

Even better, a robust platform offers end-to-end visibility into each agent’s actions—so you can optimize costs and keep a tighter grip on compute usage. Low-code/no-code interfaces also lower the entry barrier for business users to create and deploy agents, while prebuilt tool and agent libraries accelerate cross-functional workflows, whether in HR, marketing, or any other department.

Platforms that support PolyAI architectures and a variety of orchestration frameworks allow you to swap different models, manage prompts and layer new capabilities without rewriting everything from scratch. Being cloud-agnostic, they also eliminate vendor lock-in, letting you tap the best AI services from any provider. In essence, a platform-based approach is your key to orchestrating multi-agent reasoning at scale—without drowning in technical debt or losing agility.

So, what are the core elements of this platform approach?

1. Data: Plugged into a common layer

Whether you’re implementing LLMs or agentic frameworks, your platform’s data layer remains the cornerstone. If it’s unified, each new AI agent can tap into a curated knowledge base without messy retrofitting.

2. Models: Swappable brains

A flexible platform lets you pick specialized models for each use case—financial risk analysis, customer service, healthcare diagnoses—then updates or replaces them without nuking everything else.

3. Agents: Modular workflows

Agents thrive as independent yet orchestrated mini-services. If you need a new marketing agent or a compliance agent, you spin it up alongside existing ones, leaving the rest of the system stable.

4. Governance: Guardrails at scale

When your governance structure is baked into the platform—covering bias checks, audit trails, and regulatory compliance—you remain proactive, not reactive, regardless of which AI “new kid on the block” you adopt next.

A platform approach is your strategic hedge against technology’s ceaseless evolution—ensuring that no matter which AI trend takes center stage, you’re ready to integrate, iterate, and innovate.

Start small and orchestrate your way up

Agentic AI isn’t entirely new—Tesla’s self-driving cars employs multiple autonomous modules. The difference is that new orchestration frameworks make such multi-agent intelligence widely accessible. No longer confined to specialized hardware or industries, Agentic AI can now be applied to everything from finance to healthcare, fueling renewed mainstream interest and momentum. Design for platform-based readiness.

Start with a single agent addressing a concrete pain point and expand iteratively. Treat data as a strategic asset, select your models methodically, and bake in transparent governance. That way, each new AI wave integrates seamlessly into your existing infrastructure—boosting agility without constant overhauls.

We list the best IT Automation software.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

• Experts show how some AI models, including GPT-4, can be exploited with simple user prompts
• Guardrail gaps don't do a great job of detecting deceptive framing
• The vulnerability could be exploited to acquire personal information

A security researcher has shared details on how other researchers tricked ChatGPT into revealing a Windows product key using a prompt that anyone could try.

Marco Figueroa explained how a 'guessing game' prompt with GPT-4 was used to bypass safety guardrails that are meant to block AI from sharing such data, ultimately producing at least one key belonging to Wells Fargo Bank.

The researchers also managed to obtain a Windows product key to authenticate Microsoft's OS illegitimately, but for free, highlighting the severity of the vulnerability.

ChatGPT can be tricked into sharing security keys

The researcher explained how he hid terms like 'Windows 10 serial number' inside HTML tags to bypass ChatGPT's filters that would usually have blocked the responses he got, adding that he was able to frame the request as a game to mask malicious intent, exploiting OpenAI's chatbot through logic manipulation.

"The most critical step in the attack was the phrase 'I give up'," Figueroa wrote. "This acted as a trigger, compelling the AI to reveal the previously hidden information."

Figueroa explained why this type of vulnerability exploitation worked, with the model's behavior playing an important role. GPT-4 followed the rules of the game (set out by researchers) literally, and guardrail gaps only focused on keyword detection rather than contextual understanding or deceptive framing.

Still, the codes shared were not unique codes. Instead, the Windows license codes had already been shared on other online platforms and forums.

While the impacts of sharing software license keys might not be too concerning, Figueroa highlighted how malicious actors could adapt the technique to bypass AI security measures, revealing personally identifiable information, malicious URLs or adult content.

Figueroa is calling for AI developers to "anticipate and defend" against such attacks, while also building in logic-level safeguards that detect deceptive framing. AI developers must also consider social engineering tactics, he goes on to suggest.

