How to answer the AI questions in a client's vendor questionnaire
Procurement now asks about AI before the work starts. Answer each question with what your record can show, and the answers become terms you can keep.

The AI questions used to arrive after the work, in a client meeting or beside an invoice. Now they arrive first, in the vendor or security questionnaire that procurement sends before a contract is signed. A spreadsheet with a section headed AI, a dozen rows, and a column that expects yes or no.
It is tempting to treat that section like the rest of the form: answer quickly, answer reassuringly, move on. That is the mistake. The answers end up attached to the contract, and the client's security team will read them again the day something goes wrong.
What each AI question is really asking
The wording varies between clients, but the questions fall into a small number of kinds. Each one is asking about a risk, and each one has a place in your own record where the honest answer lives.
| The question | What procurement wants to know | Where your answer comes from |
|---|---|---|
| Which AI tools do you use? | Whether there is a list, and whether it is short | The tools that actually appear on your AI sessions, not the ones licensed |
| What client data enters them? | Whether confidential material leaves the systems they know about | Your policy on inputs, and the workstreams where sessions ran |
| Is our data used to train models? | Whether their material can surface elsewhere | The terms of the plan you pay for, checked, with the date you checked |
| Who reviews AI output? | Whether a named person stands behind each deliverable | Your review step, recorded per deliverable |
| Do subcontractors use AI? | Whether the answer still holds one step removed | What your freelancer agreements require, and what they report |
| What happens if AI output is wrong? | Whether you can find the reach of a mistake | Whether sessions are tied to entries, workstreams and deliverables |
Notice that none of the questions asks how much of the work the AI did. Procurement is rarely interested in that, and you should not volunteer a figure nobody can measure. The questions are about exposure, control and accountability.
Answer from the record, not from the policy
Most first drafts answer from the policy document: what the team is supposed to do. A reviewer on the client side has read hundreds of those answers and gives them little weight. What carries weight is an answer that describes what is observed, in a form the client could ask to see.
Compare two answers to the review question. "All AI output is reviewed by a qualified team member" is a promise. "Each deliverable records the person who reviewed it and the review step they applied; AI sessions are attached to the time entry of the person who ran them" is a description of a record. The second answer is shorter on reassurance and much harder to doubt.
If you cannot write the second kind of answer for a question, that is useful to know before the contract. It tells you which part of your record to fix first, and it stops you signing a sentence you would struggle to defend. Our post on what a human review actually checks is a reasonable place to start if review is the gap.
Where an absolute yes becomes a liability
Questionnaires are built for yes and no, and some AI questions invite an absolute answer you cannot keep. Qualify them in the comments column. Procurement teams expect qualifications; what they do not forgive is a yes that turns out to have been aspirational.
- "No client data is ever entered into AI tools." Unless your tools are configured to make that impossible, the honest answer names which tools are approved for client material and on which terms.
- "Our AI providers never retain data." Retention and training terms can differ between plans of the same tool. Answer for the plan you actually pay for, and say when you last checked it.
- "AI is not used on your account." Sometimes this is the right commitment. If you make it, the record has to be able to show it was kept: workstreams for that client with no AI sessions on them.
- "We can provide full logs of AI usage." Prompts and transcripts carry the client's own confidential material and your team's working notes. Offer the structured facts instead: tool, model, session, cost, and the entry it sits on.
An illustrative example
The details here are illustrative. A consultancy is bidding for an analytics engagement with a financial services client. The questionnaire has eleven AI rows. The operations lead drafts the answers in an afternoon, working from the last quarter of the firm's record rather than from its policy.
The tools row lists two assistants and one coding agent, because those are the three that appear on sessions; a fourth tool the firm licenses but nobody uses is left off. The data row says client material is used only in the two assistants on business plans, and names the date the plan terms were reviewed. The review row describes the review step for analysis and for code separately, because they are different. The subcontractor row says freelancers on the account agree to the same tool list in writing.
On one row the consultancy writes a qualified answer instead of yes: it will not commit to zero AI use across the whole engagement, but offers to keep named workstreams free of AI sessions on request and show that in the monthly report. The client's security reviewer accepts it with one follow-up question, about where the review step is recorded, and the answer is already written.
Keep the answers as a maintained document
The questionnaire is filled in once, but the answers describe a team that keeps changing. A new coding agent, a switch of model provider, a freelancer brought onto the account: any of them can make a signed answer untrue without anyone noticing. Keep your answers in one maintained document, owned by one person, and reread it whenever the tool list changes.
The monthly report is where you show the client that the answers still hold. If the questionnaire promised a short tool list and a named review step, the AI section of the monthly report is where both appear every month, and where a change appears before the client finds it. That turns a procurement form from a hurdle into the first page of a record the client will keep reading.