Why the AI section appeared
More and more, closing a deal means answering a security questionnaire, and many now include a section on AI. Buyers have extended their third-party risk review to cover the AI tools you use. A weak or missing answer can stall a sale. A clear one keeps it moving.
When a company buys from you, your risks become theirs. AI adds a few new ones: whether your tools train on the information you are given, whether a person reviews AI output before it reaches the client, and whether anyone owns the tools in use.
What the AI section usually asks
Policy and use
- Do you have an AI acceptable use policy, and who owns it?
- Can you list the AI tools and AI features your team uses?
- Do your AI tools train on client or company data, and can that be turned off?
- Does a person review AI output before it reaches a client or informs a decision?
Vendors, data and security
- How do you choose and review third-party AI tools?
- Where does data go, how long is it kept, and who can access it?
- How are prompts, outputs and connected accounts secured?
- Do you have a process for AI-related incidents, such as wrong output reaching a client or information entered into the wrong tool?
How to answer
Answer with what you do today, not what you plan to do. Point to a named, dated policy and the practice behind each answer. Mention the records you can share, such as your acceptable use policy, your list of AI tools and your vendor reviews.
Be precise about data: whether a tool trains on what you enter, whether you have turned that off, and how long it keeps information. If the honest answer is “not yet,” say so and give a date.
A policy is not proof of practice. If your policy requires review of AI-assisted work, confirm that the review actually happens before you answer “yes.”
Have every answer reviewed before it goes out. The person who owns the practice confirms the facts, and anything that commits your company goes to whoever approves client commitments.
A worked example: three reviewed answers
Meridian Creative is a fictional 42-person agency. A client’s security questionnaire asked about AI in three places. Here is what Meridian sent, the record behind each answer and who confirmed it.
| Buyer question | Reviewed answer | Evidence | Confirmed by |
|---|---|---|---|
| Do your AI tools train on our data? | No. Client information goes only into tools that do not train on it. Our meeting summary assistant trains on inputs by default, so we turned that off and use it for internal meetings only. | Vendor review of the meeting summary assistant, with a screenshot of the training setting | Operations lead |
| Does a person review AI output before it reaches a client? | Yes. AI output is a draft. A named reviewer checks facts, rights and client suitability before anything is sent. | Acceptable use policy adopted October 8, 2026, and reviewer sign-offs on recent client work | Creative director |
| Can you delete our information from your AI tools on request? | Not yet confirmed. The vendor offers deletion on request. Our first witnessed deletion test is due October 15, 2026, and we will send you the result. | Vendor review with the open action, its owner and due date | Operations director |
Each answer describes what Meridian does today and points to a record the client can ask to see. The third is an honest “not yet” with a date. A dated gap is easier for a buyer to accept than a yes nobody can back up.
When a question names a standard or a law
Some questionnaires ask how your program relates to a named standard, framework or law. Treat it like any other question: describe what you actually do and the records that support it.
Don’t claim compliance or alignment you haven’t established, and never say “certified” unless you hold the certificate. If the real question is whether a law applies to your company, take it to your legal counsel instead of improvising an answer.
Using your AI assistant
Many teams draft these answers with an AI assistant. That works when the assistant drafts only from your own sources, shows what it cannot support, and a person checks every claim.
Use only your company’s approved AI account. Don’t paste client questionnaires or confidential information into personal tools; the questionnaire is probably asking about exactly that.
These three prompts come from the free starter. Run them in order in one conversation. (The starter’s first prompt sets your sources and scope; it’s in the free download.) Replace the bracketed inputs. These instructions guide the model; they cannot guarantee that it follows them. Check the original evidence yourself.
1. Map the questions to evidence
Using the corrected source inventory and scope, map the following questions to the provided evidence.
Questions, preserving the client's exact wording and IDs:
[paste questions]
Keep each original question and ID intact. If it contains multiple requirements, list those requirements separately under that ID. For each requirement identify a relevant source ID, section/page if available, and a short supporting excerpt. Do not invent page numbers, quotations or document titles.
Classify each requirement as Supported, Partial, Missing, Conflicting, or Scope unclear. Supported means the evidence supports that specific claim within the stated scope; it does not mean a human has approved the answer. If the question asks whether a control operates, policy language alone is Partial at most. Do not infer that a control does not exist merely because no evidence was supplied.
