How long must AI decision logs be retained, and who decides?

The floor is six months, and it is only a floor. Article 19(1) of Regulation (EU) 2024/1689 requires providers of high-risk AI systems to keep the automatically generated logs referred to in Article 12(1), to the extent those logs are under their control, "for a period appropriate to the intended purpose of the high-risk AI system, of at least six months, unless provided otherwise in the applicable Union or national law, in particular in Union law on the protection of personal data." Article 26(6) puts materially the same obligation on deployers for the logs under their control. The Regulation itself names no maximum.
Which means nobody hands you the number. You derive it. The legislature fixed the lower bound, your own statement of the system's intended purpose sets what "appropriate" has to mean, sectoral law can push the period up and fold the logs into documentation you already keep, and data protection law pushes back on how long you may hold anything that identifies a person. The retention period is a documented decision with four inputs, and a supervisory authority is more likely to ask how you reached it than to recite a number at you.
A note on what this is. This article reads the text of Regulation (EU) 2024/1689 and describes how teams reason about it. It is general information. It does not constitute legal advice and has not been reviewed by qualified regulatory counsel, which is the same disclaimer Auditrek is built to embed in every report it generates. Have your own counsel or DPO review any retention decision before you rely on it.
The wording, and what it does not say
Article 12(1) is the capability obligation: "High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system." Article 12(2) explains what those logs are for, namely identifying situations where the system may present a risk or undergo substantial modification, supporting the post-market monitoring in Article 72, and supporting the deployer monitoring in Article 26(5).
Three things the text does not say are worth noticing, because teams routinely assume all three. It does not say six months from the decision rather than six months from the end of the system's use. It does not say the logs must be immutable. And it does not say a shorter period is available on grounds of storage cost. What it does say is "appropriate to the intended purpose," and that phrase carries the work.
The Regulation entered into force on 1 August 2024 and, under Article 113, applies generally from 2 August 2026, with the obligations attaching to high-risk systems classified under Article 6(1) applying from 2 August 2027.
Who decides, in four layers
The legislature sets the floor. Six months, per Article 19(1) and Article 26(6). That much is fixed by the text.
The intended purpose sets the shape. "Appropriate to the intended purpose" is a reasoning test. How long after an inference could a harm plausibly surface and be traced back to it? If the answer is a year, six months of logs will not let you investigate your own incident, and Article 73 sets short deadlines once a provider becomes aware of a serious one.
Sectoral law can override and absorb. Article 26(6) is explicit that deployers which are financial institutions subject to internal governance requirements under Union financial services law maintain these logs as part of the documentation they already keep under that law. For those firms the AI Act period is often the shortest of the applicable periods rather than the operative one.
Data protection law works the other end. The carve-out "unless provided otherwise in applicable Union or national law, in particular in Union law on the protection of personal data" cuts both ways. Storage limitation under the GDPR does not evaporate because the AI Act asked for logs. It pushes towards records that are useful for audit and thin on identifiers.
A credit scoring model, worked through
Creditworthiness evaluation and credit scoring of natural persons sits in Annex III point 5(b), outside the fraud-detection carve-out. Start with the floor of six months, then ask what an affected person can do and for how long.
Article 86 gives a person subject to a decision taken on the basis of the output of most Annex III high-risk systems, where that decision produces legal effects or similarly significantly affects them, the right to obtain clear and meaningful explanations of the role of the system in the decision. Article 85 lets them complain to a market surveillance authority. A declined applicant who complains eleven months later is not doing anything unusual, and if the logs expired at month six, the explanation rests on a decision that can no longer be evidenced.
One defensible approach many teams take is to align the decision log with the retention period for the credit file it belongs to, and to write down that this is the reason. A common pitfall runs the other way: enriching the log with the raw bureau payload so it reads better later. The lighter pattern keeps the input hash, the model id and version, the output, the confidence, the timestamp and whether a human overrode it.
