Build an AI Vendor Offboarding Plan Before You Sign the Contract

An AI vendor offboarding plan is a tested path for leaving without losing data, prompts, agent logic, evaluation history, permissions, or business continuity. Build it during procurement, not after a price increase or incident. At minimum, name an owner, inventory assets and integrations, define export formats, set deletion evidence, rotate credentials, preserve audit history, and rehearse a fallback workflow.

List source files, vector indexes, embeddings, conversation logs, prompts, system instructions, fine-tuning data, model adapters, agents, tool schemas, evaluations, safety rules, analytics, user feedback, and generated assets. Record who owns each item and whether it can be exported in a documented, usable format. A pile of JSON is not portable if no replacement can reconstruct the behavior.

Keep canonical prompts, policies, test sets, tool contracts, and workflow diagrams in a controlled repository outside the vendor. Document vendor-specific features and the acceptable degraded mode if they disappear. For critical agents, maintain a manual procedure or secondary provider for the small set of tasks that must continue during transition.

Define export assistance, formats, timing, fees, access during disputes, retention after termination, deletion of backups, subprocessor obligations, model-training exclusions, administrator logs, and written deletion confirmation. Include how the vendor handles derived data such as embeddings, fine-tuned weights, safety labels, and usage analytics—not only uploaded files.

Freeze new writes, export and validate data, transfer ownership, disable automations, rotate API keys and OAuth grants, remove service accounts, revoke user sessions, update DNS or webhooks, and monitor for residual calls. Keep an auditable checklist with two-person approval for production credentials.

Quarterly, export a representative project into a clean environment. Confirm counts, hashes, permissions, links, timestamps, metadata, and the ability to use the result. Delete a canary record and obtain evidence of how it moves through live systems, logs, backups, and subprocessors. Record the time and labor required.

Score ownership, export completeness, format usability, replacement effort, deletion evidence, credential revocation, continuity, audit retention, contract clarity, and tested recovery from one to five. Any critical category below three is a procurement risk with an owner and deadline.

Portability is a capability, not a clause. A credible AI exit plan proves that the organization can retrieve what it owns, reconstruct essential behavior, stop every connection, verify deletion, and keep serving users when a vendor relationship ends.