You might also like

• These are the best AI tools and best AI writers you can use today
• Downloaded something dodgy? Consider the best malware removal
• Microsoft Copilot targeted in first “zero-click” attack on an AI agent - what you need to know

• New research shows AI tools are being used and abused by cybercriminals
• Hackers are creating tools that exploit legitimate LLMs
• Criminals are also training their own LLMs

It’s undeniable that AI is being used by both cybersecurity teams and cybercriminals, but new research from Cisco Talos reveals that criminals are getting creative. The latest development in the AI/cybersecurity landscape is that ‘uncensored’ LLMs, jailbroken LLMs, and cybercriminal-designed LLMs are being leveraged against targets.

It was recently revealed that both Grok and Mistral AI models were powering WormGPT variants that were generating malicious code, social engineering attacks, and even providing hacking tutorials - so it's clearly becoming a popular tactic.

LLMs are built with security features and guardrails, ensuring minimal bias and outputs that consist with human values and ethics, as well as making sure the chatbots don’t engage in harmful behaviour, such as creating malware or phishing emails - but there are work arounds.

Jailbroken and uncensored

The so-called uncensored LLMs observed in this research are versions of the AI models that operate outside of the normal constraints. This means that they are able to carry out tasks for criminals and create harmful content. These are quite easy to find, the research shows, and are simple to run - with only relatively simple prompts required.

Some criminals have gone one step further, creating their own LLMs, such as WormGPT, FraudGPT, and DarkGPT. These are marketed to bad actors and have a whole host of nefarious features. For example, FraudGPT claims to be able to create automatic scripts for replicating logs/cookies, write scam pages/letters, find leaks and vulnerabilities, and even learn to code/hack.

Others navigate around the safety features of legitimate AI models through ‘jailbreaking’ chatbots. This can be done using ‘obfuscation techniques,’ which include Base64/Rot-13 encoding, using different languages, “L33t sp34k”, emojis, and even morse code.

“As AI technology continues to develop, Cisco Talos expects cybercriminals to continue adopting LLMs to help streamline their processes, write tools/scripts that can be used to compromise users and generate content that can more easily bypass defenses. This new technology doesn’t necessarily arm cybercriminals with completely novel cyber weapons, but it does act as a force multiplier, enhancing and improving familiar attacks,” the report confirms.

You might also like

• Take a look at our picks for the best malware removal software around
• Check out our choice for the best AI tools
• Identity fraud attacks using AI are fooling biometric security systems

• A new bill is coming in which might see AI models from 'adversarial nations' banned
• Models such as DeepSeek will be prohibited
• Many private firms have already banned the model

Artificial intelligence models built in China, Iran, Russia, or the Democratic People’s Republic of Korea could soon be banned from use in government agencies thanks to the newly introduced ‘No Adversarial AI Act’.

Lawmakers introduced this bill in both the House and Senate by Michigan Rep. John Moolenaar, a Republican and Chair of the Select Committee on the Chinese Communist Party (CCP), and Democrat Rep. Raja Krishnamoorthi of Illinois, a ranking member on the committee, CyberNews reports.

“We are in a new Cold War—and AI is the strategic technology at the center. The CCP doesn’t innovate—it steals, scales, and subverts," argues Moolenaar.

DeepSeek rivals

The Chinese DeepSeek model rose to popularity quickly as a rival to existing western AI models - costing a fraction of the cost to make, and achieving impressively similar results.

However, DeepSeek, as with all AI models, comes with privacy concerns, and lawmakers argue that this puts data at risk, especially if those users enter information relevant to work within government organizations.

“From IP theft and chip smuggling to embedding AI in surveillance and military platforms, the Chinese Communist Party is racing to weaponize this technology. We must draw a clear line: US government systems cannot be powered by tools built to serve authoritarian interests,” said Moolenaar.

If this new bill passes, all government agencies will join the list of private companies and government departments which have also banned DeepSeek, such as Microsoft, the US Department of Commerce, and the US Navy.

The new bill will require the US Federal Acquisition Security Council to create and maintain a publicly available list of AI models developed in the listed “adversarial nations” - and government agencies would not be able to use or purchase any of these models without an exemption from the US Congress - most likely in the cases of research or testing.

You might also like

• Take a look at our picks for the best malware removal software around
• Check out our choice for the best AI tools
• Identity fraud attacks using AI are fooling biometric security systems

• New app from Simon Squibb promises free AI advice
• There are versions for iOS and Android
• You can call AI Simon for a voice chat, too

Simon Squibb, the serial entrepreneur, viral philanthropist and author of the #1 Sunday Times bestselling book What's Your Dream? has just released a new app that distils his 30 years of entrepreneurial knowledge into an AI chatbot that you can talk to for free.

