Standard vendor questionnaires were written for software that stores and retrieves data. AI products also process, transform and sometimes learn from it, often through third-party model providers you never contract with directly. The generic questions still matter, but they miss the places where AI systems differ. These ten target those places.
Where does my data go?
- Which model providers see my data, and under what terms? Ask for the list of sub-processors, not a general assurance. Your data may pass through a provider you have never heard of.
- Is my data used to train or improve any model? Get the answer in the contract, per provider, including any default that requires you to opt out.
- Where is it processed and stored, and can I choose the region? Residency requirements apply to the inference call as well as the database.
- How long are prompts, outputs and logs retained, and can I set the period? Debug logs often hold the most sensitive text in the system.
What can the system do with it?
- What access does the system hold in my environment, and is it scoped per user? An assistant that reads everything the service account can read will eventually show someone something they shouldn't see.
- How do you handle prompt injection from documents and web content? A credible answer describes input handling, output checks and limits on what tools can do — not a claim that the model is 'secure'.
- Can the system take actions, and which need human approval? Request the list of tools it can call and the approval rules for each.
How will I know what happened?
- Do you keep an audit log of inputs, tool calls and outputs that I can export? You will need it the first time someone asks why the system said what it said.
- What is your incident process for AI-specific failures, such as a data exposure through a model or a harmful output? Ask for the notification timeline in writing.
- What happens to my data and my configuration if I leave? Deletion, export format and timelines should be specific, not 'on request'.
Vendors who have thought about these questions answer them quickly and with specifics. Vague reassurance, or answers that arrive weeks later, are themselves the answer.
Use the answers
Score each answer as clear, partial or missing, and attach the high-risk gaps to the contract as conditions, not footnotes. A short review done before signing costs a few hours; the same issues found after launch cost a relationship.