1. Data
What data will the system encounter: public material, internal documents, customer conversations or user uploads? Each source carries different authority, retention and risk.
Define what may enter a model, what must remain internal, what requires de-identification and what should not be collected.
2. Access
Not every user should see all knowledge, and not every automated task should carry the same permissions. Visibility, editing, execution and export should be scoped by role, department, project and risk.
3. Human review
AI can draft, summarise, classify and recommend, but the product must define what may be sent automatically and what requires confirmation. External commitments, contracts, policy, health, finance, personal data and high-risk decisions should not default to automatic execution.
4. Cost
Models, voice, storage, search, messaging and external APIs create ongoing cost. Define per-user, per-task and per-period limits, degradation, alerts and provider-switching pathways.
5. Failure and recovery
Production systems must assume wrong answers, timeouts, platform changes and user error. Logs, retries, human takeover, version rollback, backups and kill switches are prerequisites for trusted automation.
Conclusion
An enterprise AI project should not begin with a model name. Begin with what the system is allowed to do, what it must not do, who is responsible and how it recovers when something fails.