The most useful question to put to an AI supplier is what happens when the model they depend on is deprecated by its provider. It is specific, it has no marketing answer, and a supplier who has thought about it will tell you their plan in under a minute.
Most AI procurement fails at the same point. The buyer assesses the demonstration, the demonstration is impressive, and nobody assesses the organisation behind it or the chain of providers behind them.
What is AI due diligence?
AI due diligence is the assessment a buyer runs on a supplier's ability to build, operate and withdraw an AI system safely, as distinct from assessing whether the system currently works.
It covers four things: where data goes and who touches it, how the supplier decides a system is fit to release, what happens when the underlying model changes, and who carries the cost when an output is wrong.
Standard software due diligence covers none of them properly. A supplier can hold every certification you normally ask for and still have no answer to what happens when their model provider retires a version with six months' notice.
What should you ask an AI supplier before signing?
Ask about the processing chain, the evidence of testing, the change process and the liability position, in that order. The order matters because each answer determines whether the next question is worth asking.
On the processing chain, establish where inputs are processed, under whose jurisdiction, who else sits in the chain as a sub-processor, whether inputs are retained, and whether anything you send is used to train a model. Suppliers frequently know the answer for their own platform and not for the model provider underneath it, which is the gap worth finding.
On evidence, ask what the system was evaluated against, on whose data, and whether you can see the results. Ask what the failure rate is and how it is measured. A supplier who has never quantified their own error rate has not tested the system, they have demonstrated it.
On change, ask what happens when the model is updated, who re-runs the evaluation, and whether you are told before or after. Model updates change behaviour, and a system signed off in March can behave differently in September without a line of the supplier's code changing.
On liability, ask who carries the cost of a wrong output that reaches a member of the public. Read what the contract says rather than what the salesperson says, because those two are rarely the same document.
We have set the questions out in full, with what a good answer contains and what a weak one sounds like, in a supplier question set you can send to a bidder or drop into an existing assessment.
The answer arriving slowly is itself an answer
Where the service is genuinely well run, most of these come back in a sentence or two, because they are facts about how the thing was built rather than judgements about it.
A supplier who takes a fortnight and comes back with a document has told you something useful. They are constructing the position now. That is not disqualifying, particularly for a small supplier moving fast, but it changes what you should require in the contract and how much you should rely on the system without your own checks.
Score the specificity, not the sentiment. "We take data security extremely seriously" is not an answer to where the data is processed. "Inputs go to our application in London, then to the model provider's European region, and we hold no copy after 30 days" is.
What should an AI contract cover that a software contract does not?
Five clauses, none of which appear in a standard software agreement.
A named model and a change process, so that switching the underlying model is a contractual event with notice rather than a silent update. A retraining and retention position, stating in terms that your data is not used to train the supplier's or the provider's models, if that is what you have been told verbally.
An evaluation commitment, requiring the supplier to re-test against an agreed set after any material model change, and to share the result. A withdrawal trigger, defining the evidence that requires the system to be taken out of service and who can make that call, because the absence of this is why systems known to be producing wrong answers stay live.
And an output liability position that does not simply assign every consequence to you. Read the indemnity and the limitation of liability together rather than separately. In most AI agreements as drafted, a wrong output is your problem financially and your problem publicly, because the organisation running the service is the one a regulator writes to. Our guide to what councils need in place before using AI sets out why that accountability stays with the organisation and cannot be delegated.
Exit deserves a mention of its own. Establish what you get back at the end, in what format, and whether any fine-tuning, prompt engineering or evaluation work you have paid for leaves with you or stays with the supplier.
How do you assess AI in your supply chain?
Start with the suppliers who are already using AI in work they deliver to you, not with the ones selling you an AI product. That is where the exposure sits and it is almost never assessed.
The pattern is now documented at scale. Deloitte carried out an independent assurance review of the Targeted Compliance Framework for Australia's Department of Employment and Workplace Relations between December 2024 and June 2025. ABC News reported on 9 October 2025 that the firm would issue a partial refund to the government after using AI in a report that contained mistakes. The department's published resource page records that the report was updated on 3 February 2026 to address identified corrections, replacing the version published on 3 October 2025.
The relevant detail for a buyer is that this was not an AI product procurement. It was a professional services contract, and the AI arrived inside a deliverable without being part of what was bought.
Three questions cover it. Does any part of what you deliver to us involve generative AI. If so, where, and what checks the output before it reaches us. And will you tell us in writing when that changes.
Put them in the contract renewal rather than in a one-off survey. A survey gets answered once by whoever opens the email.
Where certification helps, and where it stops
Certification to ISO 42001 means an auditor has been through the method. That is worth having and it is not evidence about the particular system you are buying. Our guide to what an ISO 42001 certificate actually covers sets out what the standard certifies, how UK accreditation stands after January 2026, and why making it a mandatory requirement today produces a very short bidder list.
The practical position for the next year or two is to treat certification as one input, ask every bidder the process questions directly, and weight the quality of the answers more heavily than the presence of a logo.
We hold ISO 27001, ISO 27017, ISO 27018 and ISO 20000-1, certified through BSI, plus ISO 9001, ISO 14001 and Cyber Essentials Plus. We do not hold ISO 42001. Ask us the questions in this article and hold our answers to the same standard.
Before your next AI supplier conversation
Take the questions above into the next supplier call you have booked and ask four of them. Write down which ones produced a fact and which produced a position. The ratio is the assessment.
If a supplier's answers have left you unsure, send them over. We will tell you which ones were facts and which ones were positions, and the assessment is yours to use however you like.
Frequently asked questions
Should we run AI due diligence on suppliers who are not selling us AI? Yes, and that is where most of the unassessed exposure is. Professional services, research, translation, content and development suppliers are all using generative AI inside deliverables, usually without being asked and sometimes without a policy of their own. Ask at contract renewal and require notification when the position changes.
Can we rely on a supplier's certifications instead of asking these questions? No. A management system certificate confirms an audited process for governing AI. It makes no statement about the specific system you are buying, the model underneath it, or what the contract says about liability. Certification narrows the questions. It does not remove them.
What is the single best question to ask an AI supplier? What happens when the model you depend on is deprecated. It reveals whether the supplier has a dependency they have thought about, whether they can switch, what it would cost you, and how much of their service is their own work rather than a thin layer over somebody else's model.
