The refund agent
Designing an AI agent that can act, safely

The situation
A retailer wants an AI agent that handles customer refunds. Agents can now act: refund customers, change prices, approve invoices. So the real question is not whether it works in a demo. It is who said yes when it gets something wrong at 2am.
The design
- Its own identity. Every refund carries the name of the customer service colleague the agent is working for.
- Clear permissions. It can read orders and policies freely. It can issue refunds up to £50 on its own.
- A human above the line. Anything above £50, it drafts, and a person approves with one click.
- Hard limits. It can never touch prices, and it can never refund the same order twice.
- A flight recorder. Every decision is recorded, with the policy it relied on.
- A kill switch. There is a daily cap on total refunds. If the cap is hit, the agent pauses itself and alerts a manager.
- A named owner. The head of customer service owns it. Their name is on it.
The lesson
Nothing in that design is exotic technology. It is ordinary architecture discipline, applied to a new kind of worker. That is why architects need to be in the room before agents go live, not after the first incident.
Your AI Agents Can Now Act. Nobody Is Ready.
The episode this case study comes from.
Watch the episode

