The questionnaire arrives from a client's procurement or security team with a return date two weeks out, and somewhere among the familiar questions about encryption and backups is a new section on AI. At that point, the person filling it in usually has two options that both feel wrong: answer "no" to everything and hope, or describe controls the business does not have yet. There is a better route, which is mostly work you would want done anyway.
The questions that come up
The standard question sets have caught up. The Cloud Security Alliance published an AI Controls Matrix with its own questionnaire in 2025, and the Shared Assessments questionnaire that many larger clients send now includes an AI section. Where clients write their own, they tend to ask a shorter version of the same things:
- Which AI tools, if any, do you use?
- Does our information go into them?
- Is our information used to train the vendors' models?
- Where is it processed and stored?
- Do you have an AI policy, and who approves new tools?
- Does a person review AI output before it reaches us?
- Do your subcontractors use AI on our work?
- Will you tell us, or ask us, before using AI on our matters?
- How are your people trained, and what happens if something goes wrong?
Why a confident no is the riskiest answer
"We do not use AI" is the answer most businesses would like to give, and it is rarely true. Copilot Chat comes with most Microsoft 365 plans, note-takers join calls on their own and most businesses have people using ChatGPT on personal accounts.
The same goes for an answer that describes a control in the future tense as if it were in place. If the policy is being written, say so and give the date it will be approved. A security team reviewing a questionnaire is usually looking for an accurate picture and a direction of travel, and it can tell when a control exists only on paper.
What a strong answer looks like
A weak answer to "is our information used to train AI models?" reads: "No. We take the security of client data extremely seriously." It states a conclusion and gives nothing to check, so the reviewer has to come back with a follow-up.
A strong answer reads something like this, with every fact in it made true before you send it.
"Client information may be used only in the AI tools on our approved register, attached, and only on company accounts under our sign-in. The business terms for each approved tool exclude customer content from model training by default, and we review those terms every quarter. Personal AI accounts may not be used for work. Our AI policy, approved by the board on 1 September 2026, is attached."
Because the strong answer names the control, the document that shows it and how it is kept current, the reviewer has the evidence to close the question.
The documents behind every strong answer
Every strong answer leans on the same few documents, and they are worth having before the next questionnaire arrives:
- A register of approved AI tools, with the account type and the vendor's training and retention terms for each.
- A written AI policy with an owner and a review date, and the acceptable use page your people acknowledge.
- A record that sign-in to approved tools runs through your company identity.
- The result of a permissions review for any tool connected to your systems.
- The training record, and the route for reporting an AI-related incident.
- Your suppliers' own certifications. Several AI vendors now hold ISO 42001, which answers part of the supplier question for you. For the rest, checking a supplier's use of AI sets out what to ask.
Keep one set of answers
Most businesses answer the same AI questions from scratch each time a questionnaire arrives, and the answers drift. Write one set, owned by a named person, dated and reviewed on the same cycle as the AI policy. Once that set exists, each new questionnaire becomes a matter of mapping its wording onto answers you already hold.
If you use an AI tool to help draft or map those answers, treat the questionnaire like any other client document. It is often confidential, and some standard question sets restrict what their content may be used for in their terms of use, so check both before it goes into a tool.
Questions eight and nine deserve a decision
Two of the questions above are about commitments you make to clients. Whether you will tell a client before using AI on their work, or ask their permission, is a commercial position, and some clients now set it in their terms for suppliers.
Where this fits
The documents in the evidence list are the outputs of Govern, and a questionnaire with a return date is a common reason a business starts it. If the questionnaire is already on your desk, the readiness snapshot tells you quickly which tools are in use and on whose accounts, which is the part you cannot answer accurately without looking.