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

How to Run an AI Act Gap Assessment: A Practical Method

A structured AI Act gap assessment turns 100+ obligations into a sized backlog. This guide covers method, interviews, evidence, scoring and the output that gets funded.

Renomira Team · 6 min read
Clipboard checklist representing an AI Act readiness gap assessment
Photo: RyanT27 / CC0 via Wikimedia Commons

An ai act gap assessment is the bridge between reading the Regulation and knowing where you stand: a time-boxed exercise that turns an abstract list of obligations into a sized, owned, prioritised backlog. Firms that skip it end up with compliance work organised by whoever shouts loudest — usually the tool that caused the last incident. Firms that run it properly get one artefact their board can fund. This guide gives you the method: how to scope it, who to interview, what evidence to collect, how to score gaps honestly, and how to turn a current state assessment ai regulation into a plan.

Scope first: decide what is in the assessment

The assessment covers your AI population, not your IT estate. Define three boundaries before anyone opens a spreadsheet:

  • Tool boundary. A tool is in scope when it infers from input to produce predictions, content, recommendations or decisions. Business intelligence that only aggregates and reports is usually out; anything generating outputs that influence decisions is in. Write the rationale either way — out-of-scope decisions are audit evidence too.
  • Role boundary. For each in-scope tool, note whether you are provider or deployer. Your gap list for a tool you merely use is a fraction of the one for a tool you place on the market.
  • Time boundary. Fix an assessment date and a system-version snapshot. Gap assessments that chase a moving estate never finish.

A mid-market firm typically lands between 15 and 60 in-scope tools. If you find 400, your boundary is wrong: you are inventorying loggers and recommendation widgets that carry no AI Act weight. Split the inventory (a register of everything) from the assessment population (the subset that matters).

The five-step method

Step 1 — Inventory and interviews (week 1). Pull the tool list from procurement, IT and shadow-IT discovery, then run 45-minute interviews with the owning function: HR for screening tools, finance for scoring, marketing for generative content. Ask three questions per tool: what does it decide or generate, who built it, what data goes in. Interview notes are where unmanaged AI surfaces — the spreadsheet model a team calls “just a formula”.

Step 2 — Classification pass (week 1–2). Apply the two-door test (Annex I product link, Annex III use) to each tool. For deployer tools, confirm the declared intended purpose matches how the tool is actually used — purpose drift is a classification gap in itself. Record provider/deployer, high-risk yes/no, and the article reference.

Step 3 — Obligation mapping (week 2). For each in-scope tool, map the applicable obligation set: providers get the Chapter II–III lifecycle list; deployers get Article 26 duties, registration where required, FRIA under Article 27 if in its scope, AI literacy under Article 4 for everything. This becomes the checklist per tool — the ai act gap assessment checklist your assessor walks.

Step 4 — Evidence test (week 2–3). For every obligation, ask one question: if the authority asked today, what file would we send? Three states only: evidenced (the artefact exists, dated, owned), partial (something exists but stale, unowned or incomplete), absent. Self-assessments where the team grades their own work always drift optimistic — have a second person spot-check ten randomly chosen “evidenced” claims.

Step 5 — Score and prioritise (week 3). Score each gap: severity (regulatory exposure) × effort (days to close) × breadth (how many tools it touches). Output a backlog with owners and quarter-level dates — a eu ai act maturity assessment radar is useful for the board slide, but the backlog is what gets worked.

The scoring rubric that survives challenge

Keep the scale small enough that two assessors would agree:

State Definition Priority default
Evidenced Artefact exists, dated, owned, matches current version None — monitor
Partial Exists but stale, unowned, or version-mismatched Medium — fix in cycle
Absent — control The process or document does not exist at all High — unless scope contested
Absent — contested scope Gap exists but classification is disputed Resolve classification first
N/A with rationale Out of scope, with the one-line reason recorded None — audit evidence

The two traps: scoring everything “high” (which funds nothing) and letting contested-scope gaps sit unowned for a quarter (which is where the regulator finds you). Force a classification decision within two weeks of flagging a contested item — escalate to a named decision-maker with a deadline.

