Blog · ← All posts
EU AI Act Article 12 Logging: What Records High-Risk AI Systems Must Keep
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 8 min read
The short answer: Article 12 of the EU AI Act requires high-risk AI systems to be designed with logging capabilities that automatically record events over the system's lifetime, so that risks can be identified, post-market monitoring can work, and deployers can oversee operation. Article 19 sets the retention floor: logs kept for at least six months, longer where other law requires. The records that satisfy it are event-level records of functioning — what the system did, when, under what inputs in summary — not indiscriminate archives of full prompts and outputs.
Replay a recorded agent incident free → Step through a real recorded run in the replay view, or paste ledger records into the public verifier and get a VERIFIED/BROKEN verdict in seconds — no signup.
Of all the EU AI Act's technical requirements, logging is the one teams can start on today with total confidence about what 'good' looks like — because Article 12 is unusually concrete. High-risk systems must technically allow the automatic recording of events while operating. Not a policy. Not a runbook. A capability, built in, running by default.
This guide reads Article 12 alongside its neighbors — Article 19 on retention, Article 26 on deployer duties, Article 72 on post-market monitoring — and translates them into a logging design for AI agents: what events, in what form, kept how long, and usable by whom. It is an engineering reading, not legal advice; classification of your system is counsel's call.
What is Article 12 logging? the record-keeping capability Article 12 of Regulation (EU) 2024/1689 requires of high-risk AI systems: technical features allowing the automatic recording of events ('logs') over the system's lifetime, at a level appropriate to its intended purpose — sufficient to identify risks, support post-market monitoring, and let deployers monitor operation under Article 26.
The scope question comes first
Article 12 binds high-risk AI systems — those in the Act's Annex III use cases (among them certain employment, education, credit, law-enforcement, and essential-services contexts) and regulated products requiring third-party conformity assessment. A general marketing assistant is not automatically in scope; an agent that screens job applicants or evaluates creditworthiness very likely is, through its deployer's use.
Two practical consequences follow. First, providers of agent platforms cannot know every deployer's classification — so building Article 12-grade logging as the default posture is how a platform stays sellable into high-risk contexts. Second, deployers near the boundary should log as if in scope: the same records serve incident investigation, customer assurance, and every other audit regime, and retrofitting history is impossible.
What 'automatic' rules out
The Act's word is 'technically allow for the automatic recording'. Manual summaries written after incidents, optional debug flags, and logging that a routine configuration change can silence all fail the spirit and likely the letter. The recording must be a property of the system's operation: events are captured because the system ran, not because someone remembered to capture them.
For agents, that means instrumentation at the action layer: every tool call and consequential decision produces a record as a side effect of execution, written to storage the agent itself cannot alter. Agent Flight Recorder's model — the agent reports each action to an append-only, hash-chained ledger as it acts — is automatic in exactly this sense, and the chain adds what the Act assumes but does not spell out: the ability to show the records were not subsequently edited.
What events, at what grain
Article 12 ties the grain to purpose: logs appropriate to the intended purpose, enabling risk identification and monitoring. Translated to agents, the event set is the functioning record: each action's type and tool, timestamp, the inputs in summary or digest form, the outcome, and the version context (model, policy, system prompt) in force. For systems in Annex III contexts like biometric identification, the Act names specific event types — each use, each verification, input data occurrences — but the agent pattern generalizes: record decisions and actions, not raw content by default.
Full prompts and outputs deserve a deliberate decision, not a default. They routinely contain personal data, and GDPR's minimization principle pulls against warehousing them. The defensible design records structured event summaries in the long-retention ledger and keeps any full transcripts in a separate, short-retention store with strict access — a design that satisfies Article 12's purposes without manufacturing a privacy liability.
Retention: the six-month floor and its exceptions
Article 19 requires providers to keep automatically generated logs under their control for a period appropriate to the intended purpose — at least six months, unless other Union or national law provides otherwise. Deployers carry a parallel duty under Article 26(6) for logs under their control, also with the six-month floor. Sector law frequently goes longer — financial and health contexts especially — and the longest applicable period governs in practice.
Six months of event-level agent records is cheap; six months of full transcripts is a storage and privacy project. The retention math is another argument for the summary-ledger design: at a few hundred bytes per action record, even a busy agent's half-year of history is measured in gigabytes, small enough that 'appropriate to the purpose' never has to argue with the storage bill.
