Before the Ink Dries: Ten Contractual Tests for UK AI Buyers in 2026
Photo: Foreign, Commonwealth & Development Office, CC BY 2.0, via Wikimedia Commons
An AI demonstration is designed to show what might be possible. A contract must establish what will be true when the system is live, audited and paid for. For UK leaders in 2026, the gap between those two moments is where risk accumulates. The ten questions below are intended for procurement teams, legal counsel, information security and finance. Each should be answered with evidence, not assurance.
Begin with the data map, not the feature list
1. Where will our data be stored, processed and supported?
Data residency is not the same as data sovereignty. Ask for the locations of primary processing, backup, disaster recovery and remote support. Identify every subprocessor and the international transfer mechanism relied upon, including the UK International Data Transfer Addendum where relevant. A vendor that cannot produce a current data map should not progress to contract review.
2. What is the lawful basis and role allocation under UK GDPR?
Establish whether the vendor is a controller, processor or joint controller for each processing activity. The contract should include a data processing agreement, clear instructions, breach notification timelines and cooperation with data protection impact assessments. If the tool makes or materially informs automated decisions about individuals, ask how Article 22 rights and human review are handled. The Information Commissioner's Office expects accountability to be demonstrable, not implied.
3. Will our inputs, outputs or usage data train the vendor's models?
This is not a minor setting. Ask whether prompts, documents, embeddings, feedback and evaluation data can be used for model training, fine-tuning or product improvement. Confirm the default position, the opt-out mechanism and whether opting out affects service quality. Seek written confirmation that deleted data is removed from training pipelines and backups within defined periods.
4. Who owns the inputs, outputs and derived artefacts?
The organisation should retain ownership of its inputs and, where appropriate, its outputs. Check the licence granted to the vendor, restrictions on use, third-party claims and indemnities. If the vendor supplies pre-trained models or embedded components, establish what happens if those components are withdrawn or changed.
Interrogate the commercial model
5. Which costs sit outside the headline price?
AI pricing often combines seats, tokens, API calls, storage, premium support, implementation and integration. Ask for a three-year total cost model with realistic usage bands, overage rates, price review clauses and foreign exchange assumptions. Hidden usage costs are easier to challenge before signature than after the first invoice.
6. What happens when accuracy or service levels deteriorate?
The contract should define service levels, evaluation methods, revalidation triggers and remedies. Model drift is a commercial risk, not merely a technical curiosity. Ask for service credits, escalation routes and termination rights if performance remains below an agreed threshold. A vendor that offers only best endeavours has not offered a service commitment.
7. Which security controls are contractual and testable?
Certifications such as ISO 27001 and Cyber Essentials Plus are useful signals, but they are not substitutes for contractual controls. Seek penetration test summaries, incident notification periods, audit rights, encryption standards, access controls and supply chain assurance. For AI systems, add testing for prompt injection, data leakage and unsafe output handling.
Treat exit as a design requirement
8. Can we extract our configuration, prompts, embeddings and evaluation history?
Portability should be specified before implementation. Ask for export formats, transition assistance, retention periods and deletion certificates. If the organisation has invested in prompt libraries, evaluation datasets or workflow configuration, those assets should be retrievable rather than trapped in the vendor's environment.
9. How will the vendor support regulatory change and audit?
UK GDPR obligations will continue to evolve, and sector regulators will expect evidence of oversight. The contract should require notification of material model changes, subprocessor changes and security incidents. It should also oblige the vendor to cooperate with audits and regulatory enquiries within realistic timeframes.
10. What are the termination, transition and deletion terms?
Exit clauses should cover notice periods, wind-down support, data return, deletion verification and continuity if the vendor fails. Check liability caps, indemnities and any minimum term that could outlast the business case. A contract that is difficult to leave is difficult to govern.
From checklist to decision
Procurement teams should score each answer red, amber or green and attach supporting evidence. No pilot should move to production until questions 1, 3, 7 and 10 have written answers accepted by legal, security and the business owner. The purpose is not to slow innovation. It is to ensure that the system the organisation approves in 2026 is the same system it can audit, afford and exit in 2027.
