Home / Guides

Article 12 record-keeping, field by field: what an AI log has to contain

Article 12 of the EU AI Act requires that a high-risk AI system technically allow for the automatic recording of events, called logs, over the lifetime of the system. This guide takes that sentence apart field by field, shows where the regulation names specific log contents and where it leaves you to derive them, and sets out the record schema a compliance officer can hand to an engineering team.

Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.

What Article 12 says, and what it deliberately does not say

Article 12 of Regulation (EU) 2024/1689 is headed Record-keeping, and it is short. The operative sentence is that high-risk AI systems shall technically allow for the automatic recording of events, called logs, over the lifetime of the system. Two things in that sentence do most of the work. Automatic rules out a process where an analyst writes a case note after the fact. Over the lifetime of the system rules out logging that begins when a supervisory authority asks a question.

Article 12(2) then sets a purpose test rather than a format. Logging capabilities have to enable the recording of events relevant to identifying situations that may result in the system presenting a risk within the meaning of Article 79(1) or undergoing a substantial modification, to facilitating the post-market monitoring referred to in Article 72, and to monitoring the operation of high-risk systems as referred to in Article 26(5).

Article 12(3) is the only place in the article where the regulation names specific log contents, and it applies to one category: high-risk systems referred to in point 1(a) of Annex III, which is remote biometric identification. For those, the text sets a minimum of the period of each use with start and end date and time, the reference database against which input data was checked, the input data for which the search produced a match, and the identification of the natural persons involved in verifying the results under Article 14(5).

That structure is what catches teams out. If your system is a creditworthiness model, an employment screening model or a triage support tool, there is no field list waiting for you in the text. Article 12 tells you what your logs have to make possible, and you derive the fields from the intended purpose of your own system. The rest of this guide is that derivation.

The three purposes a log has to serve

Before arguing about columns, it helps to hold the three tests from Article 12(2) in view, because each one implies different evidence.

Identifying risk situations and substantial modification

Article 79(1) concerns AI systems presenting a risk, and the reference points your logging at the question of whether the system's behaviour has changed or started producing outcomes that warrant attention. Practically, that means a log where you can isolate a slice of decisions, characterise them, and show what model artefact produced them. A record that does not carry a model identifier and version cannot answer whether a shift in outcomes followed a deployment, which is the first thing anyone will ask.

Facilitating post-market monitoring under Article 72

Post-market monitoring is a continuing activity across the system's life, and it runs on aggregates drawn from primary records. If your logs are sampled, or retained only in a rolling window shorter than your monitoring period, the monitoring plan has nothing underneath it. This is the purpose that makes retention a schema question rather than an infrastructure afterthought.

Monitoring operation under Article 26(5)

Article 26 sets deployer obligations, including monitoring the operation of the system on the basis of the instructions for use. Monitoring operation means being able to say what the system did on a given day, for a given case, and whether a person intervened. That is the purpose that drags human oversight into the record, and it is the one most often missing from logs built by an ML platform team for debugging rather than for evidence.

What a high-risk AI log record carries, field by field

Record fieldWhat it evidencesWhere it usually comes from
Event timestamp, and start and end of the period of useThe anchor for every other field, and the explicit Article 12(3)(a) minimum for remote biometric identificationThe serving layer or gateway clock, captured at request and at response
AI system identifierWhich entry in your AI inventory produced this decision, so logs and registered systems reconcileYour own system register, carried as a stable key on every event
Model identifier and versionWhich artefact decided, so that a model change is visible as a change rather than as unexplained driftModel registry tag or deployment manifest, read at inference time
Input reference, as a hash rather than the raw inputThat this decision was made on this exact input, without copying applicant or patient data into a second storeA digest computed in the calling service before the request leaves it
OutputWhat the system actually produced, in the form the downstream process consumedThe inference response
Confidence or scoreWhere the output sat relative to thresholds, which is what makes a borderline case reviewable laterThe inference response
Human override flagWhether a person changed or confirmed the outcome, the operational trace of oversight under Article 14The case management, adjudication or review tool, not the model
Reference database checked, and the input that produced a matchArticle 12(3)(b) and (c), a minimum for Annex III point 1(a) remote biometric identification systemsThe matching service
Identification of the natural persons who verified resultsArticle 12(3)(d), read with Article 14(5), for the same Annex III point 1(a) categoryThe review workflow, with identifiers pseudonymised in the ledger
Sequence number and previous record hashThat the record sits in an unbroken order, so a deletion shows as a gap rather than as silenceThe ledger itself, computed on write

