Skip to content
Home/Learn/AI Act record keeping retention period
retentiongdprrecord-keeping

How long to keep AI Act records (and why GDPR keeps changing the answer)

Updated 2026-11-16·8 min read·TraceSeal research

There are three different retention clocks in the AI Act and most teams apply only one. Logs have a six-month floor, technical documentation has a ten-year clock, and the GDPR can force you to delete before either runs out.

The three clocks

RecordAI Act ruleMinimum
Automatically generated logsArt. 19 (provider), Art. 26(6) (deployer)6 months
Technical documentation and QMS recordsArt. 1810 years after placing on the market
EU declaration of conformityArt. 4710 years
Evidence packs issued to customersNo explicit ruleLife of the system plus your QMS period

Both log articles add the same qualifier: "unless provided otherwise in applicable Union or national law, in particular in Union or national law relating to the protection of personal data". That sentence is where the argument starts.

What pushes retention up

  • Employment claims. Discrimination cases often have limitation periods of one to five years depending on the member state; recruitment logs are the evidence.
  • Financial services. MiFID II record-keeping runs to five years, and member states can extend it.
  • Health and safety. Incident records frequently need ten years or more.
  • Contractual audit rights. Enterprise DPAs increasingly require AI records for the contract term plus three years.
  • Litigation hold. Once a claim or investigation is reasonably anticipated, deletion stops — and deleting after that point is spoliation.

What pushes retention down

The GDPR's storage limitation principle (Article 5(1)(e)) does not allow keeping personal data "just in case". A log containing candidate names, emails and rejection reasons is personal data. Keeping it for six years because "logs are useful" needs a documented legal basis and a proportionality argument.

The practical resolution is tiering. Keep the full record only as long as it serves a legal purpose, then reduce it: delete the payload, keep the pseudonymous event with its hash and category labels. You retain the ability to prove what happened without retaining the person's data. The hash chain keeps working because the hash was computed over the original record and is stored separately.

A retention table you can defend

LayerContentKeepWhy
Raw eventMasked summary, actor, model, categories12 monthsOperational review and incident analysis
Hash chainHashes, sequence, timestamps10 yearsIntegrity proof survives payload deletion
ApprovalsApprover, timestamp, decision, reason6 yearsEmployment and contract claims
Evidence packsPDF + JSON as issued10 yearsCustomer audit rights and regulator requests
Technical fileAnnex IV documentation, test results10 yearsArt. 18
The one rule that saves you

Write the retention period into the quality management system before the first audit, apply it automatically, and log every deletion. An auditor who finds a documented rule that is actually followed is satisfied; an auditor who finds unexplained gaps is not.

Deletion is an event too

When you delete, record what was deleted, when, why, and who authorised it. That record is not personal data — it is proof of compliance with storage limitation. It also protects you when a data subject asks why their record is gone: the answer is a documented policy, not an accident.

Two mechanics matter in practice. Deletion must cascade: backups, replicas, analytics warehouses and vendor copies all count. And deletion must not break the chain: keep the hash and the pointer to the deleted record so integrity verification still works after the payload is gone.

Retention and the evidence pack

A pack is a snapshot, and it should outlive the raw data it summarises. That is the point of issuing one: the pack states the period, the events included and the chain head, so the record remains provable after the underlying rows are minimised or deleted. Keep the pack, drop the payload.

Do the maths for your system

The free risk check tells you whether you are high-risk; if you are, log retention applies. The pack records the period and the chain so the proof survives.

Run the free risk check

Frequently asked

Is six months enough for recruitment logs?

Rarely. Employment discrimination claims can be brought years later, and the logs are the evidence. Most HR deployments should keep recruitment records for the limitation period in their member state, which is often one to five years, with the payload minimised after the operational need ends.

Does the GDPR force us to delete AI logs on request?

The right to erasure is not absolute. It is balanced against legal obligations, including record-keeping duties under the AI Act and sector law. The defensible position is a documented retention rule, a legal basis for the log, and erasure where no obligation applies.

Do backups count for retention?

Yes. A deletion that only removes the live row is not a deletion. Backup rotation schedules should be documented, and restored backups must not resurrect data that was lawfully deleted.

How long do we keep the evidence packs we issued?

No explicit AI Act rule, so align with the technical documentation clock (ten years) or the contract, whichever is longer. Packs are small and contain the integrity proof, so the cost of keeping them is low.

Sources

This article is technical information, not legal advice. It reflects Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744 and the sources listed above, as of 2026-11-16. Confirm obligations with qualified counsel for your specific system.

Keep reading