EU AI Act Article 12: record-keeping and logs for AI agents

What Article 12 of the EU AI Act asks a high-risk system to log, how long Articles 19 and 26 say to keep it, when it applies, and what that means for an AI agent's actions.

· 9 min read

Article 12 of the EU AI Act says that high-risk AI systems “shall technically allow for the automatic recording of events (logs) over the lifetime of the system.” Article 19 makes providers keep those logs, and Article 26(6) makes deployers keep the ones under their control, in both cases for at least six months. For Annex III high-risk systems this applies from 2 December 2027.

Article 12 is short and does not list fields, except for one kind of biometric system. It names the purposes the logs must serve. This page goes through those purposes and what they mean for an AI agent whose output is actions.

What Article 12 says

12(1)
The system must be able to record events automatically, for as long as it is in use. Logging is a design requirement, not something added by hand.
12(2)(a)
The events recorded must help identify situations that may make the system present a risk within the meaning of Article 79(1), or that amount to a substantial modification.
12(2)(b)
They must help the provider's post-market monitoring under Article 72.
12(2)(c)
They must help the deployer monitor the system's operation under Article 26(5).
12(3)
Only for remote biometric identification (Annex III, point 1(a)): at least the period of each use, the reference database, the input that matched, and the people who verified the results.

Paragraph 2 asks for a level of traceability “appropriate to the intended purpose” of the system. That leaves the choice of events to whoever builds it, and puts the burden of showing they are enough on them too.

Who keeps the logs, and for how long

Providers (Art. 19)
Keep the logs from Article 12(1), to the extent they are under their control, for a period appropriate to the intended purpose and at least six months, unless Union or national law, in particular on personal data, says otherwise.
Deployers (Art. 26(6))
Keep the logs the system generates automatically, to the extent they are under their control, for a period appropriate to the intended purpose and at least six months, unless Union or national law says otherwise.
Deployers (Art. 26(5))
Monitor the system's operation on the basis of the instructions for use, tell the provider about risks, and suspend use when they see one.

If you build an agent on a model API and use it yourself for a high-risk purpose, you may well be both its provider and its deployer. Then both duties are yours.

When it applies

The Digital Omnibus on AI, Regulation (EU) 2026/1744, moved the high-risk dates in Article 113. Chapter III, which holds Articles 12, 19 and 26, now applies from 2 December 2027 for Annex III high-risk systems and from 2 August 2028 for high-risk systems under the Annex I product laws. The text of Article 12 itself did not change. The details, and what was proposed but not adopted, are in our article on Article 14.

Is your agent high-risk?

Most agents are not. The question is what the agent is used for, not how capable it is. An agent falls under Annex III when its job is in one of the listed areas, for example screening job applicants or deciding on promotions, assessing creditworthiness, pricing life or health insurance, deciding on public benefits, or triaging emergency calls. An agent that answers support tickets and issues small refunds usually is not, and Article 6(3) excludes some narrow or preparatory tasks even inside those areas.

If yours is not high-risk, Article 12 does not apply to it. The same log is still what you will want the first time an agent does something nobody can explain, and what a SOC 2 or ISO 27001 auditor will ask for.

Events to log for an agent

For a system whose output is actions, each purpose in Article 12(2) points at a few concrete events:

Risk situations (a)
Every tool call that changes something, with its redacted arguments, a hash of the full ones and its result status; every refused, denied or expired approval; errors, retries, loop and cost limits being hit; a person stopping the agent.
Substantial modification (a)
Changes to the model or its version, the system prompt, the tool list, the permissions and the approval rules, each with who made it and when.
Post-market monitoring (b)
Outcomes over time: how often people deny or change what the agent proposes, how often approved actions fail, per model version.
Deployer monitoring (c)
A view of the agent's runs and decisions that the deployer can read without the provider's help, and alerts on the events above.

The fields that make each tool call traceable, and a wrapper that writes them, are in audit logs for AI agent tool calls.

The people in the log

Article 12(3)(d) asks biometric systems to record “the identification of the natural persons involved in the verification of the results, as referred to in Article 14(5)”. It is only required there, but the idea carries: where Article 14 puts a person in the loop, the log should show which person, what they were shown and what they decided. Without it, the record cannot show that oversight took place.

An approval record such as VariantGate’s keeps exactly this part: which agent and key asked, the words the person saw, who decided, the reason, and when, with every entry sealed in a hash chain that can be checked without trusting us. It is not the system’s own log in the sense of Article 12. It sits beside it and covers the oversight, and the record docs say what it keeps.

Logs and personal data

Logs of a high-risk system usually describe people: the applicant, the borrower, the claimant. The retention periods say “unless provided otherwise” in Union or national law, and the GDPR still applies to what is in them. Keep the minimum that serves the three purposes, replace content with ids and hashes where you can, and decide the retention period on purpose rather than keeping everything forever.

This is an explanation, not legal advice. Whether a system is high-risk, and who its provider is, depend on facts only you know.

In short

  • Article 12 requires logging to be built into high-risk systems.
  • It names purposes, not fields: risk situations and substantial changes, post-market monitoring, and the deployer’s monitoring.
  • Providers and deployers each keep the logs under their control for at least six months.
  • For Annex III systems it applies from 2 December 2027; for Annex I systems from 2 August 2028.
  • For an agent, log the actions and the oversight: every call that changes something, and who allowed it.

Sources

  1. Regulation (EU) 2024/1689 (the AI Act): Articles 6, 12, 14, 19, 26 and 113, and Annex III. EUR-Lex.
  2. Regulation (EU) 2026/1744 (Digital Omnibus on AI), amending Article 113. EUR-Lex.