For claims containing ALL, EVERY or BEFORE, a few example records cannot establish universal coverage. A recorded exception may disprove a universal positive claim, but do not mark the broader control absent without evidence. Treat “AI-assisted” and “generative AI” as distinct unless a source explicitly connects them. Flag that distinction as unresolved when it matters.
List what remains unverified and the role best placed to resolve it. Suggest a role, never invent an individual's name. Do not force a yes/no answer when the evidence is incomplete. Return the mapping for every input question and report input and output question counts. Stop for my review.2. Draft the answers
Using the reviewed mapping, draft answers to the questions. Use the client's required answer format and limits where supplied. Otherwise use concise plain language.
Return one entry per original question with:
1. Question ID and exact wording.
2. Draft client-facing answer.
3. Supporting source IDs and precise available locations.
4. Evidence status: Supported, Partial, Missing, Conflicting, or Scope unclear.
5. Unresolved facts and suggested reviewer role.
Separate the client-facing draft from internal notes. Do not silently combine incompatible sources or change scope. Do not invent controls, certifications, contractual commitments, data-handling settings, dates, approval, or evidence. Do not turn a policy requirement into an assertion that it is implemented.
If a material part of the answer is unknown, mark the draft HOLD FOR REVIEW and explain the missing fact in the internal notes. If a mandatory yes/no field cannot be supported, leave the proposed selection unresolved. Do not fabricate a yes, no, or not-applicable answer just to complete the form.
All answers remain drafts, including those with Supported evidence status. Do not send or submit anything.3. Challenge the draft
Review every draft answer against the original questions and the supplied sources. This is a second drafting check, not independent verification.
Check for unsupported claims; invented or mismatched citations; policy-versus-practice confusion; vendor-versus-company confusion; stale or conflicting sources; wrong service or time period; unanswered subquestions; and words such as all, never, always and guaranteed that exceed the evidence.
Specifically challenge whether sample records were generalized to ALL or EVERY, whether a review was completed BEFORE release, and whether evidence of AI assistance really establishes generative AI use.
For each issue, show the question ID, problematic wording, reason, and a narrower supported revision or HOLD FOR REVIEW. Also identify sensitive details that appear unnecessary for the client's question. Do not conceal relevant shortcomings or weaken facts to make the company look better.
For each unresolved item, draft one concrete question for its human owner and identify the evidence needed. Report input and reviewed question counts and flag any missing questions. Do not assign final approval or invent a confidence score.The free starter has all five prompts, with a response record, a fictional worked example and a human review checklist.
Have the records ready before the next one
The biggest driver of a fast, credible response is having the records ready before the questionnaire arrives. Teams that write a policy mid-deal lose days. Teams with the records on hand spend their time reviewing answers, not writing policy.
A short set of records covers most AI questions:
- An acceptable use policy with a named owner and a date
- A list of the AI tools in use, each with an owner
- A review of each AI vendor: data use, retention and access
- A simple process for AI-related incidents
If you don’t have a written policy yet, the Free Policy Starter gives you an editable draft to review.
Do this now
Draft your first answers
The Free Client AI Questionnaire Starter gives you five guided prompts, a response record and a fictional worked answer. Use your own policies to prepare a draft you can review. Free. No email.
When you need it on record
AI Questionnaire Response Kit
Answering client questionnaires regularly? Keep reviewed answers and evidence references ready for the next one, with a record of who approved each answer.
Common questions
What is the AI section of a security questionnaire?
A set of questions a buyer adds to its third-party risk review to understand how you choose, use and secure AI tools.
Do we need an AI policy to answer it?
In practice, yes. Most questions assume a written acceptable use policy and a list of the AI tools you use. Without them you are answering from memory, and buyers can tell.
What if we only use third-party AI tools?
You still answer. The questions cover how you choose, govern and secure the tools you rely on, whoever built them.
Can we use AI to draft the answers?
Yes, if it drafts from your own sources in your company’s approved AI account, flags what it cannot support, and a person checks every claim before anything is sent.
Can we say we comply with a standard or a law?
Only say what is true and what your records support. Never say “certified” unless you hold the certificate. If you are unsure whether a law applies to you, ask your legal counsel first.
Related guides
Extrasphere Media LLC. Content for general informational purposes, not legal advice and not a substitute for qualified counsel. Independent and not affiliated with ISO, the IEC, NIST, or the European Union.