The refund example is exactly where I'd start on the support side too. If it helps anyone setting this up for a non-engineering team, here's the checklist I'd use for the first rollout:
- Write the refund rules as numbers, not prose: max amount per order, max per customer per 30 days, eligible reasons, and a list of order states that are never refundable.
- Make the agent propose, and let a plain rules check approve or reject. Anything over the cap goes to a named human queue, with the ticket, order and proposed amount attached.
- Treat the ticket text as untrusted. If it says "ignore the policy", log it and flag the ticket rather than arguing.
- Keep a weekly review of overrides and rejections. They show you which rules are too tight and which are being probed.
That gives support leads something they can read and sign off without touching code.