Testing your agent
Three ways to try the agent before real customers do — in the browser, on your phone, or by typing — and what each one does and doesn't cover.
On this page
You should never point a phone number at an agent you haven't spoken to. All three test surfaces are free of setup and reachable from the agent editor header or from Test call.
The three tests
| Test | What it is | Best for |
|---|---|---|
| Web test | Talk to the agent through your browser microphone. | Realism — voice, pacing, interruptions. |
| Call test | The system rings your own phone and connects you to the agent. | Checking how it sounds over an actual phone line. |
| Chat test | Type at it instead of speaking. | Fast knowledge checks and instruction tweaks. |
Web test
The closest thing to a real call without a phone number. Allow microphone access and talk normally.
This is the one to use for tuning behaviour: does it interrupt you, does it wait too long, does the greeting land, does it recover when you change your mind mid-sentence.
Call test
Enter your own number and the system calls you. Everything about the audio path is real — the codec, the latency, the way a mobile handset colours the voice.
Worth doing at least once before going live, because agents that sound great through laptop speakers sometimes sound different down a phone line.
Chat test
A text conversation using exactly the same instructions and the same knowledge base as a real call.
Much faster than voice for checking factual behaviour: fire twenty questions at it and see which ones it can't answer. Every failure is a missing knowledge snippet.
Note
The chat test doesn't wire up booking, transfer or end-call — those are voice-only. Use it for persona and knowledge; use a voice test for anything involving an action.
What to actually test
Most people have one pleasant conversation, conclude it works, and go live. The useful test is the opposite — try to break it.
- 1
The happy path
Be a straightforward customer. Confirm it gets you to the outcome you designed for: a booking, a captured lead, an answered question.
- 2
Ask what it doesn't know
A price you never wrote down. A service you don't offer. Its answer here is the single best indicator of quality — "I'm not certain, let me get someone to call you back" is a pass; inventing something is a fail, and means the guidance in your instructions needs tightening.
- 3
Interrupt it
Talk over it mid-sentence. It should stop and listen. If it ploughs on, adjust the response speed and sensitivity in Voice & behaviour.
- 4
Be difficult
Ask for a manager. Get annoyed. Refuse to give your name. Ask if it's a robot. Every one of these happens in real life within the first week, and each one maps to an instruction section you can fix.
- 5
Go quiet
Say nothing for thirty seconds. It should prompt you, then wrap up politely rather than holding an open line.
- 6
Change your mind
Book a slot, then ask for a different one. Ask for Tuesday, then say actually Wednesday. This is where booking flows usually break.
Reading the result
Test calls appear in your call log marked as tests, so they don't skew your real statistics. Open one and read the transcript — hearing a call and reading it produce noticeably different impressions of what went wrong.
Trial limits
On the Trial plan the test surfaces are your only way to make calls, so they're metered with a lifetime allowance: roughly 20 minutes of browser testing and 10 minutes of phone testing. The remaining budget is shown on the test screen, and a test call is automatically capped so it can't overrun what's left.
Paid plans have no test allowance to worry about — testing bills like any other call.
Careful
Test calls need your email verified, like every other call. If the test button does nothing, check the verification banner first.
Testing after every change
The habit worth building: change one thing, run a chat test, then a voice test if the change affects speech. Changing five things at once and testing at the end tells you something is wrong but not what.