EU AI Act high-risk requirements checklist for mid-market regulated firms
If you are looking for the EU AI Act high-risk requirements checklist, this is the article-by-article version: Articles 9 to 15 set what the system itself must satisfy, Articles 16 to 27 allocate who owns each duty, and Articles 43 to 49 cover conformity assessment, the declaration, CE marking and registration.
Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.
What the high-risk checklist covers, and where it comes from
The obligations people mean when they say "high-risk requirements" sit in Chapter III of Regulation (EU) 2024/1689. Section 2, Articles 8 to 15, describes what a high-risk AI system has to satisfy: a risk management system, data governance, technical documentation, record-keeping, transparency towards deployers, human oversight, and accuracy, robustness and cybersecurity. Section 3, Articles 16 to 27, splits those duties between providers, authorised representatives, importers, distributors and deployers, so two firms running the same model can carry very different paperwork. Sections 4 and 5 cover notified bodies and the conformity assessment route, ending in the EU declaration of conformity, CE marking and registration in the EU database.
That is the whole surface. For a mid-market firm the practical scope is usually narrower than the article count suggests, because most of you are deployers of two to six systems rather than providers of a dozen, and because a large part of the work is documentary rather than architectural.
Three things make this checklist different from the ones that circulate as slide decks:
- It separates the requirement from the evidence. "Maintain a risk management system" is not an artefact. A dated risk register with residual risk judgements and the test results behind them is.
- It names the owner. Readiness work commonly stalls when a requirement is assigned to "the AI working group" rather than to the model owner, the DPO or the head of the first line, because a duty with no named owner tends to produce no artefact.
- It assumes someone will ask. The question compliance officers phrase to themselves as "can we prove it to the regulator?" is an evidence-retrieval question, not a policy question.
One note before the table. This page describes what the Regulation's text sets out and what auditors commonly ask for. It is not legal advice, and whether your current practices satisfy any given article is a judgement for qualified regulatory counsel who can see your systems.
The high-risk requirements checklist, article by article
| Article | What the text sets out | Evidence a reviewer can ask for |
|---|---|---|
| Article 6 and Annex III | Classification as high risk by use case, with a narrow exemption route in Article 6(3) that has to be documented | Your written classification rationale per system, and the documented assessment if you claim the exemption |
| Article 9 | A risk management system running across the lifecycle, iterative rather than a one-off exercise | Dated versions of the risk register, residual risk judgements, and the testing that supports them |
| Article 10 | Data and data governance practices for training, validation and testing data sets | Data provenance notes, bias examination records, statements of known gaps and shortcomings |
| Article 11 and Annex IV | Technical documentation drawn up before the system is placed on the market and kept current | One document set per system, structured against the Annex IV headings, with a change history |
| Article 12 | Automatic recording of events over the system's lifetime, with logging capabilities appropriate to the intended purpose | Machine-generated event records carrying timestamps and the model version in use, plus an account of their integrity |
| Article 13 | Transparency towards deployers through instructions for use that let them interpret and use output appropriately | The instructions-for-use document you actually shipped, versioned against releases |
| Article 14 | Human oversight measures designed into the system so that natural persons can oversee its operation | Oversight design notes, named oversight roles, and records of when a human intervened or overrode an output |
| Article 15 | Accuracy, robustness and cybersecurity appropriate to the intended purpose, with declared performance | Declared accuracy metrics, the test evidence behind them, and your resilience and security testing |
| Article 17 | A quality management system covering design, development, testing and post-market processes | QMS documentation, internal audit records, and evidence the procedures were followed rather than written |
| Article 19 | Retention of automatically generated logs for a period appropriate to the intended purpose, at least six months unless other Union or national law requires longer | A demonstrated retention window and the ability to retrieve a specific period without a data engineering project |
| Articles 43, 47, 48 and 49 | Conformity assessment on the applicable route, the EU declaration of conformity, CE marking, and registration in the EU database | The route decision and its rationale, the signed declaration, and the registration entry |
| Articles 26 and 27 | Deployer duties, including use in line with the instructions, assignment of human oversight, log keeping, and a fundamental rights impact assessment for the deployers named in Article 27 | Deployer-side log retention, the oversight rota, and the impact assessment where it applies |
| Articles 72 and 73 | A post-market monitoring plan and reporting of serious incidents to the relevant authorities | The monitoring plan, and an incident timeline reconstructed from records rather than from memory |
Step zero: confirm you are in scope, and in which role
Two questions decide how much of the table applies to you, and both get answered badly under time pressure.
