AI vendor termination — the exit rights enterprise buyers under-negotiate.

Why AI termination clauses are different.

Traditional SaaS termination is simple: export your data, delete your account, port to alternative. AI vendor termination is structurally more complex because the vendor holds three things that may have your data embedded in them: the raw data store, the fine-tuned model or indices, and the derivative outputs already generated.

In our experience across 340+ engagements including 50+ AI vendor contracts, the default termination clauses heavily favour the vendor on all three. Data extraction windows are short. Model artifact ownership is ambiguous. Aggregate-data carve-outs allow indefinite vendor use. Buyers who don't negotiate these specifically inherit vendor-favourable defaults that they discover only at termination — when leverage has evaporated. The fix is to settle these terms during the initial AI contract negotiation, while the vendor still wants the signature.

The three things at exit

Raw data: documents, transactions, conversations, training corpora you uploaded. Models: fine-tuned weights, embeddings, RAG indices the vendor built using your data. Outputs: text, classifications, predictions generated during the term, now embedded in your downstream systems. Each requires its own clause and its own exit mechanism.

The lock-in dynamic

AI vendors increasingly compete on lock-in: proprietary embedding formats, vendor-specific RAG architectures, fine-tuning workflows that don't port. Termination clauses that allow only data export — not model export — leave the customer with raw data they cannot use without re-investing in the model adaptation. The lock-in is structural, not contractual; the contract is the only defence.

Seven AI termination clauses every contract needs.

The clause set below produces enterprise-grade exit rights. Each clause has been negotiated successfully in our practice and should be considered baseline for enterprise AI contracts above $250K ARR.

1. Extended data export window. 180 days minimum for raw data export in standard formats. Vendor obligation to maintain platform access during the export window. No vendor right to throttle export performance.
2. Model artifact treatment. Customer choice at termination: (a) export fine-tuned weights, embeddings, and RAG indices in agreed formats, or (b) certified vendor deletion within 30 days with destruction certificate.
3. Aggregate-data carve-out elimination. Vendor right to retain aggregate, anonymised, or model-improvement-derived data limited to the contract term. Carve-out does not survive termination.
4. Output ownership confirmation. Outputs generated during the term are owned by the customer and survive termination without restriction. No vendor right to use customer outputs for marketing, benchmarking, or model improvement post-termination.
5. Derivative-work clarity. Models trained on customer data are derivative works of the customer data. Vendor may not deploy or licence the derivative models to third parties without customer consent.
6. Migration support. Vendor obligation to provide reasonable migration support including export tooling, API documentation, and technical contact for the export window. Optional paid migration services at contractual rates.
7. Audit rights on deletion. Customer right to audit vendor deletion within 90 days of certified deletion. Vendor failure to delete triggers extended liability and contractual remedies.

How major AI vendors approach termination clauses.

Different AI vendor categories produce different default termination postures. Five patterns to anticipate:

• Hyperscaler AI platforms (Azure OpenAI, AWS Bedrock, Google Vertex): Generally cooperative on data export; less cooperative on model artifact export. Aggregate data carve-outs are standard and require explicit negotiation.
• Independent foundation model vendors (OpenAI, Anthropic, Cohere): Variable. Enterprise contracts typically allow data export but model artifacts are vendor-proprietary by default. Negotiation room exists for committed multi-year customers.
• AI application vendors (Glean, Harvey, Hebbia): Data export typically supported. Model artifact treatment is rarely negotiated because the application abstracts the model. Output and derivative clauses matter most here.
• Vertical AI vendors (clinical AI, legal AI, financial AI): Often restrictive on data export because the data is the value. Fine-tuned models are core IP. Termination is more contentious; negotiation begins at signature.
• AI-augmented SaaS (Salesforce Einstein, ServiceNow Now Assist): Termination follows the host platform contract. AI-specific exit terms are usually thinner. Negotiate the AI addendum, not just the host contract.