The EU AI Act compliance checklist: 12 steps to a defensible programme
Twelve ordered steps, from scoping your role to closing the evidence loop — with the article, the owner and the proof an auditor will ask for at each stage.
Most mid-market teams do not fail the EU AI Act because the rules are hard to read. They fail because the work is done in the wrong order: training before anyone knows what is in scope, a register nobody owns, documentation started only when a customer asks. This eu ai act compliance checklist for companies reverses that order — twelve steps, each producing an artefact you can hand an auditor or a market-surveillance authority.
A step by step AI Act compliance plan: twelve steps
The whole plan on one page: the article that imposes each step, the role that owns it and the evidence an auditor will ask to see. Treat it as your ai act compliance programme checklist backbone.
| Step | Obligation | Article | Owner | Evidence |
|---|---|---|---|---|
| 1 | Scope your role: provider, deployer, importer or downstream modifier | Art. 3 (definitions), Art. 2 (scope) | Head of Legal | Signed role memo per system |
| 2 | Screen every system for prohibited practices | Art. 5 | Head of Legal with DPO | Dated screening record per system |
| 3 | Classify each system against the risk rules | Art. 6 and Annex III | Compliance officer | Classification rationale, not just a label |
| 4 | Name a single accountable owner and set governance | Duty supporting Art. 26 oversight | CISO with Head of Legal | Approved RACI and AI policy |
| 5 | Deliver AI literacy to everyone who touches AI | Art. 4 | Head of People | Training records with attendance and content |
| 6 | Build and maintain the AI register | Foundational (feeds all duties) | Compliance officer | Live register with review dates |
| 7 | Assemble technical documentation | Annex IV, Art. 11 | Product or engineering lead | Versioned documentation pack |
| 8 | Choose and run the conformity route | Art. 43 with Annex VI or VII | Quality lead | Internal-control checklist or notified-body certificate |
| 9 | Register the system in the EU database | Art. 49 | Compliance officer | Registration confirmation with timestamp |
| 10 | Meet transparency duties: disclosure, synthetic content, deepfakes | Art. 50 | Product owner | Disclosure copy and marking workflow |
| 11 | Stand up incident reporting within 15 days | Art. 73 | CISO | Incident playbook and desktop-test record |
| 12 | Close the loop: post-market monitoring and audit evidence | Art. 72; logging duties for deployers | Compliance officer | Monitoring plan and quarterly review minutes |
Two things are worth noticing: the owners are roles you already have, and the evidence column is where programmes are won — performed but not evidenced is, in an inspection, not performed.
Steps 1 to 3: scope your role, screen prohibitions, classify
Step 1 — scope your role. Almost every misapplied obligation traces back to a company that never decided whether it is a provider or a deployer. A provider places a system on the market under its own name; a deployer uses one under its own authority. Mid-market firms are usually both — deploying vendor tools today, becoming a provider the moment an in-house team fine-tunes a model and releases it under the company name. Write one memo per system, because that memo decides every article that follows.
Step 2 — screen for prohibited practices. Run the Article 5 screen before any classification: a prohibited system cannot be made compliant, only switched off. The categories that catch mid-market companies: untargeted facial-image scraping, workplace emotion inference, social scoring on behaviour or personal characteristics, predictive policing on profiling. One question per system — if yes, stop and document the stop. These prohibitions have been live since February 2025.
Step 3 — classify what remains. Article 6 plus Annex III is the classifier: biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration and justice. Article 6 also filters out narrow procedural tasks and systems that only improve a prior human activity. Record the reasoning, not the verdict: “High-risk — Annex III, employment, because the system ranks candidates and recommends rejections” survives scrutiny; “High-risk” does not.
An AI Act readiness checklist mid-market teams can run: steps 4 to 6
Step 4 — name the owner. Governance fails when it is a committee. For each system, appoint one accountable individual — not a team, not a mailbox — and publish a short RACI: legal owns classification, the CISO owns vendor security, HR owns anything touching staff, the DPO owns the data overlap. The AI policy above it should be two or three pages, board-approved, and say how classification disputes are resolved.
Step 5 — deliver AI literacy. Article 4 has been in force since February 2025 and is the most commonly missed live obligation: staff dealing with AI systems must have sufficient literacy of its risks, limitations and proper use. That includes procurement staff who buy AI tools, HR staff reading AI-generated candidate summaries and finance staff trusting automated scores. Evidence means attendance records, the material used and a short assessment for high-risk users, refreshed annually. Training the 40 people who operate high-risk systems on their actual tool is compliance; generic training for 700 is theatre.
Step 6 — build the register. One row per system: vendor, role, purpose, data categories, risk tier, classification rationale, applicable obligations, owner, next review date. The register tells you which of the other eleven steps apply to which tool — and it is what an auditor asks for first. Include shadow AI — discovery lands between twenty and sixty tools for a company this size, never what leadership expects. Review quarterly, with a mandatory entry at procurement.
