Business leaders, governments, and tech giants are locked in a high-stakes race for AI supremacy, fueled by debates over the destructive power of superintelligence. In this sort of arena, it’s essential that enterprises stay grounded and resist the urge to start a race of their own, deploying ever-larger models simply for the sake of it.
AI/ML Engineering Lead at Beyond.
In the context of enterprise AI, there is no prize for launching the most powerful model first, building the highest number of agents, or forcing AI into every workflow. True success is defined by solving genuine business problems and delivering outcomes that justify the underlying cost and complexity.
While the prospect of a single, omnipotent large language model (LLM) is inspiring, today's business realities demand a more measured approach: every problem has a right tool, and often, that tool is a simple one.
AI doesn't have to mean LLMs
Before generative AI dominated the conversation, businesses were already using machine learning for recommendations, forecasting, fraud detection and estimating things such as the lifetime value of a customer.
We have had years to understand these technologies, including how to deploy and monitor them and make predictions at scale quickly and cheaply. An LLM may be capable of tackling some of the same problems, but that doesn't always make it the better option. More sophisticated models can cost considerably more to run and introduce variability that is unnecessary for a straightforward prediction.
Sometimes AI isn't necessary at all. If a process follows known rules and needs to produce the same result every time (moving data between systems on a schedule or triggering an action when a predefined condition is met, for example), conventional automation may do the job perfectly well.
Long before agents arrived, businesses handled these tasks with pipelines, schedulers, and rules. If that approach is reliable, fast, and inexpensive, adding an AI agent doesn't make the process more intelligent; it can simply make it more complex.
Generative AI earns its place when creativity or interpretation are required: understanding language or images, for example, or handling a process where the next step changes according to context. Agents can go further by using that understanding to interact with systems and take actions, opening up processes that were previously extremely difficult to automate.
The future is likely to combine conventional automation, traditional ML, generative AI and agents, rather than one replacing everything that came before it.
The AI honeymoon is ending
These distinctions matter more as enterprise AI moves beyond experimentation. A proof of concept can demonstrate what a model or agent is capable of, but production introduces running costs, security, monitoring and the work required to keep increasingly autonomous systems under control.
An engineer once gave me an analogy that captures this well. Imagine growing a tree that you need to maintain in a particular shape. When a branch grows in the wrong direction, you cut it.
Then another appears elsewhere. The difficult part is finding all the branches. Developing an agent can be similar. Obvious behaviors can be controlled relatively quickly, but real-world use exposes edge cases that weren't visible during development.
The more freedom a system has to interpret information and make decisions, the more important evaluation, monitoring and guardrails become. Control needs to extend beyond model behavior to the data a system can access, the actions it is authorized to take, the information it retains and the point at which a human can intervene.
There is also cost to consider. The most capable models can be expensive at scale, so additional capability needs to create enough additional value to justify it. As experimentation gives way to scrutiny of ROI, businesses will become less interested in whether something can use AI and more interested in whether AI actually leads to an improved commercial or operational outcome enough to justify the cost.
Let the data decide
Avoiding over-engineering starts with the business need and the data, rather than choosing the technology first.
Data engineers and analysts are crucial because they understand where information comes from, how reliable it is and what it can actually tell you. AI and ML engineers can put mathematical models on top, but sophistication cannot compensate for data that doesn't support the task.
The data and desired outcome will often point towards the appropriate approach. Fast, reliable predictions from structured numerical data may be better suited to traditional ML.
Language, images and other unstructured information can make generative AI valuable, while processes that require human level understanding to trigger actions across different systems may justify agents.
And sometimes conventional software will remain the best answer. Starting with “we need an agent” and then looking for a problem to give it is the wrong way around.
The right people matter as much as the technology
Increasingly capable AI has also created an assumption that businesses will inevitably need fewer engineers. AI is certainly changing productivity: experienced engineers can complete work faster with these tools, while junior engineers who learn to use them effectively can become productive more quickly.
But businesses still need people who understand what is happening underneath. I like to compare it with smartphones. Most of us use a phone effectively without being able to explain the hardware, operating system and networks behind it. That doesn't mean the underlying engineering skills have disappeared. Somebody still needs to understand those systems.
AI is likely to develop in a similar way. Many people will become highly capable users, while businesses will still need engineers with the deeper technical knowledge to build, control and fix the systems underneath.
That means continuing to invest in junior talent too. Senior engineers eventually move companies, change roles or retire. Organizations still need mentorship and knowledge transfer so today's juniors become tomorrow's senior engineers.
The opportunity isn't people or AI, but the two together. Capable engineers equipped with these tools can achieve considerably more than either could independently.
The next AI advantage will be restraint
The wider battle over superintelligence and increasingly powerful models will continue, but it shouldn't dictate how enterprises make technology decisions. To make this technical restraint viable in practice, engineering leaders must actively manage executive and board-level fear of missing out, turning top-down pressure for a visible ‘AI strategy’ into discipline, value-driven execution.
Some business problems will justify sophisticated AI tools. Agents will make previously impractical forms of automation possible, while generative AI can transform how organizations work with language, images and unstructured information. Others will be solved perfectly well by traditional ML, conventional automation or straightforward software.
Knowing the difference will become increasingly valuable. Competitive advantage will come from knowing where AI materially improves an outcome, where a simpler technology is more effective and how to move the right systems into productions securely and economically.
That requires good data, strong engineering talent and enough technical judgement to resist complexity simply because it is available. In the next phase of enterprise AI, restraint will not signal a lack of ambition, instead it will be a sign of technical maturity.
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