Under the Bonnet: The Contract Terms That Let an AI Vendor Swap the Engine You Bought

Few procurement exercises in 2026 have been as awkward as buying artificial intelligence. The evaluation is done on a demo, the pilot is run on a named model, the business case is signed off against a measured output, and then the contract is signed against a schedule that describes the service in terms broad enough to cover almost anything the vendor chooses to run next quarter.

The gap between those two documents is where value quietly leaks. A supplier that reserves the right to substitute the underlying model, withdraw an endpoint or adjust a safety filter is not doing anything unusual — model lifecycles are short, and no vendor can promise a fixed engine for three years without lying. But the buyer who has not priced that flexibility into the deal is, in effect, agreeing to an open-ended variation clause on the one component that determines quality, cost and compliance posture.

The product you tested is not the product you will run

It is worth being precise about what is being purchased. A large language model is not a fixed asset in the way a finance system or a payroll engine is. Vendors retrain, fine-tune, quantise, route between cheaper and more expensive models, and apply safety interventions that can alter tone, truncate answers or refuse categories of request altogether. Some of these changes are announced with fanfare; others arrive as a release note, or not at all.

For a UK buyer, the practical consequences fall into four buckets: output quality against an agreed benchmark, unit economics where pricing is metered by token or by seat, data handling where retention periods or sub-processors shift, and regulatory exposure where the output feeds a decision that an auditor, the Information Commissioner's Office or a sector regulator may later examine.

What actually moves when the engine moves

Procurement teams are comfortable interrogating price and service credits. They are less practised at interrogating behaviour. A model substitution can change the rate at which the system hallucinates on your specific corpus, the length of the answers your staff read, the latency your customers experience, and the volume of tokens consumed to reach the same outcome. A cheaper model that requires three times the prompting is not cheaper.

Data retention is the quieter risk. A vendor that switches inference providers may also switch the logging arrangement behind it, and retention periods that satisfied your records management policy in 2025 may not satisfy it now. Where the tool touches personal data, the change may also amount to a new processing arrangement that ought to be reflected in your Article 30 records and, in some cases, in a fresh assessment under UK GDPR.

The notification window is the whole argument

Most of the leverage sits in a single number: how many days' notice you receive before a material change takes effect. Thirty days is common and usually useless for anything beyond a complaint. Ninety days gives a regulated firm time to run its change process. A twelve-month horizon for the retirement of a pinned model version is better still, and increasingly achievable where the vendor wants the deal.

Define materiality rather than leaving it to the supplier's judgement. A useful definition covers any change that alters benchmark performance beyond an agreed tolerance, changes the price per unit of work, introduces a new sub-processor or retention period, or removes a capability the buyer's use case depends on.

Prove it still works: regression testing rights

A notification you cannot act on is a courtesy, not a control. Ask for a sandbox or evaluation environment in which the successor model can be tested against a fixed set of representative prompts before general release, and for the right to withhold acceptance of the change until that test is passed. Record the baseline at go-live: a documented prompt set, expected outputs, acceptance thresholds and the date. Without a baseline there is nothing to regress against, and disputes become arguments about memory.

Where the tool is embedded in a customer-facing process — claims triage, eligibility checking, service desk deflection — the test should include a human review step and a sign-off from the business owner, not just IT.

Price variation and the small print that shifts with it

Model substitution is frequently a cost story dressed as an improvement story. Check whether the contract permits unilateral changes to metering units, minimum commitments or the definition of a billable transaction. Where the pricing schedule is tied to a named model, state what happens when that model is retired.

Retention deserves its own clause. Require notice of any change to retention periods, training use of customer data, or sub-processor location, and ensure such a change triggers your own review rather than proceeding by default. For firms in regulated sectors, align this with existing outsourcing and operational resilience obligations rather than treating it as a separate exercise.

An exit that exists on paper

Finally, exit. A termination right triggered by a material model change is only useful if it is exercisable within a realistic window and with a defined refund of prepaid fees. Ask for a pro-rata credit, assistance with data extraction, and a transition period long enough to stand up an alternative. Public sector buyers letting contracts under the Procurement Act 2023 framework should map these rights to their published contract terms rather than assuming they are implied.

Clause language worth putting in front of a supplier

The Supplier shall provide not less than ninety (90) days' written notice of any Material Model Change. A Material Model Change includes any substitution of the underlying model, retirement of a named model version, adjustment to safety filters or content moderation policies, change to data retention periods, introduction of a new sub-processor, or change to metering or pricing methodology. The Customer may, within thirty (30) days of such notice, require evaluation of the successor model in a test environment against the agreed baseline; the change shall not take effect until the Customer's acceptance criteria are met. Where acceptance is not achieved, or where the Customer objects on reasonable grounds, the Customer may terminate the affected service and receive a pro-rata refund of prepaid fees together with transition assistance for a period of not less than sixty (60) days.

Governance after signature

Contract terms are the floor, not the ceiling. Nominate an owner for AI change control, keep a register of models in use and their pinned versions, and review the vendor's release notes on a fixed cycle — quarterly is realistic for most organisations. Record every accepted change alongside the evidence that justified it. That register is what you will reach for when a regulator, an auditor or an incoming finance director asks why the system behaves differently from the business case.

The model will change. The question is whether the contract gives you time, evidence and a way out when it does.