What to ask before signing any AI vendor contract

The fastest way to waste a year of AI budget is to buy a tool because the demo was good. Demos are designed to be good. The questions below come up in every implementation that goes well, and almost never come up before the contract gets signed in the ones that go badly.

1. Whose data trains your model, and whose stays private?

Ask for this in writing. Some vendors train on customer data by default, with an opt-out. Others never train on customer data. Some intermediate tools route prompts through a third-party API with its own policy. You need to know all three layers — the product, the model provider, and any sub-processors — before you put a sensitive workflow on the platform.

2. What does failure look like, and who notices?

Every AI tool will be wrong sometimes. The question is whether the wrongness is visible. If the vendor cannot tell you how they detect quality regressions in production, your team will be the regression detector. That is fine if you have planned for it. It is expensive if you have not.

3. Can we leave?

Ask how you would export your data, your prompts, your fine-tunes, your conversation history, and your audit logs if you switched providers next year. A vendor with a clean answer is one that has thought about earning the renewal. A vendor that flinches is one whose contract you should read more carefully.

4. What changes when the underlying model changes?

Model providers ship new versions every few months. The vendor sitting on top of them either upgrades silently, lets you pin a version, or quietly degrades when the underlying model behavior drifts. You want to know which one you are buying. "We pin the model and run regressions before upgrading" is a very different posture than "We always use the latest."

5. What is the smallest version of this that we can try?

Any vendor confident in their product will let you pilot on one workflow, one team, or one month before a full commitment. The shape of the pilot tells you almost everything: what they think a win looks like, how they measure success, what their onboarding actually does. If a vendor will not pilot at a price you can absorb if it fails, you are not the customer they want.