Common findings, ranked by frequency

Across mid-market assessments, the same gaps recur: no AI register or a register that stopped being updated; classification recorded as opinion rather than rationale; vendor contracts with no AI Act allocation (who is provider, who provides documentation, who responds to authorities); Article 4 literacy untrained and unrecorded; human oversight described as policy but not assigned to a role in operating procedure; logs retained shorter than the Article 26(6) period of at least six months; incident processes modelled only on personal-data breach clocks. None of these are exotic — and four of the six are cheap to close.

What good looks like: artefacts by role

Your evidence test is only as good as your map of what should exist. For a provider, the exemplar file set is: register entry with classification rationale; technical documentation (Annex IV) at current version; risk-management record with dated iterations; data-governance methodology plus dataset statements; validation report against intended purpose; logging specification; QMS document; conformity assessment record and declaration; post-market monitoring plan; incident procedure with Article 73 timelines.

For a deployer: register entry and use-confirmation with the vendor; assignment of human oversight to named roles; instructions-for-use received and acknowledged internally; log retention configured to at least six months; FRIA where Article 27 catches you; Article 26(7) flow-down where you are a regulator or authority; Article 4 literacy training records; a procurement rule that AI vendors are asked for their declaration before signature.

If an artefact exists in neither list’s expected shape, that is the gap — no scoring debate required.

Making the output fundable

The assessment fails in the last mile if it lands as a 60-row spreadsheet nobody reads. Convert it into three artefacts: a one-page heat map (tools × obligation families, colour by state), a backlog with owners and effort estimates (the working document), and a decision list — the three or four contested classifications or funding choices that only leadership can settle. Present decisions first: the honest headline is rarely “we are non-compliant”, it is “we cannot answer whether tool X is in scope until someone decides what tool X is for”. Boards fund decision clarity faster than they fund abstract obligations.

Re-run the full assessment annually and the evidence test quarterly. The quarterly pass on a sample of ten items takes half a day and catches the true killer — not absence of controls, but controls that quietly went stale.

How to check your own position

Run a compressed version this week: take your ten most consequential AI tools and apply the four-state test to five obligations each — classification rationale, provider documentation received, oversight assigned, literacy trained, logs retained. Score the grid, count the reds, and pick the single gap that touches the most tools. If classification itself is red for more than two tools, stop and fix that first: every other score is provisional until the population is right. Record the whole exercise with its date — even a half-day pass gives you a defensible starting point, and “we assessed, found these five gaps and began here” is a far better posture with any authority than a perfect programme started too late.

Frequently asked questions

What is an AI Act gap assessment?

A structured comparison of your current AI estate against the Regulation’s obligations: inventory your AI tools, classify each, map the applicable duties, test whether evidence exists for each duty, then score and prioritise the differences. The output is a remediation backlog with owners — the basis of an eu ai act maturity assessment and the document that turns compliance from a reading project into a managed programme.

How long does a gap assessment take?

For a mid-market firm with 15–60 in-scope tools, three weeks of elapsed time with a part-time cross-functional team is realistic — roughly one week to inventory and classify, one week to map obligations, one week to test evidence and score. The first assessment is the longest; quarterly evidence tests on a ten-tool sample take half a day once the method exists.

What is the difference between an AI inventory and a gap assessment?

An inventory lists what exists — every tool with AI features, including the ones you judged out of scope. The gap assessment works the classified subset: it asks whether each applicable obligation is evidenced, partial or absent. The inventory feeds the assessment, but a register alone answers no compliance question.

Who should run an AI Act gap assessment?

A small core — typically compliance or risk with IT and one business representative — with interviews across the owning functions. Two-person verification on sample evidence is what keeps scoring honest, since self-graded assessments drift optimistic. Leadership’s role is not the legwork; it is resolving contested classifications and funding the backlog the assessment produces.

What should a gap assessment deliver to the board?

Three things: a heat map showing where the estate stands, a prioritised backlog with owners and effort, and a short decision list of the contested classifications or funding choices only leadership can settle. Leading with decisions matters — the board’s scarce attention is wasted on a 60-row spreadsheet and well spent on the three questions that unlock the rest.

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 gap assessmenthow to do a gap analysis for eu ai acteu ai act readiness assessment templateai compliance gap analysis stepscurrent state assessment ai regulationai act gap assessment checklistassessing your ai systems against the ai acteu ai act maturity assessment

Keep reading