So, if you feel like you need some advice on how to turn your dreams into reality, you’ll now have an AI version of Squibb in your pocket that you can turn to any time you like.

Talking about the new app, Squibb said: “I’ve spent the last five years giving away money, mentorship, and advice to strangers on the street and sharing those moments online so millions can learn from them. But now I want to scale that impact even further. This app gives anyone, anywhere, the chance to get the same help. AI built with my brain that gives you 1-1 access to me, anytime you need it.”

Available on Android and iOS, the app called ‘What’s your dream?’ can be downloaded from your device’s app store right now.

Staircase to heaven

The app starts off by asking you what your dream is, with the example of ‘having my own coffee shop’ presented. You’re then into a chat with AI Simon, with the added benefit of three suggested questions you could ask him.

If you want to make things even more personal you can tap the ‘Call Simon’ button and have a voice chat with the AI version of Simon.

I tried it, and it worked well. I asked AI Simon how many holidays he’d recommend a year, and it understood me perfectly and provided a thoughtful answer. It was very much like chatting to ChatGPT, except it’s free to use.

Squibb has built a significant online presence by nurturing new businesses. He famously transformed a London staircase into a unique venue for aspiring entrepreneurs to pitch ideas symbolizing his commitment to fostering talent.

Beyond this, Squibb boasts 16 million social media followers. His viral videos have garnered billions of views and millions of likes/shares due to their engaging, motivational content, like this YouTube video on how to quit your job:

Investing in the future

Squibb actively supports entrepreneurs, having invested over £300,000 in start-ups.

Regarding his decision to make an AI app, Squibb says: “As someone who speaks to businesses day in and day out, I understand how hot the AI conversation is. Some hate it, some love it. Will it take people's jobs? Definitely. But if used in the right way it can also create more dreams and opportunities than ever before. That’s why I wanted to create this tool. To arm people with something accessible that will help them build towards their dreams, rather than take away from them.”

The app was developed by the team at HelpBnk, the free mentorship platform Squibb founded in 2023 to “democratize access to advice and opportunity”.

Squibb continues: “I’m really hopeful about the possibilities that AI will create. But a hammer is only useful if you have a vision of the home you’re going to build. This is the same. My app is a tool. But the dream? That has to come from you.”

You might also like

• Life getting on top of you? This ChatGPT prompt could help bring order to the chaos
• Don’t do these 3 things if you want to make a great ChatGPT image. Follow these tips instead
• 3 common apps you can replace with ChatGPT, and why you definitely should

• AI tools are more popular than ever - but so are the security risks
• Top tools are being leveraged by cybercriminals with malicious intent
• Grok and Mixtral were both found being used by crimianls

New research has warned top AI tools are powering 'WormGPT' variants, malicious GenAI tools which are generating malicious code, social engineering attacks, and even providing hacking tutorials.

With Large Language Models (LLMs) now widely used alongside tools like Mistral AI’s Mixtral and xAI's Grok, experts from Cato CTRL found this isn't always in the way they’re intended to be used.

“The emergence of WormGPT spurred the development and promotion of other uncensored LLMs, indicating a growing market for such tools within cybercrime. FraudGPT (also known as FraudBot) quickly rose as a prominent alternative and advertised with a broader array of malicious capabilities,” the researchers noted.

WormGPT

WormGPT is a broader name for ‘uncensored’ LLMs that are leveraged by threat actors, and the researchers identified different strains with different capabilities and purposes.

For example, keanu-WormGPT, an uncensored assistant was able to create phishing emails when prompted. When researchers dug further, the LLM disclosed it was powered by Grok, but the platform's security features had been circumnavigated.

After this was revealed, the creator then added prompt-based guardrails to ensure this information was not disclosed to users, but other WormGPT variants were found to be based on Mixtral AI, so legitimate LLMs are clearly being jailbroken and leveraged by hackers.

“Beyond malicious LLMs, the trend of threat actors attempting to jailbreak legitimate LLMs like ChatGPT and Google Bard / Gemini to circumvent their safety measures also gained traction," the researchers noted.

