Article 12 of the EU AI Act requires high-risk AI systems to log events automatically so their operation can be traced. For an agent that acts, that is necessary but not sufficient. A log shows what happened and when, rarely whether the agent was allowed to do it or which limit decided. Agents need a decision record on top.
What does Article 12 require?
In substance, Article 12 requires high-risk AI systems to be technically capable of automatically recording events over their lifetime. The purpose is traceability: logs should allow the system's functioning to be followed, help identify situations that may present a risk, support monitoring after the system is placed on the market and support the deployer's own monitoring. Under Article 26, deployers of high-risk systems must also keep the logs generated under their control. Our page on Article 12 record-keeping and logging what an AI agent does covers the text in more detail, and our explainer on whether Article 12 requires AI agents to keep logs deals with scope.
Two points matter for what follows. Article 12 is a design duty about the system's capability to log. And it applies to high-risk systems; many agents will not fall in that category, yet their decisions still need to be defensible.
Why is an event log not a decision record?
An event log records what happened in the system: inputs received, outputs produced, calls made, timestamps, errors. That is valuable for tracing the system's functioning. It answers "what did it do?"
A decision record answers different questions: what was the agent asked, what did it propose, what did it rely on, which limits applied, which one decided the outcome, and who acted next. These are the questions a supervisor, an auditor or a complaining customer will ask. An event log may contain fragments of the answers, scattered across many entries, but it was not built to answer them and often cannot.
| Question | Event log | Decision record |
|---|---|---|
| What did the agent do and when? | Yes | Yes |
| What was it asked to do, and by whom? | Sometimes, as raw input | Yes, as a structured request |
| What facts did it rely on, and from which sources? | Rarely | Yes, with provenance |
| Which limits did the action meet? | No | Yes: role, certificate, sources, customer limits, platform rules |
| Which limit decided the verdict? | No | Yes |
| Was it allowed, held, refused or withheld? | Sometimes, as an outcome | Yes, as a named verdict |
| Who approved, edited or refused it next? | Rarely | Yes, as a named person |
Why do agents need the extra layer more than other AI systems?
Because they act. A classification model produces a score, and a person or another system decides what to do with it. The decision sits outside the model and is usually recorded by whatever system made it. An agent, by contrast, proposes and often carries out actions itself: sending, updating, paying, closing. The decision and the action happen in the same place.
That makes the agent's authority the central question. Was this agent permitted to issue this refund, at this amount, for this customer, on these facts? An event log of the refund being issued does not answer that. Only a record of the bounds checked and the verdict reached does.
Agents also chain actions. One request can lead to a series of tool calls, each of which changes something. If only the final outcome is logged, the intermediate decisions, any of which might have been outside the agent's authority, disappear.
Is logging the agent's prompts and responses enough?
No. Capturing the full text of every prompt and response is sometimes proposed as the answer, on the reasoning that everything the agent saw and said is then on file. It produces a large volume of text that is hard to search and still does not say which limits applied or why an action was allowed. The agent's instructions are not a control, as our explainer on why a system prompt is not enough to control an AI agent argues, so a record of them is not a record of control either.
What should the decision record capture?
At minimum, for every action the agent proposes: the request, the proposed action, the sources relied on and their provenance, the bounds checked, the deciding bound, the verdict and the next actor. Our page on what an AI agent's Article 12 record must capture lists fields in more detail.
The record must be append-only, so entries cannot be edited after the fact, and retained for as long as decisions could be questioned. It should sit alongside the event log, not replace it. Engineers need the event log to diagnose the system. Supervisors, auditors and customer-facing staff need the decision record to judge the decisions.
What will a supervisor ask?
Expect a supervisor reviewing a high-risk deployment to ask whether logs exist and are kept, as Article 12 and Article 26 require. Then expect the harder questions: for a sample of decisions, show what the agent was allowed to do, what it did, and who oversaw it. Show how human oversight worked in practice. Show what happened when the agent tried to act outside its limits. A deployer with only event logs will struggle with all of these.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What is the trade-off?
Building a decision record means structuring every action's governance data and storing it alongside the logs. That is design work, and the record grows with every action. Some of the information it holds, such as customer facts, carries data protection obligations of its own.
The alternative is a log that satisfies a narrow reading of Article 12 and leaves you unable to defend any individual decision. For agents, that is the wrong place to stop. Check whether your agents' records can answer, for any action, which limit decided it and who acted next. If they cannot, add a decision record next to the log. The wider mapping of rules to controls is in our guide to AI agent regulation, mapped to controls.
Where Pryme fits
Pryme files every verdict in an append-only record: the request, the proposed action, the sources and their provenance class, each of the five bounds checked, the bound that decided and the person who acted next. Event logs sit alongside it for engineering use, so deployers have both the trace and the decision.
Questions people ask
Does Article 12 apply to every AI agent?
No. It applies to high-risk AI systems as defined in the Act. Many agents will fall outside that category, though a decision record is still good practice for any agent that takes consequential actions.
Who is responsible for logging, the provider or the deployer?
Providers must design high-risk systems so they can log automatically. Deployers must keep the logs generated under their control. Responsibility for the decision record, which reflects the deployer's own bounds and approvers, sits largely with the deployer.
How does this relate to GDPR records of processing?
They serve different purposes. GDPR Article 30 records describe processing activities at a general level. The agent's decision record describes individual actions. Both may be needed, and the decision record itself contains personal data subject to GDPR.