How to Prepare for an EU AI Act Audit: The Evidence Pack
EU AI Act audit preparation is an evidence exercise, not a slide exercise. Build the pack, run a mock audit, and close the findings that matter first.
EU AI Act audit preparation is mostly an exercise in evidence archaeology: everything an authority can demand must exist, be dated, be owned, and be findable by someone who was not in the room when it was created. Firms that treat an audit as a presentation fail; firms that treat it as a filing exercise pass calmly. This guide covers what market surveillance authorities can actually ask for, the contents of a defensible ai act evidence pack for auditors, the preparing technical file for ai act review discipline that keeps versions honest, and a mock-audit method that finds your gaps before someone else does.
What an authority can demand
Under the market surveillance framework (Articles 74–79), a national authority may require documentation, run tests, and — in defined circumstances — restrict or prohibit a system. In practice requests cluster into four shapes:
- The population question: which AI systems do you provide or deploy in this member state, and how have you classified them?
- The proof question: show the conformity evidence — technical documentation, risk-management records, data-governance methodology, validation results, declaration of conformity, CE marking trail.
- The operation question: show human-oversight assignment, logging configuration and retention, incident records, post-market monitoring data, and your Article 73 reporting history.
- The accountability question: who is your named responsible person, when was training recorded, and where are the decision logs for contested classifications?
Two timing facts shape preparation: technical documentation must exist before market placement and be retained for ten years alongside the declaration; deployer logs need at least six months’ retention (Article 26(6)). If either is missing today, no amount of narrative fixes it — start the clock now and document the remediation honestly.
The evidence pack: contents and ownership
Structure the pack the way an auditor reads — by question, not by department:
| Pack section | Core artefacts | Owner |
|---|---|---|
| 1. Population & classification | Register, classification rationales with article references, decision log, contested-item memos | Compliance |
| 2. Provider lifecycle | Annex IV technical file (current version), risk-management record, data-governance methodology, validation report, QMS document, declaration + CE marking record | Product/Engineering |
| 3. Deployer operations | Instructions-for-use received, oversight assignment per tool, log-retention settings evidence, FRIA where applicable, staff literacy records | Business/HR |
| 4. Incidents & monitoring | Incident register, Article 73 reports filed, post-market monitoring summaries, corrective actions with dates | Quality/Risk |
| 5. Procurement & contracts | AI clauses in vendor templates, vendor declarations held, supplier review records | Legal/Procurement |
| 6. Governance | Minutes showing the monthly rhythm, training attendance, role appointments, escalation decisions | Programme owner |
Every artefact gets three metadata fields: version, date, named owner. An auditor’s first pass is whether the evidence could have existed when claimed — undated documents read as documents written for the audit, which poisons the rest of the file.
Preparing the technical file for review
For provider systems, the technical file is the centrepiece and the most common source of findings:
- Version lock. Keep the file that corresponds to each version placed on the market, not just the latest. A system updated three times with one file is three audit failures waiting.
- Traceability. Requirements → design → test → result. If your validation report does not clearly map to the risk-control measures in your risk file, an auditor will find the seam immediately.
- Data claims. State datasets, sources, representativeness work and bias examination as records — Article 10 questions are specific and statistical, not philosophical.
- Oversight design. Human-oversight measures (Article 14) with the interface features that enable them and evidence overseers were told about them.
- Gaps annotated. Where evidence is thin, an annotated note explaining why and the remediation date is stronger than a silent hole — audits reward known-and-managed over unknown.
Build a one-page index into the file: numbered sections matching Annex IV, so an authority request becomes a PDF delivery instead of a three-week excavation.
Run a mock audit before the real one
The mock is where preparation becomes real. Method: appoint an internal reviewer who did not build the file, give them the four authority question shapes, and set a clock — a competent reviewer should be able to assemble the population answer in an hour and the conformity answer in a day. Observe where they stall; every hesitation is a finding. Typical mock-audit results on first run: the register and the file disagree on version numbers; the oversight assignment exists but names a role that left the company; literacy records cover half the teams using generative tools; incident procedure exists but was never tabletop-tested. Fix findings by artefact, re-run the mock on the failed sections only, and keep both the findings and the fixes in the pack — evidence of self-correction is itself persuasive.
