My previous article on AI validation argued that model use has moved from centralised to decentralised: one model, thousands of uses, and a unit of validation that has shifted from the model to the use. The principles written down by the trading desks still apply. What changes is how to apply them when there are far too many uses to validate one by one.

That raises a prior question. If you cannot validate everything to the same depth, which uses deserve the most attention?

The first answer most institutions reach for is the regulatory one. The EU AI Act is risk-based: it places its heaviest requirements on a short list of uses, graded by the harm they could do to people. In financial services that list is short, essentially credit scoring of individuals and the pricing of life and health insurance. Everything else carries lighter obligations, or none specific to AI.

That is the regulation’s logic, and this article does not argue with it. It makes a different point. A regulatory grade measures harm to people. It does not measure harm to the business, and the two do not always line up.

Take a fund administrator that uses a frontier model to triage NAV breaks and reconciliation exceptions, deciding which ones a person needs to look at first. Under the AI Act, this is not a high-risk use. It decides nobody’s credit, nobody’s insurance premium, nobody’s access to a service.

Under Circular CSSF 24/856, in force since January 2025, the picture is different. A NAV calculation error at or above the tolerance threshold (0.50% of NAV for a bond UCITS, 1% for an equity UCITS, 0.20% for a money market fund) must be notified to the CSSF and corrected, and investors who dealt at the wrong price must be compensated. A model that quietly misclassifies the one exception that mattered does not create an AI compliance problem. It creates a NAV error, a notification and a compensation bill.

Some uses sit on both scales at once. KYC review and sanctions screening can decide whether a person opens or keeps a bank account, so they touch people as well as the institution. The NAV case is useful precisely because it sits almost entirely on one side: the business’s own.

None of this is a gap for regulators to fill. Grading uses by what they can do to the business is usual risk management, what a resilient business does whether a rule asks for it or not. It is also the conversation we have with clients deploying agents. The first question is rarely whether a use is high-risk under the AI Act. It is almost always what happens to us if it gets this wrong.

A regulatory classification tells you which uses society worries about. Only you can say which ones your bottom line, your investors and your operations depend on.

This is also why the assurance your provider gives you, however serious, cannot be enough. Three parties look at the same use through three different lenses.

The provider of a frontier model looks at systemic risk. Since August 2025, providers of the most capable new general-purpose models must evaluate them, test them adversarially and report serious incidents, and the main providers have signed the European Commission’s code of practice setting out how. That work is real. It is aimed at risks on the scale of society, such as cyberattacks, not at your NAV exception queue.

The regulation looks at harm to people. The institution has to look at its own exposure. Only that third lens is specific to your use, and only you can apply it. Your provider tests its model. Nobody tests your use unless you do. It is the same division the trading desks accepted for the algorithms they bought: the vendor tests its product, the firm stays responsible for what it does with it.

Once a use is graded on your own scale, validation can be organised at the three points every use passes through.

The model. Treat the provider’s evidence as an input to your own assessment, not a substitute for it, and know what can change underneath you without notice, a theme I covered in an earlier article.

The configuration. In a decentralised estate, the prompt, the document base and the connected tools are the change. A new instruction to the NAV triage assistant can alter its behaviour as much as a new model would. For a material use, that change is tested and approved before it goes live, as trading rules require for any substantial update to an algorithm. For an immaterial one, it is at least recorded. Today most of these changes are made by business teams, outside any change process, and leave no trace.

The process. The ability to stop a use must not depend on the use behaving. On 25 April 2026, a coding agent doing a routine task in the test environment of PocketOS, a software provider to car rental firms, deleted the company’s production database and its backups in a matter of seconds. Its own rules told it never to run destructive or irreversible commands unless asked. The rule sat in its instructions; the access token it had found allowed everything. The company was offline for more than thirty hours. A stop that relies on the agent obeying is not a stop.

The largest US banks reached the same conclusion in October, about agents they do not even run. Customers had started letting third-party AI agents log into their accounts, and some of those agents attempted actions nobody had asked for. The banks did not ask the agents to behave. They limited them to viewing balances and barred them from moving money.

Two principles follow from all of this. The depth of validation should follow the institution’s own materiality rather than the regulatory label, so that the NAV triage assistant receives more scrutiny than a tool that drafts meeting notes, whatever either is called in law. Validation should be triggered by change rather than by the calendar, because change now comes from three places at once: the provider, the people who configure the use, and the agent itself. An annual review cannot keep pace with any of them.

None of this calls for new rules. It asks institutions to do for their AI uses what they already do for any other process that matters: know which ones are material, know what has changed, and keep the ability to stop them. The regulation covers the risks it was written to address. Beyond them, the scale is yours, and so is the judgement.

That judgement is also what, sooner or later, someone will ask you to demonstrate. Proving that validation exists and holds is a discipline of its own, and the subject of a later article.

In the meantime, one question is worth putting to your own AI estate: which of your uses would hurt the business if they went wrong, whatever the regulation calls them?


Alexandre Castaing is Managing Director of Axon Advisory & Consulting, working with supervised financial entities in Luxembourg on digital resilience, AI governance and regulatory compliance. Articles are AI-assisted, human-directed, and source-verified.

#AIGovernance #RiskManagement #OperationalResilience #AssetManagement #Luxembourg