Reading the field list closely

Time is two fields, not one

Article 12(3)(a) speaks about the period of each use, with start and end date and time. For a batch scoring job or a session-based assistive system, a single event timestamp collapses a period into a point and loses the thing the text asks for. For a single synchronous inference the two may be milliseconds apart, and recording both still costs nothing and removes an argument.

The input belongs as a digest, not as a copy

The inference payloads in scope here come from credit, hiring and clinical pipelines. They routinely carry personal data, and often special-category data under the GDPR. Copying the raw input into an audit ledger creates a second population of sensitive records with its own lawful basis, residency and retention problems, and it does so in the system you least want to be interesting to a data protection authority. A cryptographic digest of the input, computed at the calling service, lets you prove later that a specific input produced a specific output, because re-hashing the retained case file reproduces the digest. It does not let you reconstruct the input from the log, which is the point.

Model version is the field that makes the log evidential

A log of outputs without artefact identity answers what happened and not what decided. When a supervisory authority or an internal model risk function asks whether a change in outcomes followed a retrain, the model identifier and version is the join key between the decision record and your change history. Record it as it appears in the registry, not as a human label that gets edited.

Human override is where Article 12 meets Article 14

Article 14 covers human oversight, and the obligation lives in system design and in the deployer's arrangements rather than in the log. But the log is where oversight becomes visible. A boolean override flag on the decision record, written by the tool the reviewer actually uses, is the difference between asserting that a human reviews borderline cases and being able to show the proportion that were reviewed in a given quarter.

The chain fields are not in the text, and they are what makes the rest hold up

A sequence number and a hash of the preceding record are not Article 12 requirements. They exist because an append-only claim is only as strong as the mechanism behind it, a point taken up further down this page.

Deriving the schema for one system

  1. Fix the system boundary and its intended purpose

    Write down, in one sentence, what this system is for and where its outputs land. Article 12(2) is a purpose test, so the intended purpose is the input to every later decision about what to record. Systems that share a model but serve different purposes are separate entries.

  2. Check whether Article 12(3) applies literally

    If the system falls under point 1(a) of Annex III, remote biometric identification, the four minimum fields are named in the text and the derivation is mostly done. If it does not, treat that list as a useful template rather than as a ceiling or a floor.

  3. Work the three purposes into candidate fields

    For each of risk identification, post-market monitoring under Article 72 and operational monitoring under Article 26(5), ask what a record would need to carry to make that activity possible a year from now. Keep the fields that survive all three questions and note which purpose justifies the rest.

  4. Decide what never enters the ledger

    Name the fields that must stay out: raw applicant identifiers, diagnosis codes, unredacted bureau data. Replace identity with a pseudonymous subject reference and the input with a digest. Write this as a rule enforced at ingestion, not as guidance in a runbook.

  5. Choose the capture point in the pipeline

    The record should be written where the decision is served, by the gateway or the calling service, so it does not depend on the model or on a migration. Capturing at the point of use also gives you the human override flag from the review tool rather than from the model.

  6. Set retention against the longest applicable clock

    Retention is fixed by the regulation's minimum read together with your sectoral record-keeping duties and your data protection retention analysis. Set the schema and storage so the longest of those, not the shortest, is the one you can satisfy.

  7. Write the verification procedure before you need it

    Decide now how someone outside the team would check that the log has not been altered, and who runs that check on what cadence. A procedure invented during an examination is worth less than one with a history of being run.

How long the logs are kept

Article 12 creates the logging capability. Retention sits elsewhere, and the two are often confused in internal memos.