Is the system high risk? Article 6 routes through two doors. One is Annex I, where the AI is a safety component of a product already covered by sectoral Union legislation. The other is Annex III, which lists use cases directly, and this is the door most financial services, insurance and healthcare firms come through. Annex III point 4 covers employment, workforce management and access to self-employment. Point 5 covers access to essential private and public services, including creditworthiness evaluation and credit scoring, and risk assessment and pricing in life and health insurance. Article 6(3) offers a narrow exemption where the system performs a narrow procedural task or does not materially influence the outcome of decision-making, but claiming it is itself a documented act, not a silent internal conclusion.
Are you the provider, the deployer, or both? If you buy a vendor model and run it as instructed, you are a deployer, and Articles 26 and 27 are your part of the table. If you put your own name on it, substantially modify it, or change its intended purpose, Article 25 can move provider obligations onto you, and the whole of Articles 9 to 15 follows. Firms that fine-tune an external model on their own book of business, or wire an off-the-shelf model into a decision flow it was not documented for, are the ones most likely to discover late that they are sitting in the provider column.
Write the answer down per system, with a date and an author. A classification rationale you can produce in thirty seconds is worth more in an examination than a better-reasoned one nobody recorded.
Article 12 in practice: what automatic logging has to be able to show
Article 12 is the record-keeping and logging obligation, and it is the requirement that reads as easy and behaves as hard. The text calls for high-risk systems to allow for the automatic recording of events over the system's lifetime, with logging capabilities appropriate to the intended purpose. Article 19 adds retention: providers keep those automatically generated logs for a period appropriate to the intended purpose and at least six months, unless other Union or national law says otherwise. Article 26(6) puts a parallel log-keeping duty on deployers.
Notice what the obligation is not. It is not a request for your observability stack. Application logs are built for debugging: they are sampled, rotated, mutable by anyone with write access, and they rarely carry the model version that produced a specific output. An examiner looking at a single contested decision from fourteen months ago wants a narrower thing, and it is usually some combination of these fields.
- A reference to the input, which for sensitive pipelines is better as a hash than as the input itself
- The model identifier and version that produced the output, not the model family
- The output, and the confidence or score attached to it where one exists
- The timestamp, in a form whose integrity you can speak to
- Whether a human overrode the output, which is where Article 12 meets the human oversight measures in Article 14
Then the question an auditor asks second, and that checklists almost never anticipate: how do you know this record has not been edited since it was written? A row in a mutable table with an updated_at column is evidence of what the database says today. A record whose integrity is verifiable independently of the person producing it is evidence of what happened. Tamper-evident approaches, hash chains and write-once storage among them, exist precisely to close that gap, and they are much cheaper to adopt before you have a two-year backlog of decisions than after.
Article 13 in practice: the document your deployers actually read
Article 13 is the transparency and provision of information obligation. High-risk systems are to be designed so their operation is sufficiently transparent for deployers to interpret output and use it appropriately, and they come with instructions for use carrying the information deployers need, including the provider's identity, the system's characteristics, capabilities and limitations of performance, known or foreseeable circumstances that may create risks, human oversight measures, and expected lifetime and maintenance arrangements.
For a provider this is a deliverable with a version number, not a paragraph in a contract. For a deployer it is the document you are measured against, because Article 26 asks you to use the system in accordance with those instructions. That has a quiet consequence worth naming: if your operations team has drifted from the instructions, the instructions themselves become the evidence against you. Three practical habits help.
- Treat instructions for use as release-coupled. When a model version ships, the instructions get a version too, and the pairing is recorded.
- Keep the limitations section honest. An understated limitation is the sentence a supervisory authority will read back to you.
- Store the instructions where the decision records can point at them, so a record from March resolves to the instructions in force in March rather than the current ones.
Articles 12 and 13 are also the pair most audit requests actually land on, which is why they are the two the Auditrek report generator is scoped to at launch.
A readiness sequence you can run without a dedicated programme team
- Build the inventory before the gap assessment
You cannot assess what you have not listed. One row per AI use case, with the business owner, the model and version, whether the output touches an individual, and whether a human reviews it. Shadow use cases in spreadsheets and vendor features count. Most mid-market inventories end up shorter than expected and less centralised than expected.
