What to Settle Before Signing an AI Vendor Contract
An AI vendor contract risk checklist should connect data use, provider changes, evidence, incidents and exit to decisions before a buyer signs the agreement.

The AI demonstration works. The business sponsor wants to move. Procurement has a price, and the vendor's standard agreement is ready for signature. The remaining questions are described as details for legal to finish.
Some of those details determine what the company is actually buying. Can the provider use submitted information to improve a model? Can it change the underlying service? Will the buyer receive enough notice to respond to an incident or move elsewhere?
An AI vendor contract risk checklist makes those questions visible before the commitment. It is a preparation tool for business owners, procurement, security, privacy and counsel, not a substitute for legal advice. The aim is to connect each important promise to an owner, evidence and an operational response.
Key takeaways
Define the permitted use, data and business outcome before negotiating general assurances.
Read the order form, product terms and incorporated policies together with counsel.
Ask how provider, model and processing changes reach the people who can act on them.
Match evidence and incident obligations to the buyer's actual use and response needs.
Test exit and portability assumptions before treating termination rights as a working fallback.
Step 1: Fix the use case before reviewing the promises
Describe what the product will do, who will use it and what it can influence. Internal drafting, customer advice and decisions affecting eligibility create different questions. Identify the data involved and whether the AI can take actions or only propose outputs.
Then identify the service and contracting entity. A favorable statement about one product tier may not cover the product being purchased. Ask counsel to trace the agreement through its attachments and incorporated terms, and identify how conflicts or later changes are handled.
Record the business owner who will decide whether the service remains appropriate. A contract file without an operational owner becomes difficult to use when the product changes.
Step 2: Resolve data use and downstream responsibilities
Ask what happens to prompts, uploaded files, outputs, usage records and feedback. Distinguish processing needed to deliver the service from training, product improvement or other secondary use. Clarify access, retention, deletion and the treatment of backups or logs where relevant.
Identify the parties behind the service and the information each receives. A vendor may provide a subprocessor list, but that alone may not answer model-routing or other dependency questions. Link the agreement to the dependency record rather than maintaining two conflicting versions of the service.
The NCSC's secure AI development guidance recommends assessing and monitoring AI supply chains across the life cycle. The practical contract question is what evidence, notice and options let the buyer manage that dependency after signature, not only during procurement.
Step 3: Make change and incident terms usable
List the changes that could alter the approved use: a new model provider, a different processing location, a material feature change or retirement of an interface. Ask what notice the buyer receives, when it arrives and which choices remain available.
A notice is useful only if someone owns it. Put the receiving address, review owner and escalation route into the vendor record. If the business cannot tolerate a change without testing, discuss how the agreement and operating process support that need.
For incidents, ask counsel and the response team to assess notification, cooperation, evidence preservation and relevant reporting obligations. Avoid a generic number of hours chosen without the facts. The agreement must work with the buyer's duties and response process, and applicable requirements vary.
Step 4: Connect evidence rights to real questions
Request evidence that covers the service, period and controls relevant to the use case. A certification badge is a starting point, not the full answer. Determine what reports can be reviewed, under what access terms and how unresolved findings will be handled.
Ask how the buyer can validate important claims when a report is unavailable or excludes the relevant function. Options may include a scoped discussion, additional documentation or another assurance route. Do not assume an unlimited audit right will be commercially available or sufficient on its own.
Where personal data is involved, counsel should determine the parties' roles and applicable requirements. The ICO's AI contract guidance discusses controller and processor responsibilities and written agreements under UK data protection law. That is a jurisdiction-specific source, not a universal rule for every AI purchase.
Step 5: Test the exit before relying on it
Termination ends a legal relationship. It does not automatically restore the business process. Ask which data and configurations can be exported, in what format, at what cost and within what period. Check whether the receiving system can use them.
Consider the outputs and working assets the business needs to retain, rights to use them, deletion confirmation and any continuing obligations. Counsel should review these questions alongside commercial liability, intellectual property, remedies and other terms for the actual transaction.
An alternative also needs capacity, approved data handling and working integrations. A contractual exit can still leave a service interruption if no one has prepared the destination.
A pre-signature decision record
This is a practical review worksheet, not proposed contract language. Replace the illustrative questions with the facts of the transaction.
| Area | Question to settle | Owner and evidence |
|---|---|---|
| Approved use | Which actions, users and data are in scope? | Business owner; use-case record and product description. |
| Data and providers | Which uses, recipients and retention practices are permitted? | Privacy and counsel; applicable terms and architecture evidence. |
| Change | What changes require notice or reassessment, and what can the buyer do? | Product owner and counsel; notice route and contractual options. |
| Incidents and assurance | What information and cooperation can the buyer obtain? | Security and counsel; response needs and scoped assurance evidence. |
| Exit | Can the business move or continue without the service? | Operations and procurement; export test and transition plan. |
An unresolved row should lead to a condition, a narrower use, further negotiation or an authorized decision about the remaining exposure. Do not bury it in an email and call the review complete.
Sources
NCSC, Guidelines for secure AI system development: secure development, supply-chain and asset guidance (2023). ICO, Contracts and third parties, UK-specific AI data protection guidance. The checklist is a practical synthesis, not a claim that these sources prescribe every commercial term.
Frequently asked questions
Is this checklist legal advice?
No. It organizes the business and risk questions that counsel should assess in the actual agreement, jurisdiction and use case. It does not supply clauses or determine whether a contract complies with law.
Is a promise not to train on customer data enough?
It answers one question only if its scope and meaning are clear. The buyer still needs to understand processing, access, retention, downstream recipients and other uses relevant to its information and workflow.
Should every AI vendor accept the same contract terms?
The review should reflect the service and its consequences. A low-impact drafting aid and an AI service embedded in a critical process may need different evidence and protections. Any remaining exposure still needs an owner and the appropriate decision authority.
What should happen if a material question remains unresolved?
Make the gap explicit. Seek evidence, negotiate, limit the use or defer the commitment. An authorized decision-maker may accept an exposure within their authority; urgency does not create that authority.
Closing
Before signing, ask the service owner to explain what the company can use, what can change and how it would leave. If those answers depend on an assumption no one has verified, that is the next item to settle, not a detail for later.
Get Board-Ready on AI and Cyber Risk
A 30-minute call. No charge. You'll leave with a clear read on where your oversight stands and what to do about it.


