EU AI Act Article 12 logging for AI agents: what it says, and when
A plain reading of Article 12, the retention rules that sit in Articles 19 and 26, and the date change from the Digital Omnibus. Every claim links to the text.
What Article 12 actually says
Article 12 of Regulation (EU) 2024/1689 (the AI Act) is titled “Record-keeping”. Paragraph 1 says high-risk AI systems “shall technically allow for the automatic recording of events (logs) over the lifetime of the system.” Paragraph 2 says the logging capability must enable recording of events relevant to (a) identifying situations that may present a risk or involve a substantial modification, (b) post-market monitoring, and (c) monitoring the operation of the system by deployers. Text: Article 12.
Paragraph 3 lists minimum log contents, but only for one category: the remote biometric identification systems in point 1(a) of Annex III. For those, logs must include the period of each use, the reference database checked, the input data that produced a match, and the identification of the natural persons who verified the results. For every other high-risk system, Article 12 sets the goal (traceability appropriate to the intended purpose) and leaves the content to the provider.
Who has to keep the logs, and for how long
- Providers keep the automatically generated logs “to the extent such logs are under their control” (Article 19(1)).
- Deployers keep them for a period appropriate to the intended purpose “of at least six months, unless provided otherwise in applicable Union or national law” (Article 26(6)).
The six-month figure is a retention floor in Articles 19 and 26, not in Article 12 itself. Many summaries, including a code comment in our own SDK, attach it to Article 12. That comment is inaccurate.
When it applies: the date changed
The original application date for high-risk obligations was 2 August 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026, moved it. The Council said the new application dates are 2 December 2027 for stand-alone high-risk AI systems and 2 August 2028 for high-risk AI systems embedded in products (Council press release, 29 June 2026). EUR-Lex shows the regulation as in force, with first entry into force on 27 July 2026. The Article 12 page of the AI Act Explorer reflects the same dates, citing Article 113(c).
Correction. Until 4 November 2026 this site said Article 12 had been “in force since 2 August 2026”. That was wrong after the Omnibus. We corrected the banner and the affected pages. If you saw the old wording, treat it as superseded.
Other AI Act dates did not move with it. The Council release lists, for example, 2 December 2026 for the grace period on transparency solutions for AI-generated content. Do not read “delay” as “nothing applies yet”.
Is your agent a “high-risk AI system”?
This is the question that decides whether Article 12 is a legal duty or a good habit, and it is a legal question. High-risk status comes from Article 6 and Annex III (use cases such as employment, credit scoring, education, essential services, biometrics) or from being a safety component of a regulated product. A general agent that refunds orders or opens pull requests is not automatically high-risk. We cannot tell you whether yours is. Your counsel can.
What the records have to be, technically
The text asks for automatic recording over the system’s lifetime and traceability appropriate to the purpose. It does not name a data format, a hash function or a storage medium. That means a hash-chained audit log is one reasonable design, not a mandated one. What auditors tend to ask for in practice (our observation, not a legal requirement): what the agent proposed, which rule or person allowed it, what was executed, and a way to show nobody edited the record afterwards.
Where MnemoPay fits, and where it does not
The SDK’s AuditChain records events with ids, sequence numbers and timestamps and rolls them into a Merkle root; buildArticle12Bundle packages a finished mission as mission.json, events.json, events.csv, chain.txt and a checksummed manifest.json with a default retention field of six months. A signed demo bundle is downloadable and verifiable in your browser.
- It does give you tamper-evident event records and a verifier. See the runnable how-to.
- It does not make a system compliant, classify your system, or replace a conformity assessment. The repo’s own EU AI Act technical file is self-prepared; no auditor has reviewed it.
- It does not prove when a record was created unless you anchor or store the root somewhere you trust. No production MnemoPay deployment anchors on chain yet.
Sources
- Article 12, Record-keeping: artificialintelligenceact.eu/article/12
- Article 19, Automatically generated logs: artificialintelligenceact.eu/article/19
- Article 26, Obligations of deployers (para. 6): artificialintelligenceact.eu/article/26
- Regulation (EU) 2026/1744, Digital Omnibus on AI: EUR-Lex
- Council of the EU, 29 June 2026: final green light
Related
- Build a tamper-evident audit log for an AI agent (runnable)
- Audit trail vs observability logs
- MnemoPay compliance overview