- Classify each system and record the reasoning
Walk each row against Annex I and Annex III, then against the Article 6(3) exemption if you intend to rely on it. Output is a dated, authored rationale per system, not a red or green cell. Reclassify when the intended purpose changes, because purpose is what the classification hangs on.
- Fix your role per system under Article 25
Provider, deployer, or both, decided per system rather than per firm. Check fine-tuning, rebranding and purpose changes specifically. This single decision determines whether you owe Articles 9 to 15 and a conformity assessment, or Articles 26 and 27 and log retention.
- Define what a decision record is for each in-scope system
Agree the field list with the model owner and the DPO in the same room: input hash rather than raw input, model id and version, output, confidence, timestamp, human override flag, pseudonymised subject reference. Disagreements here are usually about sensitive data, and they are easier to resolve on a whiteboard than in an ingestion bug.
- Decide where the record is written and how its integrity is established
The options are your data platform with write-once storage and access controls, a purpose-built append-only ledger, or both. The test to apply is not whether the record exists but whether a third party can verify it was not altered after the fact, and whether verification depends on trusting the team that produced it.
- Rehearse the request end to end
Pick a system, pick a date range, and produce the evidence pack as though an authority had asked. Time it. Note every step that required an engineer, a query written by hand, or a screenshot. Those steps are your real gap list, and they are more useful than any maturity score.
- Connect post-market monitoring to the same records
Articles 72 and 73 assume you can reconstruct what a system did in a defined window, including around an incident. If your monitoring plan reads from the same records that back your Article 12 evidence, incident reporting becomes a retrieval task. If it reads from a separate dashboard, you will be reconciling two accounts of the same week under deadline.
Where mid-market teams get stuck, and the trade-offs worth naming
Enterprise readiness programmes stall on committee time. Mid-market ones stall on four specific trade-offs.
Re-platforming versus instrumenting
The governance platforms that dominate this category want to sit around your model lifecycle, which often means moving model registration, evaluation and deployment into their world. That is a credible answer if you are staffing a model risk function anyway. If you have four systems and one platform engineer, the cost of a migration lands before any evidence does. Instrumenting the existing pipeline to emit decision records is a smaller change with a narrower promise, and it leaves the production inference path alone.
Data minimisation versus evidentiary completeness
Credit, hiring and clinical pipelines carry special category personal data under GDPR and, in US contexts, data touched by HIPAA and FCRA. The instinct to log everything for defensibility collides directly with the instinct to log nothing for privacy. When you design your own record format, the middle path most privacy teams will accept is to carry a hash of the input rather than the input, a pseudonymised subject reference rather than a name or account number, and no raw sensitive fields in the evidence layer at all, so the record shows what the system did without duplicating the data that made it sensitive. Whatever rule you and your DPO settle on, decide it before the first event is written, because retrofitting a field policy onto a populated store is the part that consumes the quarter.
Presence versus integrity
A log that exists is a project milestone. A log whose integrity can be tested is evidence. The difference only becomes visible during an examination, which is the worst moment to discover that your records are mutable and your retention was set by a default.
Retention versus deletion rights
Article 19's minimum sits alongside data subject rights and your own deletion schedules. Holding a hash and a pseudonymised reference rather than the underlying record is generally the easier posture to defend on both sides, and it is worth settling with your DPO before ingestion starts rather than during a subject access request.
Dates and penalties, stated carefully
Regulation (EU) 2024/1689 entered into force on 1 August 2024 and applies in stages under Article 113, with the prohibitions and AI literacy provisions first, general-purpose AI and governance provisions next, the bulk of the Regulation including the Annex III high-risk obligations following, and Article 6(1) systems embedded in products under sectoral legislation later still. Amendment proposals affecting the high-risk timing have been in circulation, so the only responsible instruction is the boring one: read the dates off the current consolidated text in the Official Journal and off the Commission's own framework page before you commit a board date to them. Treat whatever you find as a planning constraint rather than a countdown, because the work in the table above takes quarters, not weeks, and none of it is wasted if a date moves.
On penalties, Article 99 sets tiers rather than a single headline number: the highest cap applies to infringements of the prohibited practices in Article 5, a middle cap applies to non-compliance with other obligations including those on providers and deployers, and a lower cap applies to supplying incorrect, incomplete or misleading information to authorities. In each case the figure is the greater of a fixed amount and a percentage of total worldwide annual turnover, with the lower of the two applying for SMEs. Several figures still circulating in decks and vendor material come from draft versions of the Regulation and do not match the adopted text, which is one more reason to cite the article rather than the slide.
