EU AI Act
What Article 12 asks for
Record-keeping obligations tend to get read twice. Once by a lawyer, who says it's manageable, and once by an engineer, who wants to know what to actually build. This is the second reading.
The short version
The EU AI Act's record-keeping article asks that high-risk AI systems technically allow for the automatic recording of events over the system's lifetime, at a level appropriate to the intended purpose. For the high-risk uses listed in Annex III, such as credit scoring, hiring, and access to essential services, those obligations apply from December 2, 2027.
Two words in there do most of the work.
Automatic. Not "the team writes it down." The system produces the record as a consequence of operating, without anyone deciding to.
Over the lifetime. Not a snapshot of the last thirty days because that's what the retention policy happened to be.
What "appropriate" means for your system is a call for you and your advisers. Those two properties aren't really negotiable, and they're the ones people design around by accident.
Where agentic systems make it harder
A model that returns a score is comparatively easy to record. An agent that takes actions is not, for three reasons.
The action and the decision are separate events. Policy said the refund could proceed. Then the application issued it, or didn't, or issued a different amount. A record of the first without the second doesn't describe what your system did.
A person is often in the middle. Somebody approved the exception. Their decision is part of the operation of the system, which means it belongs in the same record rather than in an approvals inbox somewhere else.
The rules move. What was permitted in March isn't what's permitted in September. A record that doesn't capture which version applied can't tell you whether an action was allowed at the time it happened.
What that turns into, practically
If you're building against this rather than writing a policy about it, four things have to be true.
Scroll horizontally to see all columns.
| What has to be true | Why it matters |
|---|---|
| The record is a side effect of running | Written at the policy check, as part of the same operation. Nobody opts in, so nobody can forget. |
| The rule version travels with the decision | Not the policy library in general, but which rules were live at that moment. This is the one that gets skipped most often. |
| Human decisions live in the same record | Who was asked, what they were shown, what they decided, and what the system did afterward. Oversight you can't reconstruct is hard to evidence. |
| The reported outcome closes the loop | Your application says what it did. Intent and outcome can differ, and the gap between them is what an inquiry cares about. |
Whether a given obligation applies to your system is a call for you and your advisers; a record like this is what lets you reconstruct what happened either way.
See how logging and oversight connect to real decisions
Framework mappings tie requirements to the records behind an actual decision.