D&S-07: Twelve Questions to Ask Any AI Vendor Before You Sign

Part 7 of the Data & Security series. In Part 6 we handled the human side. Here is the vendor side: one checklist for the meeting before the contract.

Every AI vendor meeting sounds the same. The demo is fast, the slides say “enterprise-grade security” twelve times, and by minute forty everybody is nodding. Then the contract lands, and the security turns out to be three vague sentences.

These twelve questions are how you find out which meeting you are actually in. Ask them before you sign, ask them in writing, and notice not just the answers but the temperature. A serious vendor answers calmly, because they get asked these every week. Under each question is what a good answer sounds like.

The twelve questions

1. Do you train your models on our data? Good answer: “No,” followed by the clause in the agreement that says so. Anything that starts with “we take privacy very seriously” is not an answer.

2. How long do you keep our prompts, files and outputs? You want a number, and ideally a setting your admin controls. “As long as necessary” is a number too. It is just not yours.

3. Can we delete our data completely, and how would we verify it? Deletion on request, a defined timeline, and clarity about backups. If deletion is a support ticket into the void, note it.

4. Where is our data physically stored, and can we choose the region? Country-level answers at minimum. Regulated work may need the data to stay in the US, and “the cloud” is not a place.

5. Who are your subprocessors, and do you notify us when the list changes? Every vendor has suppliers behind them. The list should exist, be current, and come with change notices in the contract.

6. Who at your company can access our data, and is that access logged? Support access happens. The questions are whether it is limited, approved, recorded, and reviewable by you after the fact.

7. Which independent audits do you hold, and can we see the current report? SOC 2 Type II is the common one, ISO 27001 beside it. Asking to see it under NDA is normal. Being refused is information.

8. How fast do you notify us of a breach that involves our data? In hours, in the contract. This single clause has saved more client relationships than any encryption ever invented.

9. Will you sign a data processing agreement? And a BAA, if health data is involved? Standard vendors have these ready. If legal has to invent the document from scratch, you are their first serious customer, and you should price that role accordingly.

10. What audit logs do we get on our side? You need your own record of who used the tool and what it touched, exportable, kept as long as your industry requires. Their logs are theirs. Yours must be yours.

11. What happens to our data when we cancel? Export first, then deletion, on a timeline, in writing. Leaving a vendor should not mean leaving your data behind as a souvenir.

12. When your terms change, how much notice do we get, and can we leave? Terms change in this industry like weather. Notice periods and an exit right turn a surprise into a decision.

How to use the list

You do not need twelve perfect answers. You need twelve real ones, weighed against what the tool will touch. A vendor handling your public marketing drafts can miss on residency. A vendor touching client files cannot miss on questions 1, 8 and 11 at any price.

And keep the answers. The file you build in this meeting is the same file your regulator, your insurer, or your biggest client will ask for later, usually on the worst possible day. Having it is the difference between a company that uses AI and a company that can prove it uses AI carefully.

The series closes next with the whole thing on one page: a ten-point safety check you can run on your own company in an afternoon.