For merchants, the biggest risks appear when an agent presents information as a promise or takes action itself. A wrong delivery date is already a problem during comparison. Once money, customer data or orders are involved, the systems need clear limits and a traceable record of what happened.
Which four failures should I test first?
| Failure | Control in the shop |
|---|---|
| An old price still appears valid. | Verify price, stock and delivery options against the responsible systems before placing the order. |
| A product description contains instructions for the agent. | Treat catalogue and web text as data. Actions come from approved tools and rules. |
| An order arrives twice after a timeout. | Recognise the same purchase request by a unique identifier and return the existing order status. |
| A user may place orders up to a spending limit. | Check the person or role, the action and the limit in the merchant’s system before executing it. |
How do I set the limits for a pilot?
Start with read access and let the agent prepare a cart. Check whether the data and rules behind its decisions are traceable. Enable actions that write data one at a time, with a clear approval step and a way to hand over to customer service. The pilot should also show how the team can stop an action or recover an order with an unclear status.