Specify separately what happens to future access, retained copies, indexes, backups, derived datasets and trained models. Ending an account is not the same as removing every effect of prior use.
AWS explicitly notes that previously exported files remain accessible after a Files subscription ends; deletion requirements must be checked in the Data Subscription Agreement. AWS — Product subscriptions in AWS Data Exchange NIST defines machine unlearning as removing selected training influence, including approximate techniques; do not turn that concept into an unconditional promise for the buyer's model. NIST CSRC — Machine unlearning
Request an exit schedule before signature: named system owners, deadlines, backup treatment, permitted legal retention, subcontractor obligations and evidence of completion. Agree what can be tested and what depends on attestation. If an already-trained model may remain in use, make the continuing permission and payment treatment explicit.
The IAB's December 2025 transaction playbook highlights the need to align termination drafting with technical realities. IAB — AI Intellectual Property and Transactions Digital Advertising Playbook, December 2025 Where a promised remedy is central to your willingness to license, have both counsel and an independent technical reviewer assess it before release. A contractual remedy may remain important even when selective technical deletion is difficult.