Six Questions to Ask Before You Sign an AI Contract
AI vendors have learned to be very good at demos. Here is what to ask when the demo is finished and the contract is on the table.
A polished demo tells you one thing: the vendor can make their product look good under conditions they control. That is worth knowing, but it is not the decision you are making. You are deciding whether this system will work on your data, inside your operation, at a cost you can predict, for years. These six questions are the ones we ask on behalf of clients, and the answers separate real products from expensive experiments.
1. What data was this demo running on?
If the answer is anything other than “yours”, ask for a pilot on a sample of your actual data before signing anything. Demo datasets are curated to flatter the product. Your invoices, your case notes, your sensor logs are not. A vendor confident in their product will agree readily. A vendor who resists has told you something important.
2. What happens when it's wrong?
Every AI system is wrong some of the time. The question is whether the product was designed by people who accepted that. Ask what the error rate is on data like yours, how errors surface, and what the correction workflow looks like. If the vendor talks as though errors don't happen, the system has no plan for them, and your staff will be the plan.
3. Where does our data go?
Get specific answers, in writing: where data is processed and stored, whether it is used to train the vendor's models, who can access it, and what happens to it when the contract ends. Under UK GDPR you remain accountable for personal data you hand to a processor. “It's all secure” is not an answer. A data processing agreement with named subprocessors is.
4. What does this cost at our real volume?
AI pricing is frequently usage-based, and pilots run at pilot volume. Model the cost at your actual throughput, including the growth you are hoping for, before you sign. We have seen per-document pricing that looked trivial in a demo become the largest line item in a department's budget at production scale. If the vendor cannot help you build that model, build it yourself before signing.
5. How do we leave?
Ask what an exit looks like while you are still friends. Can you export your data, in a usable format, including anything the system generated or learned? What notice is required? What does the transition period look like? Vendors change strategy, get acquired and shut products down. An exit clause you never use costs nothing. The absence of one can cost you the whole dataset.
6. Who is accountable when it matters?
If the system makes a decision that affects a customer, a patient or a regulator, who answers for it? Ask whether the vendor supports auditing, whether decisions can be explained after the fact, and what their own liability position is. Regulated organisations increasingly need to demonstrate oversight of AI they use, not just AI they build. A vendor with nothing to say here is a vendor you will be defending alone.
A vendor confident in their product will answer all six questions readily. A vendor who resists any of them has answered a seventh question you didn't ask.
None of these questions requires technical depth to ask. They require permission to be sceptical in a room where everyone is excited. If it helps, that is a service we provide: an independent technical review of an AI purchase before the signature, from people with no stake in whether the deal closes.
