Most AI supplier assurance still runs on a vendor’s own security questionnaire, ticked off and filed away. That tells you what the supplier is willing to claim about itself. It does not tell you what happens to your data once it leaves your building, who else can see it, or what you’re on the hook for if the model gets something badly wrong. Proper due diligence has to go further than a marketing pack and a signature.
Start With Where The Data Actually Goes
Every AI tool takes an input and does something with it. The question due diligence needs to answer is simple: where does that input travel, and who or what can read it along the way? Some suppliers process everything on their own servers. Others quietly pass requests through third-party model providers, meaning your data crosses another company’s infrastructure before you ever see the output. Ask for a data flow diagram, not a paragraph of reassurance. If a supplier can’t produce one, that’s the answer.
Check Who Trains On Your Data
This is the point most procurement processes miss entirely. A supplier can be perfectly secure and still use your prompts, documents, or customer records to improve their model for other customers. Settling that in writing before anyone signs anything is a contractual matter, not a security one. Ask directly whether your data is used for training, whether that can be switched off, and whether the answer is the same for every tier of their product, not just the enterprise one.
Ask About The Fourth Party, Not Just The Third
Your supplier almost certainly relies on someone else’s infrastructure. That might be a cloud host, a foundation model provider, or a smaller specialist tool bolted on underneath. Supplier assurance that stops at the company you’re contracting with misses the layer underneath, and that’s usually where the actual risk sits. The NCSC’s supply chain security guidance covers exactly this kind of subcontractor risk in more depth. Ask for a list of subprocessors and what each one can access. If the supplier doesn’t know, they haven’t mapped their own supply chain, which is worth knowing before you rely on them.
Look At What Happens When Something Goes Wrong
Every supplier will tell you their uptime is excellent and their security is tight. Fewer will have a clear answer for what happens when the model produces something harmful, when there’s a breach, or when the service goes down entirely and your team can’t work around it. Ask for their incident notification timeline in writing, not as a verbal promise. Ask what counts as a reportable incident from their side, because their definition and yours might not match. This is also where you find out whether they’ve actually thought about failure, or just about sales.
Get Specific On Contract Terms, Not Just Policy Documents
A supplier’s public trust page is marketing. The contract is the only document that matters when something goes wrong. Check that liability, data ownership, and termination terms are spelled out plainly, not gestured at. If the contract says data will be “handled in accordance with best practice” without defining what that practice is, that’s a gap, not a reassurance. Good supplier assurance means reading the actual clauses, ideally with someone who knows what a vague clause looks like before it becomes a live problem. There’s a fuller breakdown of the specific questions procurement teams tend to skip in this piece on what procurement needs to ask AI suppliers.
Build A Standard Process Instead Of Reinventing It Each Time
None of this works as a one-off exercise done for a single big purchase. New AI tools get adopted constantly, often by teams who never go near procurement at all. Without a consistent process, every supplier gets a different level of scrutiny depending on who happened to ask the questions that week. A structured checklist, applied the same way every time, turns this from a judgement call into a repeatable habit.
If you want a practical starting point, work through the free AI supplier review checklist before your next vendor conversation.
