The AI vendor procurement checklist — twenty-four points.

Pricing, metering and the unit economics.

1. Pricing unit defined. Tokens? Calls? Requests per second? Compute hours? Output rows? Each vendor uses different units; the contract must name the unit and its measurement methodology.
2. Rate card transparency. Per-unit rates documented for every meter, including overage. No 'market rate' or 'then-current pricing' language without a cap.
3. Commit / overage split. Pre-paid capacity at one rate, overage at another. The overage rate matters more than the commit rate.
4. Carry-forward of unused capacity. Pre-paid tokens / calls should roll over at least quarterly. Annual-only carry-forward is the negotiated baseline.
5. Price escalators capped. Annual escalation capped at CPI or a fixed percentage, not 'vendor's then-current pricing'.
6. Discount tied to commitment, not term. Multi-year discounts that compound rather than reset at renewal.

Who owns the data — going in and coming out.

1. Input ownership. Customer retains all rights, title and interest in input data — no licence-back to the vendor for any purpose.
2. Output ownership. Outputs are owned by the customer, with vendor disclaiming any licence to outputs.
3. No training on customer data. Explicit prohibition on vendor using customer input or output to train, fine-tune or improve any model. This is the single highest-impact clause and must be hard-coded — not relegated to a privacy policy that can change.
4. Data residency. Where data is processed, where it is stored, and how that is enforced. EU buyers should require EU-only processing with named regions.
5. Logging and retention. What the vendor logs (prompts, responses, metadata), how long they retain it, and customer rights to delete on demand.
6. Sub-processor transparency. Every model provider, hosting partner and processor named — and a notification right on new sub-processors.

IP, indemnity and output ownership.

1. IP indemnity for output. Vendor indemnifies customer against IP claims arising from the output the model produces — uncapped or with a cap tied to the contract value.
2. IP indemnity for input handling. Vendor indemnifies for any breach of customer input rights.
3. Output attribution rules. Where the vendor requires watermarking, provenance markers, or attribution in outputs, the customer's commercial use rights are explicit.
4. Right to remove model from production. Vendor must give notice of model deprecation (90+ days) and provide a comparable replacement.

Versioning, deprecation and roadmap commitments.

1. Model versioning commitments. Customer can pin to a specific model version for the contract term; vendor cannot silently substitute.
2. Deprecation notice. Minimum 90 days' notice on any model removal, with documented behavioural equivalence on the replacement.
3. Performance benchmarks. Latency, accuracy or task-specific metrics committed in the contract.
4. Quality regression protection. Right to exit without penalty if a substituted model fails defined behavioural tests.

What protects the buyer after signature.

1. SOC 2 Type II + ISO 27001 in force. Annual proof, with right to review.
2. Right to audit. Customer audit right on data handling and security controls, at least annually.
3. Incident notification timeline. 48–72 hours on confirmed security incidents.
4. Exit and data return. All customer data returned in usable format within 30 days of termination; vendor data destroyed and certified.
5. Transition assistance. 90 days of vendor support during transition to a successor provider, at contract rates.