1. Define what is being supplied
- List models, APIs, agents, knowledge bases, connectors and open-source components.
- State where processing occurs and where data and backups may reside.
- Separate standard product capability, custom development and implementation services.
- Record versions, constraints, supported languages, context limits and intended uses.
- State whether the supplier may replace a model or subprocessor without approval.
2. Specify data handling
- Which prompts, responses, attachments, logs and feedback are retained, and for how long?
- Is customer data used for training, fine-tuning, evaluation or improvement of a shared service?
- How do deletion, export, tenant isolation and proof of destruction work?
- Which data classes are prohibited, and how is accidental transfer prevented?
- Who has privileged access, and how are administrative actions recorded?
3. Require measurable quality
Acceptance should use the customer’s representative cases, not a general model leaderboard. Define quality, completeness, abstention, latency, cost and safe-behaviour metrics for each workflow. Preserve the test-set version, configuration and pass threshold.
The contract should require retesting after a model, prompt, knowledge base or tool change. Material regression must allow the customer to hold an update, roll back or discontinue the affected use case.
4. Examine security and resilience
- Role separation, MFA, service identities and audit trails.
- Encryption, key management, backup and recovery.
- Protection from prompt injection, tool-mediated leakage and excessive agent authority.
- Vulnerability disclosure, remediation targets and incident history.
- A manual path, secondary supplier or other means to continue a critical process.
5. Connect the SLA to business operation
API uptime alone is insufficient. Specify latency, throughput, queue behaviour, change-notice periods, support and recovery. Define an AI incident, investigation ownership and the evidence the supplier must provide.
Pricing must remain auditable: unit of charge, caching, retries, storage, network traffic, tools and minimum commitments. Add limits and notifications before budget is exceeded.
6. Design the exit before entry
- Export prompts, documents, indexes, logs, settings and evaluation results in usable formats.
- Define export time, cost, migration support and deletion of residual copies.
- Secure rights to custom components and the ability to use them with another model.
- Require interface documentation and avoid unjustified dependency on proprietary formats.
Procurement is ready for signature when business accepts the outcome, security accepts the controls, legal accepts responsibility and data terms, IT accepts operations and exit, and the acceptance team receives reproducible evidence.