How to Answer an Enterprise AI Vendor Questionnaire in One Afternoon
The questionnaire arrives with a two-week deadline and no brief. This is the framework for answering it in an afternoon — and turning it into an asset you never rewrite again.
The email is always the same: a spreadsheet with 140 rows, a deadline of “ideally by Friday”, and a note that the deal cannot proceed without it. Someone in procurement has discovered AI.
Questionnaires are not really a test of your AI programme. They are a test of whether you can produce evidence quickly — and the difference between a four-hour answer and a four-day one is almost entirely preparation.
What buyers are actually asking
Strip the wording back and every AI questionnaire has six questions underneath it:
- Do you know what AI you use? → your inventory.
- Is any of it high risk? → your classification and its reasoning.
- Who is accountable? → a named owner, not a committee.
- How do you control it? → policy, oversight, access, logging.
- What happened last time something went wrong? → incidents and response.
- What happens when the law changes? → re-classification, vendor notification, review cadence.
That is it. Everything else — model cards, training data provenance, evaluation methodology — is a variation on one of those six.
Build the evidence pack once
Assemble five artefacts before you answer anything. Together they cover roughly 80% of any AI questionnaire, and all five already exist in a functioning compliance programme.
1. The AI register. One row per system: name, vendor, purpose, provider/deployer role, risk tier, classification rationale, data categories, owner, review date. This single artefact answers every “list your AI systems” question in the pack.
2. The classification rationale. Not the tier — the reasoning. “ChatGPT Enterprise — minimal risk; internal productivity use, no decision about a natural person; Art. 4 literacy applied” is an answer. “Low” is not.
3. The AI use policy. Two pages: approved tools, prohibited uses, data rules, escalation, review cycle. Buyers ask for the document even when they never read it.
4. The obligation map. Which articles apply to which system, by role. Article 26 for high-risk deployers, Article 50 for transparent interaction, Article 4 always. One table converts a legal question into a lookup.
5. The incident log. Dates, severity, what was affected, what changed. An empty log with a documented process is a fine answer; a missing log is not.
If you have those five, the questionnaire becomes a mapping exercise rather than a research project.
Answer blocks you write once
Most questionnaires repeat themselves badly. Draft five blocks and reuse them verbatim.
On data handling. Where the data goes, whether it trains models, retention, region, sub-processors, and the contractual basis. Have the vendor’s current position in writing — model providers change this annually.
On classification. How you assess, against which annex points, who signs off, when it was last re-run. Mention the rule-pack version if you have one; versioning reads as maturity.
On human oversight. Who can override, how overrides are recorded, what training the overseers hold. Name the role, not a person, so the answer survives turnover.
On logging and audit. What is logged, where, for how long, who can access it. Note the Article 26(6) six-month floor for high-risk deployers — it signals that you know the number.
On change management. What triggers re-assessment: a vendor release, an incident, a rule-pack update, a new regulation, annually. Buyers want to know the answer will not be stale in eighteen months.
Answering fast: the method
1. Sort by section, not by row order. Group the questionnaire into your six questions. Answering in the buyer’s row order destroys context switching.
2. Fill the confident rows first. Usually 60–70% are direct matches to your existing evidence. Bank those, then spend the time on the genuinely open items.
3. Never guess a yes. For anything you cannot evidence, answer honestly with the current state and a date: “Partially — the register covers production systems; internal prototype tools are in scope by Q2.” Buyers distrust a perfect questionnaire more than a candid one with a plan.
4. Attach rather than describe. “See AI Register (extract attached)” beats three paragraphs of narrative, and it is harder for the buyer’s reviewer to misread.
5. Log your own answer in your register. Every questionnaire is a snapshot you will be asked to reproduce at renewal. Store the date, the answers and the evidence version together.
The rows that actually stall deals
Four rows routinely cost days, and each is a known gap:
- “Provide your AI impact assessment / FRIA.” If you have no Article 27 trigger (public body, public-services provider, creditworthiness or essential-services deployer), say so explicitly with the reasoning — that is the answer. If you do have a trigger, produce the assessment.
- “Provide model cards for all AI systems.” Attribute correctly: where the model is third-party, the model card belongs to the provider. Ask them for it and reference the date received.
- “State your AI Act compliance status per system.” This is the register export. If you cannot produce one, you have found your real gap — and it is the same gap the regulator would find.
- “Who is your AI governance lead?” A named person with a named deputy. Committees do not answer questionnaires.
Turn it into a reusable asset
After the first questionnaire, save three things: the answers you gave, the evidence versions you attached, and every question you could not answer. The third list is your compliance roadmap, prioritised by what real customers asked for rather than what a framework suggests.
Two quarters of that and the four-hour response becomes a two-hour one — which is the point, because the questionnaire is a symptom. The underlying question, from the buyer and from the regulator alike, is whether you can show your work.
Frequently asked questions
What is the fastest way to respond to an AI vendor questionnaire?
Build the evidence pack once — AI register, classification rationale, AI use policy, obligation map, incident log — then answer in question groups rather than row order using reusable answer blocks. With those five artefacts a 140-row questionnaire is typically one working afternoon.
Do we have to complete a customer’s AI questionnaire to close a deal?
Not always, but a refusal usually ends the review. A better position than refusing is a scoped response: answer the questions that apply to your role, state clearly where a duty sits with the model provider rather than you, and attach the evidence you do hold.
What if we have not classified every AI system yet?
Say so, with the current coverage and a date. “92% of production systems classified as of this month; internal tooling completes in Q4” is a credible answer. Claiming full coverage you cannot evidence creates a finding at audit and a breach of contract at renewal.
Is a fundamental rights impact assessment required to answer buyers?
Only where Article 27 applies — deployers that are public bodies or private entities providing public services, plus deployers of Annex III creditworthiness and essential-services systems. Outside that set, the DPIA under GDPR is usually the assessment buyers accept, and you should state the reasoning for why no FRIA is triggered.
How much of our AI questionnaire answer can be automated?
Most of it. Every answer that maps to a register field or a policy paragraph can be generated from source; the genuinely human part is the reasoning for edge cases and the honest “not yet” rows. Automate the lookup, keep the judgement.
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.