Testing Before You Go Live
This is the part most often skipped — and the one that prevents most disappointment. An agent that “kind of works” after two calls has not been tested.
Guideline: at least ten test calls, every call reason at least once. Ideally by two different people — you know your own prompt too well and unconsciously phrase things to fit it.
Write a test guide
Prepare a list in advance. Two things per test case: the exact sentence you will say, and the check — how you’ll recognize that it went wrong.
| # | What you say | Check |
|---|---|---|
| 1 | ”Hi, I’d like to book a check-up.” | Recognizes call reason 1, asks about treatment/time frame, no more than one question per turn |
| 2 | ”I need to cancel my appointment on Thursday.” | Recognizes call reason 2, verifies the name before cancelling |
| 3 | ”How much is a dental cleaning?” | Gives the stored price in spoken form — invents nothing |
| 4 | ”How long is the wait for an appointment?” (not stored) | Uses the fallback line, doesn’t guess |
| 5 | ”Next Tuesday” (near year end: “on January 3rd”) | Resolves the date correctly, including across the year boundary |
| 6 | Give your name early, then return to the appointment later | Does not ask for the name twice |
| 7 | ”I’d like to speak to someone in person, please.” | Transfers to the correct destination, announces it first |
| 8 | Call outside opening hours | Does not transfer into the void, takes a message |
| 9 | ”Sprechen Sie Deutsch?” | Switches entirely, doesn’t mix |
| 10 | Spell out an email address with a difficult name | Reads it back correctly |
| 11 | Go silent for two seconds mid-sentence | Doesn’t hang up, asks instead |
| 12 | Talk over the agent | Returns cleanly to the flow |
| 13 | ”Are you a human?” | Answers honestly in one sentence, continues |
| 14 | Something entirely off-topic (“What’s the weather like?”) | Steers back politely, no lecture |
When new call reasons, tools, or transfer destinations are added later, extend the guide.
How to test
- Listen to the recording, not just the transcript. Pronunciation errors in numbers and email addresses look correct in a transcript and sound wrong.
- Check the Result Dashboard after every test call. Empty columns are a finding, not a coincidence.
- Check that follow-up arrives (email, webhook, message).
- Change only one thing per round and test again. If you change five things at once you won’t know which one helped — and you’ll miss the one that broke something.
- After every change, repeat two or three cases that previously worked. New rules occasionally break old flows.
Continue to Symptom → Cause → Action to map what you find to a fix.