Angst vor dem Risiko ist eine Bremse: Löse sie und fang einfach an | Umair Zaffar
Umair Zaffar October 22, 2026 0 Comments

EU AI Act Article 12: Record-Keeping, Implemented in Practice

EU AI Act Article 12 requires a high-risk AI system to be able to record events automatically over its entire lifetime. Our answer is an evidence log written per decision by the gateway, append-only, that an auditor can read nine months later without asking the developer.

Article 12 is one of the shortest articles in Chapter III and the one audits fail on most often. Almost every system logs. Almost none in a way that lets you reconstruct a single decision afterwards. The article addresses the provider: it has to build the system so that it can log. What happens to the logs afterwards is governed by Article 19 for providers and Art. 26(6) for deployers. For engineers, Article 12 is a design decision for day one, not for the audit date.

The EU AI Act has been in force since 1 August 2024. The prohibitions and the AI literacy duty have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 12 belongs to these high-risk obligations. The Omnibus moved deadlines; it removed no obligation.

What EU AI Act Article 12 requires

Three paragraphs, and only the third contains a list of fields, for a single class of systems.

  • Art. 12(1): High-risk AI systems must technically allow for the automatic recording of events (logs) over the lifetime of the system. The obligation reaches the provider through Article 16, but it is phrased as a property of the system.
  • Art. 12(2): Logging must ensure a level of traceability appropriate to the intended purpose of the system. It must capture events relevant for three purposes: (a) identifying situations in which the system may present a risk within the meaning of Art. 79(1) or in which a substantial modification occurs, (b) facilitating post-market monitoring under Article 72, and (c) the deployer’s monitoring of operation under Art. 26(5).
  • Art. 12(3): For remote biometric identification systems under Annex III, point 1(a), there is a minimum content: start and end of each use, the reference database, the input data that led to a match, and the identity of the persons who verified the result under Art. 14(5).
  • What is missing: For every other high-risk system the article names no fields. The standard is “appropriate to the intended purpose”, and harmonised standards for it do not exist yet. The provider fills the gap itself and justifies the choice in the technical documentation under Article 11.
  • The connected articles: Article 19 obliges providers to keep the logs, 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 data protection law, provides otherwise. Financial institutions keep them as part of their documentation under financial services law. Art. 26(6) places the same duty on the deployer.

Article 12 does not ask for logs to exist somewhere. It asks for logs that serve the three purposes in paragraph 2. A web server log serves none of them.

The technical example

A customer-service assistant at an insurer. Agents receive questions like: is the water damage from the burst washing machine covered under this policy? The assistant prepares an answer, with supporting passages from the policy terms. The human sends, edits or discards it.

The data flow: the question arrives from the CRM together with the policy number. The application calls the policy gateway with a use-case ID, never the model. The gateway resolves the ID in the registry, checks the data zone, replaces policy number and names with pseudonyms and scrubs the free text. Retrieval runs over the policy wordings, with access rights enforced at query time and provenance per chunk: document, version, effective from, effective to. The model receives the passages and a versioned prompt template. The answer passes a citation check, gets a support score, and the gateway returns, abstains or escalates.

Where does Article 12 bite? Honestly: a pure information assistant for policy questions is, on my reading, not a high-risk system under Annex III. It becomes one when its answers feed into claims decisions or into risk assessment and pricing in life and health insurance under Annex III, point 5(c). That boundary shifts quietly in projects, because the assistant works well and the next team wants it for the next process. So the evidence log is mandatory for every use case in our registry, regardless of risk class. Retrofitting is not an option: last year’s decisions are gone.

The case everything hinges on: nine months later a customer complains. Service told him the damage was covered, now the claim is refused. Which version of the policy terms did the assistant read? Which model, which prompt, which gateway rules applied that day? Did the agent change the answer? If the system cannot answer that, Article 12 is not met, no matter how many gigabytes of logs sit on the disk.

How we implement it

At SIFAMO the evidence log is a component of the governance layer, next to the registry, the policy gateway, tiered oversight and monitoring. Four decisions shape the design.

The gateway writes, not the application

Every model call passes through the gateway, so the gateway writes the evidence. The business application cannot forget the logging, cannot switch it off and cannot half-build it. There is one schema for all use cases, and a bug in the application does not damage the evidence.

One record per decision

We log per decision, not per HTTP request. A decision here means: one question, one prepared answer, one action by a human. Simplified, the record looks like this:

{
  "decision_id": "svc-rag-2026-11-19-00018231",
  "timestamp": "2026-11-19T10:14:52Z",
  "use_case_id": "DE-V-2026-0077",
  "registry_version": 6,
  "input_hash": "sha256:9f3c...",
  "payload_ref": "gray://svc-rag/2026/11/00018231",
  "model": {"provider": "hosted LLM", "version": "2026-09-r2"},
  "prompt_version": "service-policy-v9",
  "retrieval": [
    {"doc_id": "AVB-HR-2024", "version": "3.2", "chunk": 41,
     "hash": "sha256:1b7a...", "effective_from": "2024-07-01",
     "effective_to": null, "access_rule": "role:service-agent"}
  ],
  "policy_version": "gw-policy-2026-10-14",
  "output_hash": "sha256:c0de...",
  "output_ref": "gray://svc-rag/2026/11/00018231/out",
  "signals": {"support_score": 0.91, "citation_check": "passed",
              "abstain": false, "escalation": null},
  "reviewer": {"action": "edited", "role": "service-agent",
               "actor_ref": "pseud:ag-5531", "timestamp": "2026-11-19T10:15:40Z"},
  "prev_hash": "sha256:77e1..."
}