Vertical detail, and what to read next
This page is deliberately horizontal, because the checklist is the same regardless of sector while the arguments you will have internally are not. Two sibling pages in this pillar go a level deeper into the systems that come up most.
The credit scoring page works through what an append-only record behind every automated decision looks like when the same model output has to satisfy an Article 12 evidence request, a model validation review and an adverse action question. The underwriting and pricing page covers the same ground for pricing models, and specifically the problem of one record set that an actuary and an external auditor can both read without translation.
If your next step is internal rather than external, run the rehearsal in the steps above this week and bring the timing to your next risk committee. A measured retrieval time is a more persuasive artefact than a readiness score, and it tends to end the argument about whether the current logs are enough.
What ready looks like on the day someone asks
A useful definition of readiness is not a completed spreadsheet. It is the set of things you can produce in an afternoon, without an engineer writing a query and without anyone reconstructing a timeline from email.
- A current inventory of AI use cases with a dated classification rationale and a named role for each
- Technical documentation per high-risk system structured against Annex IV, with a change history
- Instructions for use in force on any given past date, resolvable from a record made on that date
- Decision records for a chosen system and date range, carrying model version, timestamp and human override, with their integrity verifiable by someone who does not work for you
- A retention statement you can evidence rather than assert
- A post-market monitoring plan and an incident timeline drawn from those same records
Nothing on that list requires a new model platform. Most of it requires deciding, once, what a decision record is in your firm, writing it somewhere that cannot be quietly edited, and rehearsing the retrieval before anyone asks you to perform it.
Questions
What logs does Article 12 actually require?
Article 12 requires high-risk AI systems to allow for the automatic recording of events over the system's lifetime, with logging capabilities appropriate to the intended purpose. It does not hand you a field list, so in practice teams settle on a reference to the input, the model identifier and version, the output and its confidence, the timestamp, and whether a human overrode the result. Article 14's human oversight measures and Article 26's deployer duties are what turn that last field from nice to necessary.
How long do we have to keep AI logs?
Article 19 requires providers to keep the logs automatically generated by their high-risk AI systems for a period appropriate to the intended purpose and at least six months, unless other applicable Union or national law provides otherwise. Article 26(6) places a parallel log-keeping duty on deployers. Sectoral retention rules in financial services or healthcare frequently run far longer, so the six months is a floor rather than a target, and your own record retention schedule usually decides the real number.
Can a checklist tell us whether we meet the requirements?
No. A checklist maps the text and the artefacts examiners commonly ask for, which makes it a good planning instrument and a poor conclusion. Whether your specific practices satisfy a specific article is a judgement that needs qualified regulatory counsel with visibility into your systems. This page is descriptive, and so is every report Auditrek generates: each one carries a footer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel.
We fine-tune a vendor model. Are we a provider or a deployer?
It depends on what you changed, and Article 25 is the article to read. Putting your own name on a high-risk system, substantially modifying it, or changing its intended purpose can move provider obligations onto you, which brings Articles 9 to 15 and a conformity assessment route with them. Deciding this per system and recording the reasoning is step three of the sequence above, because everything downstream, including how much documentation you owe, hangs on the answer.
Is our existing observability stack enough for Article 12?
It is usually the wrong shape rather than the wrong idea. Application logs are sampled, rotated and mutable by whoever holds write access, and they often omit the model version that produced a given output. The gap that matters is integrity: an examiner asking about one contested decision from last year wants a record whose non-alteration can be tested independently, not a query result from a system your own team can edit.
Do we need ISO/IEC 42001 or the NIST AI Risk Management Framework as well?
Both are separate instruments, not the Regulation. ISO/IEC 42001 is a management system standard and the NIST AI Risk Management Framework is a voluntary framework, and neither one is Regulation (EU) 2024/1689 or a substitute for reading it. The obligations in the table above come from the Regulation's own articles, so treat them as the source of record, and take the question of whether either instrument earns a place in your programme to your second line and your counsel, reading each document directly rather than through a vendor summary.
Does one report cover every AI system we run?
Not in Auditrek. A report covers one chosen AI system and one date range, in PDF or HTML, mapped to Article 12 and 13 fields, with a single regulation and format set at launch. That is deliberate: audit requests arrive scoped to a system and a period, and a per-system report is the artefact that matches the request. The dashboard is where you see the whole estate, showing decisions logged, systems connected, tamper-verification status and when each report was last generated.