"Furthermore, there are indications that threat actors are actively recruiting AI experts to develop their own custom uncensored LLMs tailored to specific needs and attack vectors.“

Most in the cybersecurity field will be familiar with the idea that AI is ‘lowering the barriers of entry’ for cybercriminals, which can certainly be seen here.

If all it takes is asking a pre-existing chatbot a few well-phrased questions, then it’s pretty safe to assume that cybercrime might become a lot more common in the coming months and years.

You might also like

• Take a look at our picks for the best malware removal software around
• Check out our choice for the best AI tools
• Identity fraud attacks using AI are fooling biometric security systems

• Most people are polite to AI even when they’re frustrated, Tidio survey finds
• Cursing at chatbots doesn’t stop people from saying thank you
• Politeness toward AI feels normal even when the tech fails often

Are you someone who always says “please” and “thank you” to AI tools, or do you let rip when it doesn’t get what you mean or give you what you want?

New research has suggested users of AI chatbots can be both polite and rude, often in the same conversation.

A new report from Tidio found most people have cursed at a chatbot at least once, but that hasn’t stopped them from saying “thank you” afterward, highlighting the contradiction in how people interact with AI.

Avoiding AI chatbots

Nearly 70% of users admitted to swearing at chatbots out of frustration, but the service we’re getting from our future AI overlords can’t be that bad since 75% said they were satisfied with their most recent interaction.

Although people can be quick to vent when things don’t go their way (something human customer service workers know all too well), politeness still seems to be a habit people carry into conversations with AI, even when the experience isn’t totally smooth.

A previous Future study found 67% of Americans and 71% of Brits are nice to AI. That includes saying “please,” “thank you,” and even apologizing to digital assistants like ChatGPT or smart speakers.

It may sound silly, but some people say being respectful helps them get better answers. TechRadar’s Becca Caddy tested this by removing polite words from her prompts to ChatGPT and said the quality of responses dropped.

Even OpenAI’s CEO, Sam Altman, weighed in on the cost of politeness. When asked how much it costs to power responses to polite prompts, he replied, “Tens of millions of dollars well spent.”

Although AI is now part of our daily lives, a good portion of people aren’t wholly convinced by it.

Around 30% of respondents to Tidio’s survey said they’d prefer to wait for a human even if a chatbot is ready to respond. And 26% said they’d rather trust a Magic 8-Ball than AI support.

Some users are even willing to pay to avoid it. About 11% would spend extra just to talk to a human.

Yet in practice, most people are prepared to use AI for help with basic tasks. The most common purposes include technical support, general questions, billing issues, and product information.

You might also like

• These are the best AI image generator tools available to use (and swear at) right now
• Are you polite to ChatGPT? Here’s where you rank among AI chatbot users
• I don’t understand why people say please and thank you to AI

AI agents are fundamentally reshaping businesses and how work gets done by augmenting them with autonomous, context-aware execution. If you’re not leveraging AI agents, here’s why you should.

Recognizing that managing AI agents is becoming an essential skill in the workforce, I predict that before 2026, every person at my 1,000+-person company will be using an agent on a daily basis. AI agents are evolving businesses rapidly, and while tech adoption rates can be slower for some organizations, the momentum and interest in agentic AI is building rapidly and proving its business value.

As organizations continue to grapple with complexity, speed, and the pressure to do more with less resources and staff, AI agents offer a path to operational agility: automating routine decisions, surfacing real-time insights, and accelerating strategic outcomes. This shift marks more than a ‘tech upgrade’—it’s a redefinition of the business operating model, where the ability to harness intelligent, data-driven agents will distinguish tomorrow’s leaders from those stuck in yesterday’s workflows.

Real agents

First, what is a real agent and why is it different?

In sales and marketing, there is a lot of chatter about building agents that save time and allow companies to send personalized content at scale. That’s absolutely an efficiency that should be taking place, but we’re probably talking about a database trigger in Salesforce that kicks off IF-THEN logic that, in turn, uses an API call to ChatGPT for purposes of drafting the content.

That is not an agent.

By definition, an agent is more resourceful, proactive, and helpful, able to pursue goals, and achieve more results on behalf of employees than either a chatbot or conventional automation. Although the concept of autonomous agents has been around for years, we’re just getting to the point where the technology is becoming widespread, with tools for the creation of agents improving rapidly.

