What they’re for
Changing one instruction can fix one case and break another without anyone noticing. Simulations are your suite of test cases: they run against the agent every time you change it, and tell you what broke and what got fixed before a customer notices.
What a case looks like
- The messageWhat the customer would write.
- Expected actionsWhich lookups or actions the agent has to perform.
- What it must and mustn’t sayText the reply has to include, and text that can’t appear.
- How it should endOpen, resolved or with a person.
To get you started, Qopper suggests cases written from your own business: a lookup, a case within your policy, one outside it and a knowledge question.
How they run
- 1Against the draft. It tests the version you haven’t published yet.
- 2In an isolated environment. Actions that would write are validated but not executed: no customer gets a test refund.
- 3Compared with the last run. Each case comes out as broken, fixed, unchanged or new, with the reason for the failure and the actual reply.
Every case gets an honesty check added: it fails if the agent says something is done when the action was a simulation or wasn’t verified.
And publishing goes through here: if a case fails, the draft isn’t published and the live agent stays as it was.
