The EU AI Act vendor questionnaire, answered question by question.

Enterprise buyers now add an EU AI Act section to their vendor questionnaires. It usually arrives inside the security review, between the SOC 2 questions and the data processing addendum, and it is often written by someone who has read the Regulation once. The questions are predictable. This guide lists the ones that come up, what each is really asking, how to answer it in a sentence or two, and what to attach so the answer holds up.

The buyer is usually a deployer under Article 3(4). Your answers become part of their own compliance file. Short, specific, dated answers with an Article reference close the review. Long reassurance without evidence reopens it.

Before you answer anything

Every question below depends on three facts. Write them down once, date them, and reuse them across every questionnaire:

• Your feature inventory. Which features run a model doing inference, and which are plain rules. The Act regulates AI systems (Article 3(1)), feature by feature, not companies.
• Your role per feature. If you ship it under your own name, you are its provider (Article 3(3)), even when the model underneath is someone else's.
• Your risk tier per feature. Prohibited (Article 5), high-risk (Annex III or Annex I), limited (Article 50 transparency) or minimal. Most B2B SaaS features land in limited or minimal.

If you do not have these yet, that is the first gap to close. The free risk-tier scan covers the tier in about two minutes; a full written position takes longer.

The questions procurement asks, and how to answer them

1. "Does your product use AI within the meaning of the EU AI Act?"

What they want: to know whether they need to assess you at all. How to answer: yes or no per feature, with the definition named. "Features A and B are AI systems under Article 3(1). Feature C is deterministic and is not." Do not answer "no" because you call a third-party API: the system you ship is still an AI system.

Attach: your feature inventory, one line per feature. An inventory template with a worked example.

2. "Are you a provider or a deployer?"

What they want: to know whose obligations sit where. How to answer: "We are the provider of features A and B under Article 3(3). You are a deployer under Article 3(4) when you use them." Then say what would change that: under Article 25, a customer who puts its own name on the system, substantially modifies it or changes its intended purpose can become the provider. Buyers who white-label or fine-tune need to hear this.

Attach: a role statement per feature. The provider vs deployer guide explains the edge cases.

3. "What risk category does your AI fall into?"

What they want: to know whether they are buying a high-risk system, which would bring deployer duties under Article 26. How to answer: the tier per feature and the reason in one line. "Minimal risk: summarises support tickets, takes no decision about a person." The free high-risk checker runs the Annex III questions for one feature. If a feature sits in an Annex III area but you rely on the Article 6(3) exemption, say so and say that the assessment is documented, as Article 6(4) requires.

Attach: a dated classification memo. If you rely on Article 6(3), attach that assessment too.

4. "Do you use any prohibited AI practices?"

What they want: a clean no under Article 5, which has applied since 2 February 2025. How to answer: "No. We have checked our features against each Article 5 practice." If you offer anything near the line, such as sentiment analysis on employees or students, explain why it is not emotion recognition in the workplace or in education, which Article 5 prohibits.

Attach: the Article 5 check as part of the classification memo.

5. "Which AI models do you use, and who is responsible for them?"

What they want: to know the general-purpose AI model provider and whether its duties are met. How to answer: name the model provider for each feature. Obligations for general-purpose AI model providers under Article 53 have applied since 2 August 2025 and sit with the model provider, not with you, as long as you call the model rather than substantially retrain it. Say which of those is true. Where the GPAI line sits.

Attach: your sub-processor list and links to the model provider's published documentation.

6. "How do you meet the Article 50 transparency obligations?"

What they want: to avoid exposing their own users to an undisclosed bot. How to answer: per paragraph. Article 50(1) chatbot disclosure has applied since 2 August 2026: say where in the interface the user is told they are talking to an AI system, and quote the wording. Disclosure templates and examples. Article 50(2) machine-readable marking of synthetic content applies from 2 December 2026: say whether you generate audio, image, video or text output and who marks it, you or the model provider. Article 50 in full.

Attach: a screenshot of the disclosure at first interaction, and a note on content marking.

7. "How do you ensure AI literacy among your staff?"