An emergency triage model, worked through
Emergency healthcare patient triage systems sit in Annex III point 5(d). The tension is sharper here, because the payload is health data and the temptation to log the clinically interesting fields is strong.
One pattern that makes the retention question tractable is separating the two records. The clinical record stays in the clinical system, under the retention schedule the organisation and its regulators already set. The decision record proves only that a specific model version produced a specific output for a specific pseudonymous subject reference at a specific time, and whether a clinician overrode it. That second record can be held for the longer period the intended purpose argues for without dragging special-category data along, which is the kind of trade-off the Article 19(1) data protection carve-out appears to anticipate. Whether the split works in your setting is a question for your DPO and counsel.
Writing the decision down
A retention decision that survives examination tends to fit on one page per system and name:
- the ai_system, its Annex III classification, and whether you hold it as provider, deployer or both
- the statutory floor you are working from and the article it comes from
- the period you chose and the few sentences of reasoning that got you there
- any sectoral or national retention schedule that absorbs or extends it
- what a single record contains, and what it deliberately does not contain
- who approved it and on what date, and when it gets revisited
This is a starting structure to work through with counsel or your DPO, not a compliance requirement and not a statement of what your own obligations are. The reasoning is the artefact. The number without it is a guess that happens to be written down.
What the log layer has to survive
Long retention raises a question short retention hides: how do you show the records were not edited during the years you held them? A retention period is only as meaningful as the integrity of what sits inside it.
When Auditrek launches, it is built to take inference events over REST or webhook, capturing the input hash, model id and version, output, confidence, timestamp and human override flag, and to store them as SHA-256 hash-chained append-only records per ai_system, where each record's hash is computed from the previous hash and the payload. Verification is built to recompute the chain server-side from sequence one and return a tamper status with the first broken sequence number, rather than reporting a stored flag that could have been set by the same process that broke the chain.
The ingestion layer is built to reject payloads carrying recognisable raw sensitive fields, including SSN patterns, ICD-10 style diagnosis codes and unredacted credit bureau fields, before anything is written, and to pseudonymise subject identifiers by default. That is a retention control as much as a privacy control: a ledger you intend to hold for years should never have contained those fields, because deleting a record from a hash chain is not a clean operation. No system can be connected, and no API key or webhook secret issued, until a signed DPA confirmation is on record. The report generator is designed to support documentation of Article 12 and Article 13 obligations for a chosen system and date range, in PDF and HTML, with a disclaimer on every page stating that the output does not constitute legal advice and has not been reviewed by qualified regulatory counsel.
None of that decides your retention period. It makes the period you decided on worth something.
Is six months a minimum or a maximum for AI log retention?
It is a minimum. Article 19(1) and Article 26(6) of Regulation (EU) 2024/1689 require a period appropriate to the intended purpose of the high-risk AI system, of at least six months. The Regulation sets no maximum, so the upper bound comes from data protection law and from your own sectoral retention schedule.
Who keeps the logs, the provider or the deployer?
Both, each for the logs under their control. Article 19(1) addresses providers of high-risk systems and Article 26(6) addresses deployers. If you buy a model and run it on your own inference infrastructure, the events generated there are typically under your control as deployer, though the allocation is worth confirming with counsel.
Does a GDPR erasure request override the AI Act retention obligation?
The two are meant to be reconciled rather than ranked, and Article 19(1) expressly defers to Union law on the protection of personal data. One practical route is to keep decision records that carry pseudonymous subject references and input hashes instead of raw personal data, so the audit record can persist without holding identifying content. How that applies to a specific request is a question for your DPO.
Should the retention clock start at the inference or at system retirement?
The Regulation does not say, which is why the decision has to be reasoned rather than looked up. Article 12(1) frames logging over the lifetime of the system, so some teams set retention per record from the decision date and separately commit to holding the full chain for a defined period after the system is withdrawn. Whichever you choose, write down the reasoning and have counsel review it.