The Digital Operational Resilience Act has applied to EU financial entities since 17 January 2025. It does not mention AI coding agents anywhere, and that is precisely the problem: its ICT change-management expectations were drafted for a world in which a named person wrote each change and could be asked about it afterwards.
What the regulation actually asks for
DORA requires financial entities to maintain an ICT risk-management framework that includes documented procedures for changes to ICT systems — covering authorisation, testing, approval and the ability to reconstruct what was deployed. Supervisors read this as a records requirement: not “did you have a process” but “show me this change going through it.”
Nothing in that obliges you to know whether a model drafted the code. It does oblige you to be able to describe the change, justify it and identify who approved it.
Where an agent breaks the assumption
Change management assumes the approver and the author are different people, and that asking the author “why did you do it this way” is possible. When the author is a model:
The rationale is not in the commit. It was in a prompt and a response that no system retained. The commit message describes what changed, not what was asked for.
The reviewer may be the requester. One engineer prompts, reads the output and merges it. Segregation of duties has not formally broken, but it has thinned considerably.
Volume changes the shape of review. A reviewer who reads forty lines carefully and skims four hundred is behaving rationally; the control record cannot tell the two apart.
What a defensible answer looks like
An examiner asking about a specific production change wants four things: what was requested, what was produced, who accepted it, and evidence that the account has not been edited since. The first two are the ones most organisations cannot produce today, because the prompt and the response were never written down anywhere durable.
Capturing them is not a regulatory obligation in itself. It is the cheapest way to answer a question you will be asked whether or not anybody has written it into a rule yet.
Register the gap before somebody finds it
If AI-assisted development is in use on systems in scope, the honest position today is that the change record is partially reconstructible and the gap is known. That is an ordinary risk-register entry with an owner and a date. It is a much better place to be than discovering during an examination that nobody had considered the question.