What they want: evidence for Article 4, which has applied since 2 February 2025 to providers and deployers alike. How to answer: who was trained, on what, and when. "All staff who build or support our AI features completed a session on capabilities, limits and failure modes on [date]. We repeat it when the features change." What Article 4 actually requires.

Attach: the dated training record. Minutes are enough; a certificate is not required.

8. "Describe human oversight, logging, accuracy and robustness."

What they want: the controls Articles 12, 14 and 15 require of high-risk systems. How to answer: if your feature is high-risk, answer against each Article, and note that Annex III obligations apply from 2 December 2027 and Annex I from 2 August 2028. If it is not high-risk, say so first, then describe the controls you run anyway: human review of outputs, logs you keep, how you test accuracy. Do not imply a conformity assessment you have not done.

Attach: a short controls description. For high-risk systems, the instructions for use that Article 13 requires.

9. "Do you train models on our data?"

What they want: mostly a GDPR and confidentiality answer, asked under the AI Act heading. How to answer: a plain yes or no, what data, under which contract term, and whether the model provider retains inputs. Point to the data processing addendum rather than restating it.

Attach: the DPA and the model provider's data-use terms.

10. "Do you hold an EU AI Act certificate or CE marking?"

What they want: a shortcut. How to answer: honestly. Conformity assessment under Article 43 and CE marking apply to high-risk systems only. There is no AI Act certificate for limited-risk or minimal-risk systems, so the correct answer for most vendors is: "Not applicable. Our features are not high-risk. Our classification is documented and dated, and available on request." A self-assessment is not a certification, and should never be presented as one.

11. "How will you tell us about changes and incidents?"

What they want: to know they will hear before a feature's classification changes. How to answer: commit to notifying material changes to AI features, and to re-issuing your classification when a feature changes its purpose or tier. For high-risk systems, serious incident reporting under Article 73 applies on top.

Attach: your changelog or release-notes link, and the notification clause you are willing to sign. Which AI Act contract clauses to sign, and which to narrow.

The evidence pack to keep ready

Keep one version of each, dated, and send the same set to every buyer. Inconsistent answers across two questionnaires are what make a reviewer dig.

Answering well

• Name the Article. "Article 50(1), disclosed in the chat greeting" closes a question. "We take transparency seriously" does not.
• Date everything. Deadlines moved with the June 2026 Omnibus. A dated answer shows which rules you assessed against. Every date in one table.
• Answer per feature. One high-risk feature does not make your whole product high-risk, and one minimal-risk feature does not clear the rest.
• Say "not applicable" with a reason. Most Chapter III questions do not apply to most SaaS. Say why, once, and reference the classification memo.
• Do not overclaim. "Compliant with the EU AI Act" is a claim a reviewer can test. "Classified as limited risk, Article 50(1) disclosure in place since [date]" is a fact they can check.

A Reglynn gap report gives you the classification memo, the role analysis, the obligations that apply to your answers, each with its post-Omnibus deadline and the Article references, in one dated document you can attach to a questionnaire. See a sample report for a fictional product before you decide. It is a self-assessment, not a certification and not legal advice.

Frequently asked questions

Do I need an EU AI Act certification to pass a vendor review?

No. Conformity assessment under Article 43 and CE marking apply only to high-risk AI systems. For limited-risk and minimal-risk systems there is no AI Act certificate. Buyers can expect a dated classification, a role statement and evidence of Article 50 and Article 4 compliance instead.

We only call an OpenAI or Anthropic API. Can we answer that the AI Act does not apply to us?

No. The model provider carries the Article 53 duties for the general-purpose AI model, but the feature you ship is your AI system and you are its provider under Article 3(3). Answer per feature: its tier, its Article 50 disclosure and your Article 4 training record.

What should I attach to an EU AI Act vendor questionnaire?

A feature inventory with your role per feature, a dated classification memo including the Article 5 check, any Article 6(3) assessment you rely on, a screenshot of your Article 50(1) disclosure, your Article 4 AI literacy record, and your sub-processor list with the model provider's terms.