Logs are for someone: monitoring and post-market duties
Article 12 does not exist for its own sake. Its outputs feed the deployer's duty to monitor operation (Article 26), the provider's post-market monitoring system (Article 72), and serious-incident analysis and reporting (Article 73, with its short deadlines). Each consumer needs the same thing from the logs: a trustworthy, ordered, queryable account of what the system did, producible on a regulator's clock rather than an engineer's.
That producibility is where integrity pays. An incident report built from a chain-verified ledger — timeline cited by record, verification attached — carries a different weight with a market-surveillance authority than a spreadsheet assembled from application logs. Build the export path before you need it: serious-incident timelines are measured in days, and reconstruction-from-scratch does not fit inside them.
Five actions you must take today
Whether your classification is settled or pending, these five steps put an agent system in an Article 12-defensible posture:
- Classify honestly, log generously. Document your high-risk determination with counsel; build event-level automatic logging either way — the records serve every regime that will ever ask.
- Make recording a property of execution. Actions generate records as they happen, into append-only storage the agent cannot alter; no flags, no manual summaries.
- Record functioning, not content. Structured event summaries — action, tool, outcome, reasoning, versions — in the long-retention ledger; full transcripts, if kept at all, in a separate short-retention store.
- Set retention to the longest applicable clock. Six months is the floor, not the target; map sector and national overlays and keep the ledger for the longest.
- Build the incident export now. A chain-verified timeline, producible in hours, mapped to the fields a serious-incident report will demand.
Worked example: a hiring-screen agent prepared before classification closed
TalentBridge, a fictional HR-software provider, sells an agent that drafts first-pass screening summaries for high-volume roles — squarely in an Annex III employment context. Before its conformity work finished, the team instrumented the agent to record every action: 27,540 records in the first quarter of deployment at its pilot customer, each carrying the posting ID, the criteria version applied, the summary produced in digest form, and the human reviewer's decision recorded downstream.
When the deployer's works council asked how screening decisions could be reviewed for bias, the answer was already built: the ledger supported sampling any week's decisions, replaying the agent's steps against the criteria version in force, and verifying the whole quarter's chain — PASS, end to end — to show the history being analyzed was the history that occurred.
Retention was set at the deployer's sector standard, comfortably above the Article 19 floor, at a storage cost the team measured in single-digit gigabytes — because event summaries, not transcripts, were what the ledger kept.
FAQ
Does Article 12 apply to AI agents specifically?
It applies to high-risk AI systems, whatever their architecture. An agent used in an Annex III context — recruitment, credit evaluation, access to essential services — is a high-risk system in that use, and its actions are exactly the 'events' Article 12 contemplates. The agent form factor makes automatic logging more necessary, not less: an autonomous system's functioning cannot be reconstructed from user testimony.
Must we store full prompts and outputs to comply?
The Act requires logs appropriate to the intended purpose — supporting risk identification, monitoring, and oversight. It does not mandate full-content archives, and GDPR minimization argues against them. Event-level records of decisions and actions, with inputs in summary or digest, are the proportionate reading that practitioners converge on; document the choice in your technical file so the reasoning is auditable too.
How long must Article 12 logs be kept?
Article 19 sets the provider floor at six months for automatically generated logs under the provider's control, appropriate to the purpose and subject to longer periods under other law; deployers carry a matching six-month floor under Article 26(6). Sector rules — finance, health, employment record regimes — frequently extend this. Design retention as a configurable period per deployment, defaulting above the floor.
Who must be able to use the logs?
The deployer, for monitoring operation and fulfilling their Article 26 duties; the provider, for post-market monitoring under Article 72 and incident analysis under Article 73; and market-surveillance authorities on request. Practically: logs must be exportable in a form a third party can verify and read without your live system — one reason a hash-chained export with independent verification travels better than a dashboard login.
When do the logging obligations start applying?
The Act phases in: most high-risk obligations apply from August 2026 for Annex III systems (with some product-embedded systems later). But logging has a property the calendar cannot fix: history cannot be backfilled. Systems deployed before the application date will be asked for records from their operating lives; instrumenting now is the only way those records exist when the questions arrive.
Start logging your agents free → The free tier gives you the full hash-chained ledger for 1 agent and 1,000 actions a month, with chain verification and replay included.