Start with authority, not personality
Define the job in terms of allowed actions, required information, and decisions that belong to a person. A polished conversation is not evidence that a system should be allowed to issue a refund, change a price, or commit the business to an exception.
Keep state and history visible
Operational work lasts longer than one message. The system should preserve the current task state, the sources it used, the actions it took, and the reason it stopped or escalated. That record gives a person enough context to review the next step.
NIST's AI Risk Management Framework treats governance, documentation, and defined human oversight roles as continuing parts of responsible AI use—not a one-time launch checklist.
- State what the system may read and change
- Set monetary and policy limits
- Require approval for exceptions and high-impact actions
- Record actions, sources, and reviewer decisions
- Provide a safe stop when required information is missing
Test real exceptions
Before expanding authority, test ambiguous requests, missing information, conflicting policies, unusual amounts, and attempts to bypass the rules. The important result is not that every task completes; it is that the system knows when it should not continue alone.