Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

AI Act Log Retention: How Long You Must Keep Records Under Articles 12 and 26

Six months is the AI Act's floor, not its answer — sector law pushes retention up, GDPR storage limitation pulls it down. What Articles 12 and 26 actually require, what to log, and a retention schedule that survives an inspection.

Renomira Team · 10 min read
A narrow archive aisle lined with shelves of storage boxes and binders
Photo: Samuel Zeller samuelzeller / CC0 via Wikimedia Commons

“Keep the logs for six months” is the version of the AI Act’s record-keeping duties that circulates in slide decks, and it is both right and dangerously incomplete. Article 12 tells providers of high-risk systems to ensure automatic event logging over the system’s lifetime; Article 26 tells deployers to keep records of use under their control; where those rules set a minimum retention — expressed in months, not years — sector law, litigation holds and GDPR storage limitation all set their own boundaries around it. The artefact you will actually be judged on is not a duration. It is a schedule: which record, kept how long, on what legal basis, deleted by whom.

What the Act actually requires

Two logging obligations, one for each side of the value chain:

  • Providers (Article 12): high-risk systems must automatically record events relevant to identification, understanding and detection of risks, and of output discrepancies — across the system’s lifetime. The logs must be retained for at least six months unless a more specific Union or national law sets another period, in a configuration that lets you supply them to market-surveillance authorities.
  • Deployers (Article 26): you must keep the logs generated under your control by the high-risk system — typically your inputs, overrides and decisions — for at least six months where applicable, and longer where sector law requires. Logs that contain personal data remain subject to GDPR on top: the Act sets a floor, GDPR storage limitation sets the ceiling.

The subtle point is “under your control”: a SaaS deployer’s records live in the vendor’s infrastructure, yet the duty is yours. If your contract gives you no export or retention path, you hold the obligation and not the means — which is why log access clauses belong in every AI vendor agreement you sign.

The three clocks that interact

Clock Sets Effect
AI Act Art. 12 / 26 Minimum retention for high-risk system logs Six-month floor (unless sector law says otherwise)
GDPR Art. 5(1)(e) Maximum proportionate retention for personal data Forces deletion once purpose is exhausted
Sector law (finance, health, employment) Longer evidentiary periods Pushes years where records prove decisions

Reconciling them is a documentation exercise, not a legal paradox: state the record class, take the longest applicable minimum, then check it is still proportionate under GDPR (with a justification if it is), and hard-code the deletion. A retention matrix that shows this reasoning for each record class is exactly what an inspector or DPO asks to see.

What to log, concretely

Logging regimes fail in two directions — everything and nothing. The minimum useful set for a high-risk system:

  • Event and decision records: timestamped inputs (or their provenance reference), the output delivered, the model/system version that produced it.
  • Human-oversight actions: overrides, rejections, escalations — who, when, why. This is the evidence that your Article 14 human oversight was real rather than decorative.
  • Input-data checks: validation results and data-quality flags raised before processing — your Article 26 input-data duty leaves fingerprints here.
  • Access and changes: who accessed the system and configuration changes — the security log around the decision log.
  • Incident-relevant events: anything that later becomes a serious incident under Article 73 — the 15-day reporting clock is only meetable if the raw material was captured at the time.

Design rule: logs exist for three consumers — incident responders (days), auditors (months), authorities (years). Classify each record class by its longest consumer, and set retention from that, not from storage cost.

Where the retention matrix comes from

A working matrix for a mid-market deployment fits on one page:

Record class Minimum basis Practical retention Deletion owner
High-risk system event logs AI Act Art. 12 / 26 floor 6 months (more where sector law applies) Platform owner
Oversight overrides & rationales Audit defence + sector rules 12–24 months typical starting point Compliance
Input-data quality flags GDPR accountability Until purpose exhausts + challenge window DPO
Incident files Art. 73 support Per incident hold; ≥ statutory investigation periods Security
Subject-access / complaint evidence Limitation periods Case duration + national limitation period DPO

Fill the “practical retention” column with your sector’s answer, attach the legal basis, and review annually. The matrix doubles as the register annex and the DPIA supporting evidence — one table serving three artefacts.

Delivery to authorities: readiness beats duration

Retention without retrieval is theatre. Article 12 expects logs to be provided to authorities in a common, interoperable format on request; Article 21 cooperation duties mean the request will carry a deadline measured in days. Three operational checks separate prepared teams from the rest: can you export a date range in one operation, can you demonstrate the export is complete and untampered (hash or write-once storage), and does someone own the process when the request arrives at 4pm on a Friday? Rehearse one export per quarter; it takes an hour and exposes the gaps while they are cheap.

Common failure modes

  • Vendor-held logs with no export path — the deployer duty survives, the means do not; fix in contract before go-live.
  • “We keep everything forever” — an unbounded retention policy is a GDPR violation, not a bonus; the Act’s floor has no corresponding ceiling exemption.
  • Logs that record outcomes but not overrides — the moment an auditor asks how humans exercised oversight, outcome-only logs answer nothing.
  • Silent log-format changes on model updates — a new schema can break a two-year evidence chain; version the logging schema with the model.

How to check your own position in three minutes

The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves — mapped to the dates that apply. Confirm your email and the full report unlocks.

Frequently asked questions

How long must we keep AI system logs under the EU AI Act?

The floor is six months: providers under Article 12 and deployers under Article 26 retain relevant logs for at least six months unless Union or national sector law specifies otherwise — which it frequently does in finance, health and employment contexts. Six months is the minimum you may keep them, never a reason to delete them sooner than other obligations allow.

Does the AI Act retention duty conflict with GDPR?

No — they are floor and ceiling. GDPR storage limitation requires personal data to be kept only as long as proportionate for its purpose; the AI Act requires certain logs to be kept at least six months. Where they meet, document the justification for the longer period, bound it with a deletion date, and keep evidence of the deletion. The conflict only appears when neither rule is written down.

Our logs live in the vendor’s system — whose obligation is this?

Yours as deployer for the logs generated under your control, theirs as provider for Article 12 records of their system. In practice you need contractual export rights, retention configuration and deletion support; without them you carry a duty you cannot perform. Check this before go-live, because it is far harder to obtain after an incident.

What does an authority actually ask for?

A date range and completeness: the events in the period, in a readable, common format, with assurance nothing was altered. That is why retrieval, integrity (write-once storage or hashing) and a named owner matter more than the raw retention duration — a six-month archive you cannot export quickly fails the practical test.

Is six months enough for incident investigations?

The serious-incident clock under Article 73 runs to 15 days for notification depending on severity, but investigations, litigation and sector enquiries stretch far longer — and once an incident is known, related records should move to litigation-hold status immediately so normal deletion cycles cannot touch them. Treat the six-month floor as the routine case, and your incident policy as the override.

Do these rules apply if we only use off-the-shelf AI tools?

To the extent you operate high-risk systems as deployer, yes for the logs under your control — which is precisely where shadow AI creates invisible gaps: tools nobody registered generate records nobody retains. Reconciling the AI register against actual usage, then securing log access for everything in scope, is the precondition for every retention answer above.

Next step

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
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

ai act log retention periodrecord keeping duration high risk aiai act article 12 logging requirementshow long to keep ai system logs euai system records retention policyrecord retention deployer ai act audit trailaudit trail requirements ai acteu ai act documentation retention schedule

Keep reading