1. Limit what it answers from
Give it your approved sources: help articles, macros and a short list of facts. Check that they are current and correct, because the AI repeats whatever it is given. Remove outdated articles first.
2. Decide what it must never answer alone
Write a short list, and test that it hands these over:
- Legal, safety and medical questions.
- Complaints about a person on your team.
- Threats, abuse and anything involving self-harm.
- Disputes over money above a threshold you set.
- Any customer who asks for a person.
3. Limit what it can do, not just what it can say
If the AI can refund, cancel or change an account, set a maximum amount and require a person to approve anything above it. Start with read-only access and add actions one at a time.
4. Make it say when it is unsure
The best behavior is to hand over with a note explaining why, rather than to guess. Test it with questions your articles do not cover, and check that it does not invent an answer.
5. Put a cap on it
A cap on how many answers it can send in a month bounds both the cost and the damage from a mistake. Decide beforehand whether hitting the cap should stop the AI or keep billing.
6. Keep a record and a way to undo
You should be able to see every answer it sent, what it was based on, and mark a wrong one so it does not count. A weekly read of a random ten answers finds problems early.
7. Keep an off switch
One place, usable by an admin in seconds, that turns it off. Know where it is before you need it.