Steps 7 to 9: documentation, conformity and registration
Step 7 — technical documentation. Annex IV requires an intended-purpose description, design choices, development and validation methods, the risk assessment, the data list and human-oversight measures. Deployers should require this pack contractually and record where it lives; providers — including any team that materially modifying a bought model and releasing it under the company name — must produce it. This is the longest-lead item and the one that stalls conformity assessments.
Step 8 — choose the conformity route. Article 43 fixes the route: some Annex III systems follow an internal-control procedure where the provider self-certifies; others need a notified body to examine the quality-management system and technical documentation before CE marking. The route depends on the Annex III area, so classify first, check the route second. Either way you can start with the common groundwork: quality-management system, change-control record, data-governance documentation, logging design and human-oversight specification.
Step 9 — register in the EU database. Article 49 requires registration of high-risk systems before they are placed on the market or put into service, with additional entries where public authorities use them. Treat it as a controlled step: capture submission, confirmation and timestamp against the register row. A live-but-unregistered system is an easy finding — your register and the public database disagree visibly, with no inspection needed.
The AI Act audit readiness checklist: steps 10 to 12
Step 10 — transparency. Article 50 imposes what a customer sees directly: people must be told they are interacting with AI unless it is obvious; synthetic content must be machine-marked; deepfakes must be disclosed; emotion-recognition and biometric-categorisation systems must inform the people exposed. For mid-market firms this usually means chat-widget copy, metadata on generated marketing assets and a disclosure line on AI-assisted communications — store the before-and-after screenshots as evidence.
Step 11 — incident reporting. Article 73 requires providers of high-risk systems to report serious incidents to the market-surveillance authority within 15 days of awareness — sooner for the most severe cases. Deployers must report serious incidents or malfunctions causing serious harm to the provider, and directly to the authority where the provider sits outside the EU. Build it now: a “serious incident” definition in your incident taxonomy, a named duty holder per system, an authority-contact list and a one-page timeline — then tabletop-test it. Teams that first practise during a real incident miss the deadline.
Step 12 — close the loop. Providers need a post-market monitoring plan feeding real-world performance back into documentation and risk assessment; deployers keep the logs and review oversight when use changes. Add a quarterly evidence review — for each register row, can you produce the rationale, literacy record, documentation or attestation, registration, transparency copy and incident playbook? Keep the minutes: they are evidence the programme runs.
First obligations checklist after the AI Act applies: 2026 versus 2027
The Act switches on in tranches, and sequencing to those tranches is the difference between a calm programme and a panicked one.
- Live now. Article 5 prohibitions and Article 4 literacy have applied since 2 February 2025 — no screen or literacy evidence means you are already behind.
- From 2 August 2026. The general applicability date brings transparency duties, enforcement powers and the penalty regime running to €35 million or 7% of global turnover for prohibited practices — the year the checklist turns from preparation into exposure.
- 2027 for high-risk duties. The full high-risk set — documentation, conformity, registration, deployer duties — was delayed by the Digital Omnibus process. Plan against 2 August 2027 as your working date, but confirm it against the published regulation before finalising any board paper.
A practical consequence: run steps 1 to 6 against the year the duty bites — scope, the screen, ownership, literacy and the register cost little and de-risk everything else; documentation, conformity and registration are the expensive steps with the most lead time.
How to check your own position in three minutes
Renomira’s free exposure scan asks twelve questions about how your organisation actually uses AI and how it is governed, then returns an exposure score, the ranked exposures with the exact article that fired on each one, and the first three moves mapped to the dates that apply. Confirming an email address unlocks the full report.
Frequently asked questions
Do we need all twelve steps if we only buy off-the-shelf AI tools?
Buying rather than building changes your role — you are primarily a deployer — but it does not remove the checklist. Steps 1 to 6 apply in full; steps 7 to 9 shift to contracting, needing vendor attestations and documentation rights; steps 10 to 12 still apply to customer-facing deployments. Only creating a model yourself falls away.
Which steps are legally required today, right now?
Two duties are already live: the Article 5 prohibited-practices screen and the Article 4 AI-literacy obligation, both applicable since 2 February 2025. Everything else prepares for obligations switching on at later commencement dates, with transparency duties and penalties under the general applicability date and the high-risk set following in 2027. Preparing early is voluntary; being caught without a screen or literacy evidence is not.
Who should own this programme internally?
Give one named person overall accountability — usually the Head of Legal or the DPO — and distribute step ownership to the roles that control the work: CISO for vendor and incident steps, Head of People for literacy, engineering leads for documentation, compliance officer for register and classification. Splitting the work is realistic; splitting the accountability is not.
What if our AI vendor will not share its technical documentation?
Ask in writing and record the request — a refusal is itself a data point at audit. Many vendors supply a conformity statement instead of the full pack, which is workable for a deployer. Silence on Article 43 conformity, Article 49 registration and Article 73 incident cooperation is a procurement risk: a high-risk system from a vendor who cannot demonstrate its route to conformity is a liability you inherit.
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
Build a live AI register in minutes — upgrade only when you need the audit pack.