Article 12 of the EU AI Act requires high-risk AI systems to record events automatically over their lifetime. The logs exist so problems can be found, traced and fixed, but a log nobody can read protects no one. Staff who oversee agents must be trained to find a decision, reconstruct it, and act on what it shows.
What does Article 12 require?
Article 12 sits in the Act's requirements for high-risk AI systems. In substance, such systems must technically allow events to be recorded automatically over their lifetime. Logging must support three things: identifying situations where the system may present a risk or has been substantially modified, supporting the provider's monitoring after the system is on the market, and supporting the deployer's monitoring of the system in operation.
Article 12 is primarily a duty on providers: they must build the logging in. Deployers pick it up through Article 26, which requires them to keep the logs generated under their control for an appropriate period and to monitor the system's operation. Article 14 requires human overseers to be able to interpret the system's output, and Article 4 requires sufficient AI literacy among staff. Put together, the logs must exist, be kept, and be understood by the people who rely on them.
For the logging requirement itself, see Article 12 record-keeping and logging what an AI agent does, and for whether it applies to your agent, does Article 12 require AI agents to keep logs.
Who does it apply to?
Providers and deployers of high-risk AI systems, such as those used for creditworthiness assessment and credit scoring of individuals, or risk assessment and pricing in life and health insurance, among the uses the Act lists. An agent used for such tasks is a high-risk system for that purpose. Agents used elsewhere may not be caught by Article 12, but most organisations keep the same record for every agent, because the reasons to read it do not change.
On the deployer side, the people who need to read logs are:
- reviewers, who need to see why an action was held before deciding it;
- team leads, who investigate errors and complaints;
- risk and compliance staff, who monitor operation and report incidents;
- internal audit, who reconstruct decisions independently.
What does a governed agent's record contain?
On a governed platform, every verdict is filed in an append-only record. A useful entry answers six questions:
- What was attempted? The capability, the action and its inputs.
- At what depth? The depth the agent held for that capability at that moment.
- On what facts? Each fact used, with its provenance: sealed, signed, declared or external.
- What was the verdict? Allowed, held, refused or withheld.
- Which bound decided? Role scope, certificate, sources, the deployer's own limits, or the platform guardrails. The lowest bound wins and is named.
- Who acted? For a held action, the named person who approved or declined it, and their reason.
The detail of what must be captured is set out in what an AI agent's record must capture for Article 12.
What must staff learn to do with the logs?
Reading a log is a skill, and it differs by role.
A reviewer should be able to open the entry behind a held action and see, in under a minute, the facts and the bound that held it. If the deciding fact is external, such as an emailed invoice, they should know it is the weakest class and cannot on its own authorise a payment.
A team lead investigating a complaint should be able to find every entry about one customer across a period, follow the chain from first action to final outcome, and identify where a person approved something they should have questioned.
A risk analyst should read the record in aggregate: a rise in refusals after a source change, a cluster of held actions on one capability, an agent whose certificate lapsed and dropped it to depth 1. These are signals that the system may present a risk, which is exactly what Article 12 logging exists to surface.
An auditor should be able to reconstruct any decision without help from the provider.
Teaching method matters. The most effective approach is to train on real entries from the organisation's own agents, with exercises that start from a complaint and end at the decisive entry. How to train staff to read an AI agent's decision record sets this out in detail.
How do you meet Article 12 and its people duties?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Automatic logging over the lifetime (provider) | Platform files every verdict in an append-only record | The record itself, with integrity checks |
| Deployer keeps logs under its control (Article 26) | Retention set and access controlled by role | Retention configuration; access logs |
| Deployer monitors operation (Article 26) | Risk staff trained to read the record in aggregate and escalate signals | Monitoring reports; escalations raised |
| Overseers can interpret output (Article 14) | Reviewers trained to read the entry behind each held action | Training records; review decisions with reasons |
| Staff AI literacy (Article 4) | Log-reading included in role-based training | A training record per person, dated before access |
What will a supervisor or auditor ask?
Expect a reconstruction test. The examiner picks a decision, perhaps from a complaint, and asks you to show what the agent did, on what facts, under what bound, and who approved it. They then ask who on your staff can do that, and how they were trained. They may ask what your monitoring found last quarter and what you did about it, and how long logs are kept and why. A team that needs the provider on the phone to answer has a training gap.
What does Article 12 not cover?
It does not decide how long a deployer must keep logs in every case; other law, including data protection and sector rules, also bears on retention. It does not prescribe log format beyond what the Act requires. It does not apply to systems that are not high-risk. It does not require staff training on its own; that comes from Articles 4, 14 and 26. And logs containing personal data remain subject to the GDPR.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where the Academy fits
Pryme Intelligence Academy teaches log-reading as part of reading the workspace and handling what goes wrong, using labs that run in a free Build workspace so learners work with real records rather than screenshots. Our guide to AI literacy and certification for people who work with agents shows where these skills sit for each role.
Questions people ask
Can staff edit or delete log entries?
No. An append-only record cannot be altered, which is what makes it trustworthy. Corrections are made by adding new entries that reference the original.
Should every reviewer be able to read the full record?
Reviewers need the entries behind the actions they decide. Wider access should follow role, both for privacy and to keep the record focused.
How do logs help when an agent's certificate lapses?
The record shows the lapse, the drop to depth 1, and every action held or refused afterwards. That lets the team see what manual work the lapse created and confirm nothing ran at a depth the agent no longer held.