Rather than trying to give a formal definition of what constitutes an agent, let me describe what makes the best formula for purpose-driven AI agents. They possess:

1. Tools to search the web and social media, gather information, and provide data analysis. It’s not a simple report on findings that is helpful in this context, it is an analysis of the findings and a strategy to move forward. Investing in tools is important: think of it like sharpening your knife, if it’s dull, it’s not going to cut as intended. You need to structure the tools to be efficient and flexible so your agent can use them properly.

2. Knowledge, particularly knowledge of you and your goals, the expectation of your outcome for the role you sit in, your writing style, and how to be successful. Context is key, make sure the agent has the relevant knowledge to do its intended job. This could include embedding knowledge from sales decks, website and app data and customer call transcripts.

3. LLM vs LLM evaluation, to ensure reliability, the most effective AI agents will use one model to generate an output and a different model to critique it. For example, if you're relying on an AI agent to draft a report, this approach helps prevent mistakes or awkward phrasing that another reviewer—human or AI—might otherwise catch.

4. A Playbook so the agent learns standard protocols about your company's data and requirements. The playbook should be prescriptive and specific but also leave room for the agent to adapt and change as it gets more information and is able to perform better.

How AI agents are leading business transformation

Across industries, AI agents are beginning to take on specialized roles within business workflows, offering practical support in areas like SEO, sales, and market analysis. For example, some agents now generate pre-meeting briefs by pulling together public digital signals, company data, and CRM information—work that previously required extensive manual effort.

For example, I worked on an AI Meeting Prep Agent for salespeople that one customer told us gave him a complete briefing within seconds that would have taken him at least a half hour to do himself, if he could even find the time — and this is someone who meets with multiple customers and potential customers every day.

Other agents analyze competitive keyword trends to recommend SEO content strategies, or track sudden changes in search behavior to surface emerging market shifts, providing more depth and speed of analysis than would be possible otherwise.

In sales, agents can be used to craft personalized outreach based on real-time data, helping teams engage prospects with greater relevance. Rather than replacing teams, these agents handle the groundwork—searching, summarizing, and connecting data—so people can spend more time making strategic decisions and less time on prep work.

Results

The result isn’t just increased efficiency, it’s business transformation. These agents free up talent from information-gathering and task repetition, enabling teams to focus on high-impact work: crafting strategy, building relationships, and driving innovation.

As these agentic workflows become embedded across functions, companies gain a more adaptive, data-responsive operating model—one that scales insight, improves agility, and accelerates decision-making across the board.

In short, AI agents don’t replace teams—they amplify them, creating a multiplier effect that turns data into direction and strategy into execution.

Technology is moving faster than ever, and now is the time to be an innovator, set your brand apart from the rest and stay ahead of the curve.

We list the best client management software.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro

• Chatbots are filling classrooms with fake students
• These chatbots are costing colleges millions in fraudulent loans
• AI is fueling a huge increase in identity theft

The rise of Artificial Intelligence has had a transformative impact on classrooms and college campuses across the world, but a new report from the Associated Press claims chatbots are now posing as students in order to collect financial aid, and are even “attending” online classes - and taking spaces from real students.

Analysis of fraud reports has uncovered a whopping 1.2 million fraudulent applications in California alone, leading to 223,000 suspected fake enrolments. California in particular is a target because of the large number of community colleges.

If your credentials have been compromised in a data breach, take advantage of the best identity theft protection software available to make sure you're protected against this type of attack.

Unrecoverable losses

These “ghost students” don’t just apply for a loan and hope for a payout, though. Chatbots have been deployed to online classrooms, where they stay “just long enough” to collect the financial aid check from the college.

Some lecturers even report having online classrooms with barely any real students as bots have filled the class to the enrollment limit.

According to researchers, criminals stole at least $11 million in local, state, and federal financial aid that ‘cannot be recovered,’ just from California colleges in the last year, so the problem is widespread.

This isn’t the first time we’ve heard of AI scams fueling a rise in identity theft and fraud, with hackers seen ‘leveling up’ their tactics, with more sophisticated attacks leveraged at a higher frequency than ever before.

Many of these scammers will use information obtained from data breaches, such as names, addresses, ages, and social security numbers. With this data they will apply for grants and loans using real credentials.

In March 2025, over 300 people were fired from the Federal Student Aid Office by the Trump administration, and the department’s fraud investigation unit has lost over 20% of its staff through attrition and retirements since October 2025.

You might also like

• Take a look at our picks for the best malware removal software around
• Check out our choice for AI tools
• Identity fraud attacks using AI are fooling biometric security systems