How to answer AI questions in a supplier security questionnaire

Written and last checked 9 October 2026 against Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744.

Answer each AI question for named systems, in five parts: a direct answer, the control you run, the systems it covers, the evidence with its date, and the exceptions. Give every answer one of four statuses and never claim more than the evidence shows.

Who receives these questionnaires, and what do they look like?

Suppliers to banks, insurers, the public sector and large companies: anyone whose customer runs a vendor-risk review. It arrives as a spreadsheet or a Word document. Some customers use a standard one, such as SIG from Shared Assessments or the CAIQ from the Cloud Security Alliance, which also publishes an AI Controls Matrix (AICM). Many write their own.

The AI questions rarely cite an article, so the first job is to find which provision, if any, a question is about. Part of what the buyer collects is evidence for its own file: a deployer of a high-risk system must take appropriate technical and organisational measures to use it in accordance with its instructions for use (Article 26(1)), and it can only do that with what you send.

How should each answer be built?

Five parts, in this order. The example answers question 10 of the fifteen, on reviewing the AI's output, for an invented product; it is an illustration, not a real supplier's answer.

Direct answerYes, no or partly, first, for the systems asked about.

ExampleYes, for the two AI features that draft text for users.

ControlWhat you actually do, as a practice, not a policy title.

ExampleThe user accepts, edits or rejects every draft before anything is sent.

ScopeWhich systems the answer covers, and which it does not.

ExampleReply drafting and ticket summaries. Not the spam filter.

EvidenceThe record or screenshot that proves it, with its date.

ExampleA screenshot of the review step and the log entry of one accepted draft, dated 1 October 2026.

ExceptionsWhere the control does not apply or is not in place yet.

ExampleThe spam filter moves messages without review; each move is logged and can be undone.

The order matters. A reviewer decides on the first words, and an answer that reaches "partly" in its fourth sentence reads as evasive even when it is honest.

Which status should each answer carry?

• ImplementedThe control runs today for every system in scope, and you can show the evidence.
• Partially implementedIt runs for some systems or some steps. Name them in the scope.
• PlannedNot in place yet. Give an owner, and a date only if someone has committed to it.
• Not applicableThe duty does not reach the system. Say why, and attach the classification.

A later date in the law is not a status. A duty that applies from 2 December 2027 under Article 113, third paragraph, point (c), is still something you either do or plan; put its date, with its article, in the exceptions.

Which AI questions come up, and where are they in the Act?

Fifteen recur, each with the provision it is about, in which AI questions supplier questionnaires ask. Rows 8 to 14 there apply only to high-risk systems; for any other system they are Not applicable, with the classification as the reason.

What are the common mistakes?

• Answering for the company. Article 6 classifies systems, not companies.
• "Yes" with no evidence. If you cannot attach it, the status is Partially implemented or Planned.
• A policy title as the control. It says nothing about what happens to an output.
• Mixing up the roles. Article 50(1) is a provider duty; Article 26 is the deployer's. Which one you are is set per system: see provider or deployer.
• Not applicable with no reason. The reason is the classification.
• Dropping the exception. The summary that leaves out "unless" is the one that is wrong.

How does an answer trace back to its evidence?

An answer that holds up ends a chain: the question, its provision, the system and its classification, the control and the evidence with its date and owner. A missing link is a status: no evidence makes it Partially implemented, no control makes it Planned, and no classification makes the answer a guess. Classify first, in the order how to classify your AI system sets out, and the high-risk rows answer themselves for every system that is not. The four Article 50 disclosures cover rows 5 and 6 of the fifteen provision by provision.

Where is the template?

Download the Excel template. No email is asked. It carries the fifteen questions with their provisions, a column for each of the five parts, a status column limited to the four statuses, and an owner column.

Send us a questionnaire you received (remove the names). We'll show you, for free, how many answers you can already prove: hello@complipath.io.

Complipath is adding questionnaire answers to the app. It is not available yet.

What this means for you

If you build the AI (provider): the transparency rows of the fifteen questions are yours under Article 50(1) and (2), and for a high-risk system so are rows 8 to 13. Keep each system's classification with its date; it is the evidence behind every Not applicable.

If you use someone else's AI (deployer): Article 4 is yours whatever the tier, rows 5 and 6 are yours where Article 50(3) or (4) reaches you, and for a high-risk system so are rows 9, 10 and 14 under Article 26. Ask your own providers the same questions; their answers are your evidence.

Complipath records each system's classification and the articles it rests on: see what a risk classification records.

FAQ

We only use a general-purpose AI tool internally. Do the AI questions apply to us? Yes, the ones about you as a deployer. Article 4 applies whatever the tier, so the literacy questions need a real answer. The high-risk rows are Not applicable for a tool that is not high-risk, and the classification of that tool is the evidence you attach to say so.

Can we answer Not applicable to every high-risk question? Only for systems that are not high-risk, and only with the reason. Answer per system: the classification, the route it rests on, and its date. If one of your systems is high-risk under Article 6(2) and Annex III, its rows need real answers, with dates from Article 113, third paragraph, point (c).

Is there an AI Act certificate we can attach? Not for a system that is not high-risk: the conformity assessment of Article 43 concerns high-risk systems. For those in points 2 to 8 of Annex III, Article 43(2) sets an internal-control procedure with no notified body. What you can attach is each system's classification, the articles behind it and the version of the law.

Should Planned answers carry a date? Only a date someone has committed to. A Planned answer with an owner and no date is honest; one with a date nobody owns becomes a promise the buyer will hold you to. Where the law sets the date, as Article 113 does for the high-risk duties, give that date with its article.

Where this question meets the product: the questionnaire answer library.