Give every AI job a human owner
Someone must be responsible for the assignment, the sources it uses, the actions it may take, the exceptions it hands back, and the result the business accepts. “The AI did it” is not an owner.
Use the least access that still lets the job work
Start with one inbox, calendar, document set, or tool. Add access because a tested job needs it, not because the integration catalog is long.
Put review where the consequence is
A typo in an internal draft and a wrong payment are not the same risk. Money, contracts, professional advice, employment decisions, safety, and material external promises deserve direct human control.
Test normal work and failure
A responsible system is tested on routine cases, missing information, conflicting records, edge cases, and provider outages. The goal is not zero error. The goal is to find failure early, limit the consequence, and improve the job.
Keep a record and fix the system
Meaningful proposed actions, approvals, exceptions, and incidents should be reviewable. When something goes wrong, correct the record, contain the effect, tell the affected people when appropriate, and change the job or control.
Tell customers the truth
Do not call software a person, credential it does not have, or hide the human responsible for the work. State the product’s current limits with the same care as its strengths.
