What is an AI Act evidence pack? (and what an auditor actually checks)
Most teams answer the AI Act question with a screenshot of a dashboard. An auditor cannot use that. An evidence pack is the opposite: a dated, scoped, signed file that answers who, what, which model, which data and who approved — and proves it has not changed since.
If you deploy or provide a high-risk AI system in the EU, the question you will hear in 2027 is not "are you compliant?" It is "show me." Procurement, customers, insurers, notified bodies and market surveillance authorities all ask the same thing in different words: give me a record I can check without trusting you.
That record is the evidence pack. It is not a policy PDF, not a dashboard export and not a legal opinion. It is a snapshot of one system over one period, with enough structure that a third party can follow it, and enough integrity that they can tell it was not edited afterwards.
The short answer
An evidence pack contains five things, in this order: what the system is and how it was classified under the AI Act; which controls apply and where each one stands; what the system actually did in the period (logs); who approved what (human oversight); and a cryptographic seal over the whole thing.
| Part | Question it answers | Typical artefact |
|---|---|---|
| Classification | Is this system prohibited, high-risk, limited or minimal? | Annex III reference, Article 5 screening, Article 50 triggers |
| Control status | Which obligations apply and what is still open? | Checklist per Article 9–15, 17, 72, 73 with owner and evidence pointer |
| Activity log | What did it do, with which model, on whose data? | Normalised event records with timestamps and model version |
| Human oversight | Who decided, when, and on what basis? | Approval records with approver identity and time |
| Integrity | How do I know this file is the file you issued? | Hash chain, Merkle root, pack hash, signature, verification page |
Why "we have logs" is not evidence
A log is a by-product. Evidence is a statement you are prepared to defend. The gap between the two is made of four questions that a plain log file cannot answer.
- 1Was the log complete for the period, or was it sampled after the incident? A pack states its period and seals every record inside it, so a gap is visible rather than arguable.
- 2Could the log have been edited? A database row with an UPDATE grant is not evidence. A hash chain is: each record commits to the previous one, so removing or changing a row breaks every hash after it.
- 3Who was accountable? Model output alone does not satisfy Article 14. The pack has to name the human who reviewed or approved the outcome, with a timestamp.
- 4Which model version produced this output? Models change. A record that says "GPT" without a version cannot be tied to the evaluation you ran six months earlier.
Hand the file to someone who does not trust you, with no access to your systems. If they can verify the seal, read the control status and see who approved what, it is an evidence pack. If they need your login, it is a dashboard.
The six questions an auditor asks
| Question | Where it is answered | What fails the check |
|---|---|---|
| What is the system and what does it decide? | System description and classification | A generic description that could be any chatbot |
| Which legal basis applies? | Article 5, Annex III or Article 50 determination | Calling an HR screening tool "minimal risk" without analysis |
| Which controls are in place? | Checklist with status and evidence pointer | All items marked "done" with no artefact behind them |
| What did it do in the period? | Event log with model and data categories | Only aggregated counts, no individual records |
| Who was in the loop? | Approval record with identity and timestamp | A policy that says "a human reviews outputs" |
| Has the file changed? | Pack hash, signature, verification page | A PDF emailed around with no seal |
What the delay to December 2027 changes
Regulation (EU) 2026/1744, the AI Digital Omnibus, entered into force on 27 July 2026. It moved the Chapter III obligations for Annex III high-risk systems to 2 December 2027, and for AI embedded in Annex I products to 2 August 2028. The stated reason was the delay in harmonised standards from CEN and CENELEC.
It did not touch Article 5 prohibitions (in force since 2 February 2025), Article 50 transparency duties (from 2 August 2026) or the general-purpose AI obligations (from 2 August 2025). It also did not remove the obligation to keep records: Article 12 and Article 19 apply as part of Chapter III, so the clock moved, the content did not.
Time to build the file, not permission to skip it. Buyers and procurement teams have already started asking for AI Act documentation in 2026, and a supplier who cannot produce a pack loses to one who can, regardless of the statutory date.
How to build one without a consultant
Three passes, in this order. Each pass produces something you keep, not something you rewrite.
- 1Classify. Answer the Annex III question honestly and write down why. If the answer is "high-risk", the obligations in Articles 9 to 15 apply and the checklist is fixed. If it is "limited", only Article 50 transparency applies. The reasoning matters more than the label.
- 2Produce the evidence as you work. Log every model call with the actor, model version and data categories; record approvals with a name and a timestamp; keep the checklist status current instead of reconstructing it the week before an audit.
- 3Seal and date it. Generate the pack for a defined period, hash the contents, sign the hash, and publish a verification page. A pack you cannot verify is a document; a pack you can verify is evidence.
Twelve questions classify the system, then the checklist, logs and approvals feed a signed PDF and JSON pack. Free to check, €490 for the pack.
Run the free risk checkWhat the pack does not prove
It does not prove the model is accurate, that the data was lawful, or that the system is safe. It proves what was configured, what happened and who approved it. Accuracy and lawfulness are separate evidence: test results, a DPIA, a data processing agreement, a risk register with a named owner. The pack tells the auditor where to look; it does not replace the underlying documents.
It is also not a conformity assessment and not a legal opinion. A notified body or a lawyer still has to reach the conclusion. What the pack removes is the most expensive part of that conversation: reconstructing the facts.
Frequently asked
Is an evidence pack legally required by the AI Act?
Not as a named artefact. The AI Act requires record-keeping (Article 12), technical documentation (Article 11 and Annex IV), human oversight (Article 14) and post-market monitoring (Article 72). A pack is a practical way to package those obligations so they can be shown to a third party. Regulators ask for the content, not the format.
How long should we keep it?
Article 19 requires providers to keep automatically generated logs for at least six months, and Article 26(6) sets the same floor for deployers. Sector law and contracts often require longer. A practical default is to keep every issued pack for the life of the system plus the retention period in your quality management system.
Can we just build it in Word?
You can write the narrative in Word. You cannot produce the two things that make it evidence: records that are provably complete for a period, and a seal that survives being emailed. Those need to come from the system that produced the records.
Does the December 2027 delay mean we can wait?
For the statutory deadline, yes for Annex III Chapter III duties. For customers, no: procurement questionnaires already ask for AI Act documentation, and the delay explicitly does not cover Article 50 transparency (August 2026) or the Article 5 prohibitions (February 2025).
Sources
Keep reading
A practical reading of Article 12 of the EU AI Act: automatic event recording, the minimum fields, retention under Articles 19 and 26(6), and the traps that make logs unusable as evidence.
Which AI systems Annex III makes high-risk, how the Article 6(3) filter works, the full obligation checklist under Articles 9–15, 17, 43 and 72, and the timeline after the AI Digital Omnibus.