Become a TechRadar Insider
For enterprise technology leaders, the current debate around Europe’s AI rules can sound deceptively simple. If regulators delay some high-risk AI requirements or soften parts of the compliance burden, it can appear as though deployment should become easier, with fewer reporting layers and fewer obstacles between proof of concept and production.
However, it’s not that simple.
Solutions Lead at UnlikelyAI.
When visibility into high-risk AI use becomes weaker or slower, the risk does not disappear. It moves downstream, onto the organizations actually deploying the systems.
Brussels’ direction on implementation makes that especially clear: regulatory flexibility on paper does not translate into reduced accountability in practice. Businesses are still responsible for what their AI does.
That matters because most organizations are not deploying AI tools in a vacuum. They are using it in customer communications, operational workflows, compliance checks, document handling, claims processes and internal decision support, where outputs have real consequences and where “the model got it wrong” is not a defensible answer.
Boards, risk teams and operational owners will still need answers to the same basic questions regardless: why did the system produce this output, what shaped that decision, what happens when it is uncertain, and can its reasoning be reviewed after the fact?
The burden is shifting from compliance paperwork to operational proof
The old assumption was that regulation would tell businesses exactly what “responsible AI” looked like. In reality, many technology leaders are discovering that compliance is only part of the problem. The harder challenge is proving that an AI system is dependable enough to use in workflows where mistakes carry serious consequences.
Most of the AI now being deployed in enterprise settings is built on large language models (LLMs). These systems are powerful, but they are probabilistic by design: they generate the most likely next output based on patterns in data, rather than reasoning through a problem in a transparent, rule-bound way.
That makes them useful for drafting, summarizing and handling ambiguity, but much less suited to workflows where decisions need to be consistent, traceable and easy to justify after the fact.
This is why human-in-the-loop is often a weaker safeguard than it first appears. If the human reviewer is simply being asked to sense-check an output from a black-box model, one that cannot explain how it reached its answer, then the organization has not solved the trust problem.
It has just inserted a manual backstop into an unreliable process. This may reduce legal exposure in the short term, but it does not improve productivity, accountability or confidence. It also scales poorly, as human oversight of every AI output defeats the purpose of automation.
In practice, this means CIOs, CTOs and AI governance leads need to think less about whether a model looks impressive in a demo, and more about whether it can survive scrutiny in production.
What enterprise buyers should prioritize instead
There are four questions worth bringing into any procurement or deployment decision, and the answers point increasingly towards architectures that combine probabilistic and deterministic reasoning: what practitioners call neurosymbolic AI.
First, can the system explain how it arrived at an answer in a way a non-specialist reviewer can follow? Not just produce a plausible summary, but expose the logic, rules or constraints that shaped the outcome.
Second, does it know when not to answer? In high-stakes settings, a useful AI system is not one that always responds fluently. It is one that can recognize ambiguity, defer, escalate or say “don’t know” when confidence is too low. An LLM will rarely do this as it is optimized to respond, even when it doesn’t know the answer.
Third, can it be audited after the fact? If a regulator, customer or internal reviewer asks why a decision was made, teams need more than a confidence score and a generic disclaimer. They need a trail.
Fourth, is the architecture suited to the type of problem being solved? This is where neurosymbolic AI becomes directly relevant. Neural systems – LLMs – are powerful at pattern recognition and language flexibility. Symbolic systems are strong at rules, constraints, consistency and auditability.
When a spreadsheet calculates a result of a formula, nobody checks twice whether it may have hallucinated an alternative answer. That is the standard enterprises need from AI in regulated workflows.
Neurosymbolic AI combines both, using neural capability to interpret language and extract information, while applying symbolic reasoning to determine and explain outcomes. Organizations including Lloyds Banking Group are already piloting these approaches in regulated environments.
The real enterprise risk is opaque outputs, not regulation
For years, the industry has tended to treat transparency as something that can be added once a system has already been built – through disclosures, warnings or compliance dashboards. The reality of enterprise deployment is exposing the limits of that approach. If a system is opaque by design, no amount of paperwork will make it truly trustworthy.
That is why the current debate in Europe should matter to business technology leaders. Even if policymakers allow more time or reduce certain formal requirements, the underlying accountability does not go away.
The organizations that move most effectively from pilot to production will not be the ones taking the most permissive view of compliance. They will be the ones choosing architectures, controls and operating models that can stand up to scrutiny from the start.
If transparency obligations weaken, enterprises don’t escape accountability, they absorb it. The question is whether the systems they’ve deployed can withstand that scrutiny.
We've featured the best AI chatbot for business.
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