AI due diligence should survive the signature
A supplier can pass an AI review today and operate a materially different service six months later. Models, prompts, retrieval sources, safety measures, subprocessors and intended uses can all change while the contract remains in force.
Procurement teams therefore need contract terms that preserve visibility and options after approval. The aim is not to freeze every technical detail. It is to define which changes matter, what evidence remains available and what the buyer can do if the risk position changes.
The seven areas below are a commercial and governance checklist, not model legal wording. Contract language should be tailored with qualified legal advice to the service, jurisdiction and risk.
1. Material AI system and model changes
Define the changes that require advance notice or renewed assessment. Examples may include replacing a foundation model, introducing new training or retention uses for customer data, changing a material subprocessor, removing a safety control or expanding the system into a new decision context.
The clause should set a usable notice period and provide enough information for the buyer to assess impact. A bare statement that the service is being improved does not support meaningful review.
2. AI incident and regulatory-notification duties
Set expectations for notifying the buyer about material AI incidents, harmful or unlawful outputs, security events affecting AI components, significant performance failures and contact from relevant authorities where it affects the service.
Clarify timing, initial information, continuing updates, evidence preservation and cooperation. The contractual promise should match an incident process the supplier can actually operate.
3. Evidence, assurance and audit access
Specify the evidence the supplier must maintain and provide, proportionate to the system. This may include intended-purpose records, risk assessments, testing summaries, transparency measures, human oversight design, incident records, certifications and control-to-evidence mappings.
Use a tiered mechanism where appropriate: routine assurance reports first, targeted evidence for material concerns, and stronger audit or expert-review rights when agreed triggers occur. This can protect legitimate confidentiality while avoiding an assurance right that exists only on paper.
4. Customer data and output-use boundaries
State whether inputs, outputs, prompts, feedback and usage data may be retained or used to train, fine-tune, evaluate or improve models. Address permitted purposes, retention, segregation, deletion and any opt-in or opt-out mechanism.
Do not leave AI-specific uses hidden inside a broad service-improvement term. The buyer should be able to reconcile the contract with its data map, privacy position, confidentiality commitments and customer promises.
5. Model providers and AI subprocessors
Require visibility over material model providers and other AI dependencies, including the process for additions or replacements. The buyer should understand where responsibility sits when the supplier relies on an upstream provider for core behaviour or evidence.
Where objection rights are negotiated, make them operational by defining notice, the information supplied and the available remedy. A list that can change without meaningful notice offers limited control.
6. Performance, monitoring and human oversight
Translate important performance and safety expectations into measurable commitments or governance obligations. Depending on the use, this may include accuracy thresholds, monitoring, testing cadence, known-limitation notices, human review features and support for investigating disputed outputs.
Avoid treating a benchmark as a universal guarantee. The contract and operating process should recognise the intended context, buyer configuration and responsibilities allocated to users or deployers.
7. Suspension, remediation and exit
Define options when a material AI risk cannot be accepted: feature suspension, restricted use, remediation plans, replacement models, data export, transition support and termination where necessary.
Exit planning is especially important when the AI service is embedded in a customer workflow. The buyer needs a realistic way to stop or replace the capability without losing essential records or creating a new operational risk.
Turn the clauses into an operating control
Assign an owner for each post-signature right and record the relevant dates, supplier contacts and evidence location. A notice right has little value when product teams do not recognise a material change or procurement cannot find the agreed route for raising it.
Connect contract management to the AI inventory and supplier register. Model changes should trigger review; incidents should update risk records; new subprocessors should refresh the dependency map; and unresolved evidence gaps should remain visible to accountable owners.
AI Act Ready helps buyers and suppliers turn these commitments into maintained evidence, actions and review cycles rather than leaving them buried in executed contracts.