For providers, Article 19 covers automatically generated logs: providers keep the logs referred to in Article 12(1) that their high-risk systems generate, to the extent those logs are under their control, for a period appropriate to the intended purpose of the system and of at least six months, unless Union or national law provides otherwise, in particular Union law on the protection of personal data. For deployers, Article 26 carries a parallel duty over the logs under their control, with the same at-least-six-months framing, and a specific provision directing deployers that are financial institutions to keep the logs as part of the documentation they already maintain under Union financial services law.

Two readings of that follow. First, six months is a floor conditioned on a period appropriate to the intended purpose, not a target. A creditworthiness model whose decisions are contestable for years does not have a six-month evidence problem. Second, the data protection carve-out cuts both ways: it can require you to keep less, and a minimised record schema is how you avoid a direct conflict between a retention duty and a minimisation duty. A ledger of digests, model versions, outputs and override flags can be retained for years with a defensible rationale. A ledger containing raw clinical or bureau fields cannot.

Tamper evidence, and why the argument is evidential rather than textual

Article 12 does not use the words immutable, tamper-proof or WORM. Anyone telling you the regulation mandates a hash chain is reading something into the text that is not there. The argument for tamper evidence is different and, for a compliance officer, more useful.

A log is evidence. Evidence is weighed by whoever is examining it, and the weight depends on how hard it would be for the party presenting it to have altered it. An application-level table that the team operating the model can update in place is evidence of a weak kind, because the answer to how do we know this record was not edited after the outcome became inconvenient is a statement about access control and intentions rather than about mathematics.

A hash chain moves that answer into the record. Each entry stores the hash of the one before it, and its own hash is computed over that previous hash and the entry's payload. Changing any field in any historical record changes that record's hash, which breaks every hash after it. You are not claiming that records cannot be altered, which no storage system can honestly claim. You are claiming that alteration cannot be silent, and a verification pass recomputing the chain from the first sequence number will locate where it happened.

The practical detail that decides whether this is real: verification has to recompute the chain rather than read a stored status flag, and every field you intend to rely on has to be inside the hashed payload. A timestamp held outside the hashed payload is a timestamp that can be moved without breaking anything, which in a record-keeping context is precisely the field an examiner cares about. When you evaluate any logging approach, including ours, ask what exactly is covered by the digest.

Four readings of Article 12 that cause rework

That the Article 12(3) list is the requirement for every high-risk system

It is the minimum for high-risk systems under point 1(a) of Annex III. Applying it verbatim to a hiring or credit system produces fields that do not exist, such as a reference database, while leaving out the model version that your own monitoring depends on.

That observability tooling already satisfies it

Application performance monitoring and ML observability stacks capture much of the payload, and they are built around retention windows measured in weeks, sampling to control cost, and mutable indices. Each of those design choices is correct for their purpose and wrong for a record-keeping obligation. The question to ask of an existing stack is not whether the fields are present but whether all of them are retained, unsampled, for the full period, and whether anyone can alter them.

That the log has to sit inside the inference path

Nothing in Article 12 requires the record to be written by the model or by the serving container. Writing it from the gateway or the calling service, after the response is returned, keeps latency and failure modes out of production inference while still capturing the decision at the point of use. It also means a change of model platform does not invalidate the audit trail.

That record-keeping is a provider obligation only

Providers and deployers both hold duties over logs under their control, and in most mid-market cases the organisation buying a model and putting it into a decision process is a deployer with real obligations of its own. Establishing which role you occupy for each system, with counsel, is the step that determines which articles you are reading.

Where Article 12 hands over to Article 13, and what to do next

Article 13 covers transparency and the provision of information to deployers, and it is the article behind instructions for use. It governs what a deployer is told about the system: its capabilities and limitations, the conditions of use, the human oversight measures, expected performance. Article 12 governs what the system recorded while running.

The two meet at examination time. A supervisory authority reviewing a decision will usually want the instructions for use that framed how the system was meant to operate, alongside the records of how it actually operated in the period under review. That is why documentation output is worth structuring against both articles at once rather than assembling an Article 12 extract and an Article 13 pack separately. Our page on the audit report generator covers what that combined output contains and how the date range and system selection work.