The input hash and the payload reference separate proof and content: the hash stays verifiable, the pseudonymised content lives in the gray data zone and can be deleted on its own schedule. Model and prompt version say what was running that day, the sources with document versions which policy terms were read, the policy version which rule the gateway applied. The signals serve Art. 12(2)(a): support score, citation check, abstention, escalation reason, from which monitoring reads drift and risk. The reviewer action is the oversight under Article 14, as a pseudonym with a timestamp.

Append-only with tamper evidence

The log knows one operation: append. Every record carries the hash of its predecessor, forming a chain per use case. Once a day the latest hash is signed and stored in a system engineering cannot write to, usually held by risk or compliance. The storage is configured write-once, and only the retention job has delete rights. Tampering is not impossible, but it is detectable.

Retention by purpose, not by storage cost

The retention period sits in the registry record, signed by the business owner, risk and IT. The six months in Article 19 and Art. 26(6) are our floor, not our target. At insurers and banks we align the retention of the decision record with the retention of the related business records under financial services law, as Art. 19(2) envisages for financial institutions. The pseudonymised content follows a shorter period derived from data protection. Separate storage is what makes that possible.

The test: nine months later

Before go-live we draw random decision IDs from the staging environment and hand them to a person in the auditor role who did not build the system. They get read access to the log and the document store, nothing else. Can they explain why this answer came out this way, which sources were read and what the human did with it, without asking an engineer? If not, the log is incomplete and the gate stays shut. The test is part of the quality gates and is repeated at every substantial change. In operation the auditor sees a read-only view per decision ID with chain verification and a pointer to the registry version in force at the time.

Where the article leaves room, we also log the cases where nothing happened, meaning abstentions and refused requests, because that is exactly where the signals for Art. 79(1) live, and every resolution of a pseudonym as an event of its own. How we introduce this in projects is under AI Governance.

Typical mistakes

  • The application logs itself. Every team builds its own format, the logs live in the application database and vanish with the next migration.
  • Only the answer is stored. Without the retrieved sources with document versions and without the prompt version, the answer cannot be explained. You know what was said, but not why.
  • Plain text in the log. Names, policy numbers and health details end up in the log, the log becomes the biggest data protection problem of the system and gets deleted for that reason. Article 12 is breached.
  • No versions. The log says “model” and “prompt”, not which ones. Models change every few months, prompts every week.

Article 12 checklist

  • Every use case in the registry has a defined log schema and a justified retention period.
  • The evidence is written by the gateway, not by the application, and cannot be switched off from there.
  • Per decision the following is stored: input hash, payload reference, model and prompt version, sources with document versions, policy version, output, signals, reviewer action.
  • The log contains no real names; pseudonyms are resolved separately and every resolution is logged.
  • The log is append-only, with a hash chain and an anchor outside engineering access.
  • Abstentions, escalations and refused requests are recorded as events and read by monitoring.
  • A person in the auditor role has reconstructed random decisions before go-live without help from the developers.
  • The description of the logging capability is part of the technical documentation under Article 11.

Open questions

  • What is “appropriate”? The harmonised standard on logging and traceability has not been published yet. Until then our schema is a reasoned choice, not a presumption of conformity.
  • Logging versus data minimisation. Reconstructing a decision about a person is itself processing of personal data. Hashes and pseudonyms soften this but do not resolve it. Article 12 and Article 19 point to data protection law without settling the conflict.
  • Reconstructing is not reproducing. Language models are not deterministic, the same input does not yield the same answer. Our log explains what happened, not why the model internally decided that way. Whether interpretability research can ever put that into a log is open.

Frequently asked questions

Does this apply to us as deployers too?

Article 12 addresses the provider who builds the system. As a deployer you are caught by Art. 26(6): keep the logs, to the extent they are under your control, for at least six months, and under Art. 26(5) monitor operation with them. Anyone who substantially modifies a system or puts it into service under their own name can become a provider under Article 25. Then Article 12 applies directly.

Is the model vendor’s logging enough?

No. The provider of a hosted model logs its API calls, not your decision. It knows neither the document versions that were read nor the policy nor the human’s reaction. Those logs are also not under your control. The evidence has to be created on your side.

How long do we have to keep the logs?

At least six months under Article 19 and Art. 26(6), longer where the purpose or other law demands it. I consider six months too short in regulated processes, because complaints and audits arrive later. The period belongs in the registry record, with a justification.

This post belongs to my series EU AI Act in practice. Here you find all articles at a glance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment