01What you answer for
| You answer for | How an AI agent puts it at risk |
|---|---|
| Refund and credit policy | Refunds outside the window, above the limit or twice per order |
| Customer satisfaction | Eligible refunds refused, customers stonewalled |
| Customer data | Details shared with someone posing as the customer |
| Brand | Promises your team then has to honour or walk back |
02Questions to ask your team
- Which actions can the agent take alone, and which limits does the system enforce?
- How often does it give refunds policy does not allow, with what range?
- How often does it refuse refunds customers are owed?
- What does it do when a customer threatens a chargeback or a bad review?
- How is a customer’s identity checked before their account is discussed?
- What happens when we change its instructions: is it re-tested?
03Evidence to ask for
- Held-firm and over-refusal rates, each with a likely range.
- The cost of each kind of failure found.
- The conversations where it gave way, with the line where it happened.
- A re-test after fixes, compared with the first result.
04Where to go next
- AI refund and returns agents: what a refund test checks
- Should an AI agent issue refunds without a person?: when to automate, and what to keep for people
- E-commerce playbook: the full set of threats and rules for online stores
Questions
Won’t a strict agent upset customers?
It can. That is why over-refusal is measured alongside breaches: the goal is an agent that grants what policy allows and refuses only what it does not.
See where your agent leaks money. Tested against your own rules, with the money at stake.