FIELD GUIDE

AI Contracts and Model Changes: Protect the Right to Re-evaluate

A supplier can change its model, safety filters, subprocessors or data processing terms while the customer’s business process remains unchanged.

A supplier can change its model, safety filters, subprocessors or data processing terms while the customer’s business process remains unchanged. Contracts need a realistic response mechanism.

List change types with consequences

Model version updates can alter outputs, latency or refusal behavior. Retrieval, logging, region and subprocessor changes can alter privacy exposure. Define which changes warrant notice, what will be disclosed and whether customers can pause, test, roll back or renegotiate. Marketing phrases such as always latest cannot substitute for version control.

Specify evidence and access

Require timely changelogs, supported retention and deletion commitments, incident notification channels, audit cooperation and export paths appropriate to the service. Avoid demanding detailed secret model weights if equivalent outcome and assurance evidence is feasible. Separate enterprise contractual commitments from public documentation.

Plan a regression gate

For consequential uses, keep fixed evaluation cases and acceptance thresholds. Test before moving to a newly routed model where the service permits pinning or staging. If automatic upgrades are mandatory, use an isolation layer, health checks and human escalation. An unexplained change must not silently alter high-risk eligibility decisions.

Preserve data processing clarity

The ICO’s UK controller–processor materials explain documented instructions and subprocessor safeguards for covered contracts. But those statutory provisions and remedies should not be described as binding worldwide. Transfer terms, jurisdiction and the actual controller/processor relationship matter.

Example

A customer-support AI changes its underlying model overnight and begins inventing refund exceptions. A contractual change notice plus automated scenario tests flags the regression. The client routes disputed refunds to trained staff and documents financial effects while requiring the vendor’s correction plan.

Sources & Further Reading

CONTINUE EXPLORINGExplore related files ↗