There is a step before any of this that is easy to postpone. Routing inference events from credit, hiring or clinical pipelines into any third-party ledger is a processing arrangement, and the data processing agreement is a precondition for it rather than a document that follows the pilot. Our page on the DPA gate explains why connection credentials should not exist until that agreement is recorded.

A short sequence for the next quarter

Start with the inventory, because you cannot derive a schema for a system you have not named. For each named system, establish your role, fix the intended purpose, and derive the fields using the steps above. Pick one system, ideally one with a real decision volume and a real review workflow, and log it end to end for a month before touching the rest. The first month will tell you which fields your pipeline cannot actually produce, and that is the finding worth having early.

This guide describes the structure of the regulation and of a record. It is not legal advice, and whether any particular logging arrangement satisfies an obligation is a question for qualified regulatory counsel who knows your systems.

Questions

What does Article 12 require logs to contain?

For most high-risk systems, Article 12 does not name fields. It requires that the system technically allow for the automatic recording of events over its lifetime, and that the logging enable risk identification, post-market monitoring under Article 72 and operational monitoring under Article 26(5). Article 12(3) names a minimum set, the period of each use, the reference database checked, the input data that produced a match, and the persons who verified results, for high-risk systems under point 1(a) of Annex III, remote biometric identification.

How long must AI logs be retained?

The regulation sets a floor of at least six months, expressed as a period appropriate to the intended purpose of the system, in Article 19 for providers and in the equivalent deployer provision in Article 26, unless Union or national law provides otherwise, in particular data protection law. Deployers that are financial institutions are directed to keep the logs as part of the documentation they already maintain under Union financial services law. Read the floor together with your sectoral record-keeping duties and your retention analysis.

Does Article 12 apply to deployers or only to providers?

Article 12 is a requirement on the high-risk system itself, which falls to the provider to build. The duty to keep the resulting logs attaches to both roles over the logs under their control, to providers through Article 19 and to deployers through Article 26. Organisations that buy a model and put it into a decision process are frequently deployers with obligations of their own, so establishing your role per system is the first step.

Does the EU AI Act require tamper-proof or immutable logs?

Article 12 does not use those words and does not mandate any particular storage technology. The case for tamper-evident storage is evidential rather than textual: a record that the operating team can silently edit carries less weight in an examination than one where any alteration breaks a verifiable chain. If you adopt a hash chain, check which fields are inside the hashed payload and whether verification recomputes the chain rather than reading a stored flag.

Can AI logs contain personal data?

Inference events from credit, hiring and clinical pipelines routinely carry personal data and often special-category data, so the ledger is a processing activity governed by the GDPR in its own right. The common design answer is to store a cryptographic digest of the input rather than the input, and a pseudonymous subject reference rather than an identifier, which preserves the ability to prove that a given input produced a given output without duplicating sensitive records.

What is the difference between Article 12 and Article 13?

Article 12 is record-keeping: what the system recorded while it ran. Article 13 is transparency and the provision of information to deployers, which is the basis for instructions for use covering capabilities, limitations, conditions of use and oversight measures. An examination usually draws on both, which is why documentation output is worth structuring against the two together.

When do the high-risk obligations start to apply?

The regulation as adopted phases obligations in over several dates, with the high-risk provisions applying later than the prohibitions and the general-purpose model rules. The timeline has been the subject of amendment proposals since adoption, so check the current consolidated text on EUR-Lex and the Commission's own regulatory framework page rather than a secondary summary, and confirm the applicable date for your system category with counsel.

Does a logging tool need to sit in the inference path?

Nothing in Article 12 requires the record to be written by the model or the serving container. Capturing the event from the gateway or calling service after the response is returned records the decision at the point of use while keeping latency and failure modes out of production inference, and it means changing model platforms does not invalidate the trail.

Sources

Join the waitlist

Leave your work email and we will send the payload schema, the verification endpoint documentation and a sample Article 12 and 13 report before you talk to anyone. Pilot cohort conversations start with your own pipeline, not a demo script.

One email when it launches, and one now to confirm your address. Unsubscribe any time. Privacy.