5 questions to ask any AI vendor before you sign

Most AI tools are sold on capability and demoed on a good day. The risk shows up later — in where your data goes, what the model was trained on, and who's accountable when it's wrong. You don't need a security team to ask the right questions. You need five.

If a vendor can't answer these clearly and in writing, that's your answer.

1. Where does our data go — and is it used for training?

Ask exactly what happens to the data you put in: where it's stored, who can access it, how long it's retained, and whether your inputs or outputs are used to train the vendor's models. "We take privacy seriously" is not an answer. A data-flow diagram and a retention number are.

2. What's the model behind this, and what changes when it changes?

Is it their own model or a third party's (OpenAI, Anthropic, Google, open-source)? That matters because your risk now includes their vendor too. Ask how you'll be told when the underlying model is swapped or updated — silent changes can break your workflows and your assumptions overnight.

3. How do you handle hallucinations and wrong outputs?

Every AI tool gets things confidently wrong. The question is whether the vendor designed for it: citations, confidence signals, human-in-the-loop steps, and clear limits on what the tool should be trusted to decide. If their answer is "our model is very accurate," they haven't.

4. Who is accountable when it causes harm?

Read the contract, not the website. What does the vendor indemnify? What's explicitly excluded? What are your obligations? Many AI contracts quietly push the entire risk of a bad output onto the customer. Know that before you sign, not after.

5. Can you show me your security and compliance posture?

Ask for the basics: independent audits (SOC 2, ISO 27001), how they handle access and encryption, their incident-response and breach-notification commitments, and any certifications relevant to your industry. Real vendors have this ready. The rest stall.

None of this requires you to be technical. It requires you to make the vendor put their claims in writing — which is exactly where marketing and reality separate.