Ten questions every UK organisation should ask before signing an AI vendor contract, plus five red flags that deserve a renegotiated multiple.
Every UK board meeting in 2026 has the same slide: "we need an AI strategy." Every vendor demo has the same script: "our AI-powered platform delivers transformative outcomes." Every procurement team has the same problem: how do you tell the serious vendors from the ones that bolted a prompt onto an existing product last quarter?
This guide is a buyer's checklist for UK organisations evaluating AI vendors. It is deliberately boring, in the way that useful things usually are.
What "AI-powered" actually means, in 2026
Three very different architectures are all currently sold under the "AI-powered" banner. Knowing which one the vendor is pitching you is the single most important question in the entire conversation.
The first is a classical machine-learning model trained on the vendor's domain data: risk scoring, fraud detection, recommendation engines. These have existed for a decade; what changed in 2024 and 2025 is that they are now a routine offering. These are generally mature, measurable, and worth considering on a straightforward feature-functionality basis.
The second is a large language model, used behind the scenes to generate content, summarise documents, extract structured data, or power a chat interface. These systems are less mature than they appear. They are also non-deterministic, which means the output you see in the demo is not the output you will always see in production. The question to ask is not "does it work?" but "what happens when it does not?"
The third is a thin wrapper over a public API — OpenAI, Anthropic, Google — with the vendor's product adding a small amount of orchestration and a user interface. This is not inherently bad, but it is important to know, because the vendor's margins, data flows, and regulatory exposure all depend on it.
Ten questions to ask in the first meeting
- 1Which of the three architectures above powers the AI in your product? If the sales engineer cannot answer in a single sentence, you are being sold the second category dressed as the first.
- 2Where does our data go? Specifically: does our data cross into the model provider's infrastructure? Is it used for training? For how long is it retained? What is the contractual position?
- 3How do you measure quality? Is there an evaluation suite? How often is it run? Can we see the last three results?
- 4What happens when the model is wrong? Who is liable? How is a wrong output detected, flagged, and corrected? What is the user experience of a hallucination?
- 5What is the cost model? Per seat? Per token? Per inference? What governs a surprise bill?
- 6Can we bring our own model? If yes, you have optionality and leverage. If no, you are permanently locked into whatever this vendor's margin structure becomes.
- 7What is the fallback? If the model provider has an outage, is the product unusable or gracefully degraded?
- 8What data from our organisation is required for the product to work? The answer "none, it is pre-trained" is a warning sign — it usually means the product is generic.
- 9What audit trail exists? For a regulated organisation this is non-negotiable. You need to be able to reconstruct who did what, when, with which inputs, and what the model returned. "The model decided" is not an answer an auditor accepts.
- 10Can we speak to three customers of the same size and sector as us, who have been using the product for at least twelve months? A vendor who cannot, or who only offers you their beta customers, is signalling something.
Red flags worth naming
Demo-in-a-box: the vendor shows a scripted, impressive demo but will not let you try a parallel scenario with your own data. A real product performs; a demoware product performs on the demo data.
Unspecified model: the sales deck says "powered by advanced AI." Push until you find out which model, which version, which provider. Any serious vendor will tell you; a vendor who cannot has either not standardised it or is embarrassed by it.
No evaluation story: the vendor cannot show you accuracy, precision, recall, or any equivalent numbers for their system. In 2026 this is increasingly a choice rather than a limitation.
"We don't need to worry about compliance because you own the data": the user owns the data, perhaps, but the model provider often processes it, and that processing is a matter of data protection law regardless of ownership. Vendors who wave compliance away are vendors who have not read the law.
Opaque pricing: AI costs are not stable. If the vendor cannot give you a credible worst-case cost under realistic usage, you are being asked to sign a blank cheque.
How the written contract should look
Three clauses, at minimum, that are often absent or weak in a first draft.
A data processing addendum that is specific to the AI feature, not just the host platform. It should name the model provider, the sub-processors, the data flows, the retention periods, and the audit rights.
An uptime and quality SLA that is conditioned on model behaviour, not just infrastructure. "99.9% uptime of the hosted service" is useful; it is also not the metric that matters when the model starts returning nonsense.
A clear termination and data-return clause. If the service stops working as described, or if a quality regression is unresolved, you need an exit route that does not leave your data in the vendor's warehouse.
When to build, when to buy
A short version of a long question.
Buy, when the problem you are solving is common across many organisations, the vendor has deep domain data, and the integration into your workflow is shallow. Content moderation, email classification, standard document extraction, basic customer-service automation — these are commodity AI purchases now.
Build, when the problem is specific to your organisation's workflow, the data advantage comes from your proprietary data rather than a vendor's generic corpus, and the output needs to integrate deeply into systems you already operate. Custom risk models against your customer base, bespoke internal tooling, workflow automation across several of your own systems — these do not have good off-the-shelf answers and buying one often locks you into generic patterns that stop fitting as you grow.
Combine, when the base capability (a model, a vector store, a retrieval system) is commoditised but the glue that puts it into your workflow is where the value lives. This is where most 2026 enterprise AI projects sit. Buy the model; build the glue.
Closing
The AI market in 2026 is a market for discernment. The technology itself is extraordinary; the vendor landscape is uneven. The buyers who will win are not the ones with the biggest AI budgets — they are the ones who ask the boring questions early, who read the DPA before they read the pitch deck, and who treat non-determinism as a product characteristic rather than a demo caveat.
If you would like independent help structuring a procurement, AANDA Labs runs AI adoption reviews on a fixed fee: four to eight days, a written report, three ranked opportunities, and a vendor shortlist. We are a UK-registered Ltd and we do not take vendor referral commissions — the report is yours, independent, and signed.