DEV Community

Ventse
Ventse

Posted on Originally published at traceseal.io

EU AI Act Article 12 logging requirements: what AI agents must record, and what it leaves out

If an AI agent you build or run falls into one of the EU AI Act's high-risk categories, Article 12 requires it to keep logs. The article is short, it lists fields for only one kind of system, and the Commission gives no date for the harmonised standard that should fill in the detail. This piece covers what Article 12 requires, who keeps the logs and for how long, when it applies, how to turn it into an event list for an agent, and what it leaves out.

What Article 12 says

The core obligation is one sentence. Article 12(1) states: "High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system."

That makes logging a design requirement. The capability must exist, it must not depend on someone remembering to use it, and it must last as long as the system does.

Article 12(2) says what the logs are for. Logging capabilities "shall enable the recording of events relevant for":

  1. identifying situations that may result in the system presenting a risk within the meaning of Article 79(1), or in a substantial modification;
  2. "facilitating the post-market monitoring referred to in Article 72"; and
  3. "monitoring the operation of high-risk AI systems referred to in Article 26(5)", which is the deployer's duty to monitor the system on the basis of its instructions for use.

Only remote biometric identification systems get a minimum list of contents. Article 12(3) requires them to record, among other things, the "start date and time and end date and time of each use" and "the identification of the natural persons involved in the verification of the results".

For every other high-risk system, including any agent, the Act sets purposes and leaves the fields to the provider. That is the main practical difficulty with Article 12.

Does it apply to your agent?

An agent is not high-risk because it is an agent. It is high-risk if it falls under Article 6, most often because its intended use appears in Annex III, whose headings include critical infrastructure, education, employment, essential services, law enforcement, migration and the administration of justice.

Employment is where agents are most likely to land. Point 4(a) covers "AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates". An agent that screens applications fits that plainly. A coding agent working in your repository almost certainly does not.

Article 6(3) lets some Annex III systems fall outside the high-risk category, for example where they perform a narrow procedural task, but a system that profiles natural persons is always high-risk. Classification is a legal assessment, and this piece is not legal advice.

When it applies

The Digital Omnibus on AI moved the high-risk dates. The European Commission's AI Act page says the Act "became applicable on 2 August 2026, with some exceptions", that rules for high-risk use cases "will apply from 2 December 2027", and that "For systems integrated into products such as lifts or toys, the rules will apply from 2 August 2028."

For an agent whose logging today is whatever its framework prints to standard output, a little over a year is not long.

Who keeps the logs, and for how long

Article 12 says what the system must be able to record. Two other articles say who keeps the result.

Article 19 requires providers to keep the automatically generated logs, to the extent they are under their control, for a period appropriate to the intended purpose and of at least six months, unless Union or national law, in particular data protection law, provides otherwise.

Article 26(6) puts the same duty on deployers: they "shall keep the logs automatically generated by that high-risk AI system to the extent such logs are under their control", for "at least six months, unless provided otherwise in applicable Union or national law".

"Under their control" matters for agents. One run can touch a model provider, an agent framework, tool servers and the deployer's own infrastructure, each holding part of the record. Each party keeps its own part, so nobody necessarily holds the whole run. Deciding where the complete record lives is part of meeting the obligation.

Turning the purposes into events

Because Article 12(2) gives purposes rather than fields, ask for each purpose what you would need to reconstruct afterwards. For an agent, we suggest the list below. It is our reading, not text from the Act.

  • Each run. Start and end time, agent and model versions, the instructions and configuration in force, and what started it. This mirrors the per-use timing Article 12(3) asks of biometric systems.
  • Each tool call. Tool, arguments, result or error, exit status. In our experience, the risk in an agent sits mostly in what it does, not in what it writes.
  • Each human decision. Who approved, rejected or overrode what, and when. Our Article 14 piece covers what an oversight record should show.
  • Changes to the system. New models, prompts, tools or permissions. Article 12(2)(a) names substantial modification, and an agent's configuration changes more often than its code.
  • Refusals and failures. Blocked actions, sandbox denials, timeouts. Post-market monitoring needs near misses as well as incidents.

Arguments and results often contain personal data. One option is to log hashes of the content and keep the content under separate, stricter access. Our audit trail piece covers storage and retention in more detail.

What Article 12 does not say

Article 12 says logs must exist and what they are for. It says nothing about protecting them from alteration or about how anyone outside the organisation could check them. A log the operator can quietly edit still satisfies the wording.

That matters once a log is used for more than debugging. If a candidate disputes an agent's decision, or an authority asks how an incident happened, the log is the operator's account of its own system.

Harmonised standards may say more. The Commission's standardisation page lists record keeping among the ten areas CEN and CENELEC were asked to cover, and says companies that apply harmonised standards "are presumed to be compliant with the legal requirements." It gives no timeline for the remaining deliverables, so plan without it.

How to read this: Article 12 asks whether you can reconstruct what a high-risk system did. Whether anyone else should believe your reconstruction is a separate question, and the text leaves it to you.

Where signed records fit, and where they do not

Signed receipts address that second question. The Traceseal receipt specification defines a JSON record of a run that holds SHA-256 hashes of inputs and outputs and is signed with Ed25519, so a third party can check it from the file alone and detect any change to the signed content.

The limits are real, and the specification states them. A receipt is "an attestation by the operator, not a proof of execution". It records what the operator claims the sandbox was, not that the sandbox enforced it. Without the operator's audit log or an earlier receipt, a verifier "cannot detect replays". Hashes of guessable values can be reversed by guessing. And a receipt says nothing about events that never reached the signer: signing part of a run produces a complete-looking record of an incomplete one.

A signature also fixes content, not time. For that you need something outside the operator, such as a trusted timestamp or a witnessed transparency log.

Article 12 requires none of this, and adopting it will not make a system compliant by itself. What it changes is the answer to the question an auditor will eventually ask, which is not "do you have logs?" but "why should we believe them?"

What to take from this

  • Check classification first. Being an agent does not make a system high-risk. Recruitment and other Annex III uses can.
  • The dates have moved. 2 December 2027 for Annex III systems, 2 August 2028 for systems in regulated products.
  • Purposes, not fields. Write down which event serves which purpose in Article 12(2).
  • Six months minimum, for both parties. Agree who holds the complete record of a run.
  • Existence is not integrity. If your logs may be used in a dispute, decide now how an outsider would verify them.

For why a log on its own is weak evidence, start with our primer on signed execution receipts.

Originally published at traceseal.io.

Top comments (0)