Ask where data goes
Confirm whether prompts, files, outputs, tool traces, and feedback are stored, used for training, retained for abuse monitoring, or processed in specific regions.
For agent tools, ask about connected systems too: repositories, browsers, cloud accounts, ticket systems, document stores, and messaging platforms.
Review contractual and operational terms
Check service levels, support channels, model deprecation policy, pricing change policy, audit logs, admin controls, indemnity, and incident notification timelines.
A model that performs well but lacks predictable enterprise operations can create hidden risk for production teams.
Plan exit and migration
Document how prompts, evaluations, logs, embeddings, files, and workflows can move if the provider changes pricing or deprecates a model.
Avoid building irreversible assumptions around a single model name. Prefer interfaces that allow routing and replacement.
Practical checklist
- 1Review data retention and training use.
- 2Check regional and access-control requirements.
- 3Confirm support and incident terms.
- 4Document model deprecation exposure.
- 5Create an exit plan before rollout.
Related comparisons