When the request actually lands
The clock starts the day the letter arrives, and the response shape matters as much as the substance. First 48 hours: acknowledge receipt, name an internal coordinator, and diarise every date the authority gave — missed deadlines read as evasion. Second: answer in the authority’s own structure. If they asked four questions, reply with four numbered headings, each opening with the direct answer and following with the supporting artefacts; authorities repeat requests for firms that bury answers in narrative. Third: scope discipline. If a question covers only one system, do not volunteer the entire estate — but never contradict your register either, because inconsistency between your reply and your published register converts a routine check into an investigation. Fourth, and most delicate: if the review surfaces a genuine gap, state the fact, the remediation owner and the date. Corrective action offered before it is compelled is treated very differently from a gap discovered under pressure — the cooperation posture is itself a statutory consideration in how enforcement proceeds.
The findings that actually get cited
Mock audits and published enforcement patterns point to the same shortlist: classification without recorded rationale; technical file version drift against the system in production; logging shorter than six months or untested for retrieval; oversight described in policy but unassigned in operation; literacy records absent for the teams doing the work; incidents recorded nowhere because the definition of “serious” was never operationalised; vendor declarations never requested at procurement. Notice that six of seven are documentation failures rather than engineering failures — which is precisely why audit preparation is a filing discipline. The firms that struggle are rarely running non-compliant systems; they are running compliant systems without the receipts.
How to check your own position
Take the authority’s four question shapes and set a stopwatch. Can your population answer — every in-scope system with classification and rationale — be produced in an hour? Can your conformity answer — the technical file and declaration for your highest-risk provider system — be produced in a day? If either stalls, write down exactly where: the delay point is the finding. Then check dates on the three artefacts most likely to be stale: the register, the oversight assignments, and the last incident tabletop. Give the pack a single index document with numbered sections matching the authority’s questions, and re-run the stopwatch after one fix cycle. A team that can answer all four shapes inside a day is not merely audit-ready — it is the kind of respondent that closes requests in one exchange.
Frequently asked questions
What happens in an EU AI Act audit?
A national market surveillance authority asks which AI systems you provide or deploy in its member state and how they are classified, then requests the supporting evidence: technical documentation, risk-management and data-governance records, validation results, declaration of conformity, oversight assignment, logging and incident history. Follow-up can include testing and, in serious cases, restrictions on the system. Your defence is the evidence pack, organised by their questions.
What documents do I need for an AI Act audit?
At minimum: the AI register with classification rationales and decision log; for provider systems, the Annex IV technical file at the correct version, risk-management record, data-governance methodology, validation report, QMS document and signed declaration of conformity; for deployer systems, oversight assignments, log-retention evidence, literacy records and FRIA where applicable; plus the incident register and governance minutes. Every artefact needs a version, a date and a named owner.
How long must I keep AI Act technical documentation?
Ten years for the technical documentation and the EU declaration of conformity after placing the system on the market — and deployer-side logs must be retained for at least six months. Version-lock the file to the system version that was on the market at the time: authorities may request the historical file, not the current one.
What is a mock audit and do we need one?
A rehearsal where an internal reviewer who did not build the file applies the authority’s question shapes under a clock. It is the cheapest way to find version mismatches, stale assignments and missing records before they become findings. Re-run only the failed sections after fixes — two cycles of this puts most mid-market firms in a position where the real audit is a document delivery.
Can a market surveillance authority shut down our AI system?
Yes. The Regulation gives authorities powers to require corrective action, restrict or prohibit non-compliant systems on the market, and order withdrawal — scaled to the risk and the operator’s cooperation. That tail risk is why audit preparation belongs to the risk register, not the marketing calendar: the cost of the pack is trivial against the cost of a suspension.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.