Define the permitted use
Distinguish model training, fine-tuning, evaluation, retrieval, redistribution, and human review. Define whether derived datasets, model weights or outputs may be shared. Avoid relying on a phrase such as AI purposes when the parties mean different things.
Term and exclusivity
Write down when access begins, how long the licence lasts and what survives termination. An exclusive licence should state the field, geography, customer group and exceptions. Discuss how exclusivity affects your own operations and existing licences.
Payment and acceptance
Specify the delivery package, objective acceptance tests, inspection period, permitted rejection reasons and cure process. Identify when payment becomes due and whether there are holdbacks, usage reports, audits or renewal payments.
Security, onward sharing and deletion
Name the authorised recipients and allowed processors. State storage protections, incident handling, subcontracting limits and what happens to evaluation copies if negotiations end. Discuss separately whether trained models are subject to any deletion or remediation obligation.
Warranties and liability
Be careful with absolute warranties about every person, document or possible legal claim. Review indemnities, liability caps, insurance and responsibilities for buyer transformations with a qualified lawyer. A checklist is not a substitute for a negotiated contract.