Plan requirement
| Subscription | Any plan |
| Also required | AI agents - Advanced |
| Access | Admin |
Test the call first, then the conversation around it. Separating those two is what makes a fault findable.
Test it
- Open the action and run it with known values.
- Check what comes back is what you expect.
- Run it with values that should fail.
- Then test it inside a conversation.
Four cases, every time
- A value that works. The straightforward case.
- A value that does not exist. An order number that is not there.
- Nonsense. Letters where a number belongs, an empty answer.
- The system unavailable. Harder to arrange and the most important, because it will happen.
Watch what the customer sees
An action failing is not the problem; an action failing silently is. Run each failure inside a conversation and read what the agent says. A technical error reaching a customer is worse than no answer.
Check what it reveals
Confirm the reply contains only what the customer should see. Actions often return more than is needed, and a field that seemed harmless in the response can be repeated back in the conversation.
Test as somebody without permission
Where the action shows personal data, verify that a customer who cannot be identified gets nothing. This is the test that is skipped and the one that matters most.
Retest after any change
To the action, the procedure using it, or the system at the other end. The last of those is the one nobody remembers, and a field renamed elsewhere breaks an action silently.
Comments
0 comments
Article is closed for comments.