AI vendor due diligence for a UK SME: the twenty questions, and the answers that should end the conversation.
By Zain Masroor · 16 October 2026 · 11 min read
Before a UK SME signs for an AI product or a supplier’s build, twenty questions settle whether it is safe to proceed. On data: are we the controller and you a processor under a DPA, is our content used for training, where is it stored and processed, who are the sub-processors, what is retained and for how long. On the model: which provider and model, what happens when it changes, and can we see the prompts. On accuracy: what is the measured accuracy on our kind of input, how is it tested, what is the exceptions path. On security: least privilege, prompt injection controls, attestations, incident notice. On cost: the running cost per unit, the ceilings, what happens at ten times the volume. On lock-in: who owns the code, prompts, data and accounts, and how we leave. Each has an answer that is fine and an answer that ends the conversation, listed below.
Talk it through: book a free 30-minute call →Data: five questions
The five that decide whether a product can hold your data at all. Ask them first; the rest of the conversation depends on the answers.
| Question | Acceptable | Ends the conversation |
|---|---|---|
| Are we the controller and you a processor under a written DPA? | Yes, DPA supplied, ICO-aligned | No DPA, or “our terms cover it” |
| Is our content used to train your or anyone’s models? | No, by default, in writing | Yes, or opt-out only, or unclear |
| Where is our data stored and where is inference run? | Stated per region, with the transfer mechanism | “In the cloud” |
| Who are your sub-processors, and will you tell us when they change? | Published list, change notice | Will not say |
| What is retained, for how long, and can we delete it? | Stated retention, deletion on request, zero-retention option where relevant | Indefinite, or cannot delete |
The model: three questions
Which provider and which model does the product run on, and does that change without notice? A vendor that will not say is hiding either a cheap model or a dependence it does not control. What happens to accuracy and cost when the provider retires a model, and who tests the replacement? And can we see the system prompts and the rules the product applies to our data, or at least the categories of them? “Proprietary” is an acceptable answer for a large vendor’s internals; it is not an acceptable answer to whether your client’s data is sent to a third model provider you have not recorded.
Accuracy: three questions
What is the measured accuracy on inputs like ours, and can we test it on fifty of our own before signing? A vendor confident in its product agrees to the test; one that offers a demo on its own examples instead is telling you something. How is it monitored after go-live, and what does drift look like? And what is the exceptions path: when the product is unsure or wrong, does it stop and ask a person, or does it guess? A product without an exceptions path is a product that guesses.
Security: four questions
What access does the product need to our systems, and is it the least that does the job? What are your controls against prompt injection, given the NCSC’s position that it cannot be fully prevented? Which attestations do you hold (SOC 2, ISO 27001, ISO 42001, Cyber Essentials) and will you share the reports? And how quickly will you tell us of an incident affecting our data? A vendor that asks for full admin access to your mailbox or CRM, has not heard of prompt injection, and has no attestations is a vendor to walk away from, however good the demo.
Cost: two questions
What is the running cost per unit, per conversation, per document, per call, and what is included in the subscription before overage begins? And what does the bill look like at ten times our volume, with what ceilings? Credits-based pricing, which several platforms have moved to, needs a worked example at your volume, because the headline seat price and the effective price can differ by a large multiple. A vendor that cannot give a per-unit figure has not measured its own product.
Lock-in and the contract: three questions
For a built system: who owns the code, the prompts, the data and the accounts, and are they in our name from day one? For a product: how do we export everything, in what format, and what happens to our data after we leave? And what is the notice period, the price-change clause and the minimum term? The acceptable answers are ownership in your name, a documented export, and a term you could actually leave at. Our guide on your data when a supplier builds your software sets out the clauses for the build case in full.
What to keep on file
The twenty answers, the DPA, the sub-processor list, the attestations, the accuracy test results on your fifty examples, and the worked cost example, in one folder with the contract. That folder is most of what a client’s security questionnaire, an ISO 42001 assessor or an ICO enquiry will ask for, and having it is the difference between an afternoon and a fortnight.
Common questions
What should I ask an AI vendor before signing?
Whether you are the controller and they a processor under a DPA; whether your data trains models; where it is stored and processed; the sub-processors; retention and deletion; the model used; measured accuracy on your inputs; the exceptions path; access needed; prompt-injection controls; attestations; per-unit running cost; ownership and export.
Which answer should stop me buying an AI product?
Training on your data by default, no data processing agreement, will not name the model or sub-processors, no accuracy test on your own examples, full admin access requested, no per-unit cost, or code and accounts not in your name.
Should I test an AI product on my own data before buying?
Yes, on fifty real examples including the messy ones, against an accuracy threshold you set. A vendor that refuses the test is answering the question.
What security certifications should an AI vendor have?
SOC 2 and ISO 27001 for security; ISO 42001 for AI management; Cyber Essentials for UK suppliers. Ask for the reports, not the logos.
How do I avoid lock-in with an AI supplier?
Ownership of code, prompts, data and accounts in your name from day one for a build; documented export and a leavable term for a product; a runbook so another supplier can take over.
Does this apply to ChatGPT and Claude business plans?
Yes, and both answer most of the twenty questions on their published trust pages: processor under a DPA, no training, published sub-processors, attestations. Record the answers rather than assuming them.
Want the twenty questions answered for a system we build?
Every proposal we write answers them in advance: processor position, model, measured threshold, least privilege, per-unit running cost, and ownership in your name.