Twelve questions to ask any AI vendor

About the work itself

1. Which of my processes would you start with, and why that one?A vendor who answers before understanding your business is selling a product. The right answer names one process and gives a reason grounded in what you told them.
2. What will exist at the end of the first engagement that did not exist before?"A roadmap" is not an answer. Look for working software touching real data.
3. Show me something you have built that is in daily use.Not a demo environment, not a slide. Something a real person opens on a Monday.
4. What does this need from my team, in hours, in the first month?Every honest answer is larger than you expect. A small number means they have not thought about discovery.

About risk

5. What can the system see, and who decides that?You should be able to get a written list of every source it reads, and change it without calling the vendor.
6. Where does our data physically go?If any part of it reaches a model vendor's servers, that should be stated plainly and named, not described as "the cloud".
7. What happens when it is wrong?The most revealing question on this list. Look for: how it is detected, what it does instead of guessing, whether the action can be undone, and who finds out.
8. What happens if you disappear?Ask who owns the code, the prompts, the retrieval index and any tuned models when the relationship ends. Ambiguity here is lock-in.

About the money and the relationship

9. What is the total first-year cost, including hardware and your time?Build cost, running cost, hardware, and the internal hours. A quote that covers only the first of those is not a budget.
10. How many other clients are you delivering right now?A small firm that says "plenty" is either lying or about to be late. Capacity is a legitimate thing to ask about.
11. Who actually does the work?The person in the meeting, or someone you will never meet. For a small firm the honest answer is usually "me", and that has both an upside and a risk you should weigh.
12. What would make you tell us not to do this?A vendor with no disqualifying answer has never turned down work, which means they will take yours whether or not it should be done.

How to read the answers

Specificity beats confidence. On every one of these, a firm that has shipped something answers with detail and a caveat; a firm that has not answers with enthusiasm. You are not testing whether they know about AI — everyone does now. You are testing whether they have watched their own work survive contact with a real business.

Related

How we work · AI agents for your ERP · Private LLM deployment in India