On this page
Voice agent testing, the guide
Where to start when you test an AI phone agent. What to check, how to run the calls, and where each topic is covered in more depth.
Short answer. A voice agent can read well in a transcript and still fail on the phone. Test it the way a caller meets it: ring it, talk over it, hang on a noisy line, and see what it does when something goes wrong. Then keep the evidence.
Start here. Pick the section that fits your problem. Each links to guides with sources.
What to test
- Turn-taking and latency. Does the agent stop when the caller speaks, wait when the caller pauses, and answer quickly enough? See turn-taking and latency.
- Phone audio, accents and noise. Does it still work on a narrow, noisy line with a regional accent? See phone audio, accents and noise.
- Identity checks. Does it let the right caller in and keep the wrong one out? See testing identity checks.
- Your own platform. What changes for your stack. See test guides by platform.
How a test works
- Write the call as a short script with one thing to check.
- Ring the agent with a simulated caller and keep the audio and transcript.
- Score the call against written yes or no rules, with the quote that earns each mark.
- Run it more than once, because the caller and the agent both vary.
- Fix what failed, then run the same calls again.
Why more than once. One passing call proves little. Report a rate across runs, and say how many runs sit behind it.
Read next
The handful of things that decide whether a caller gets what they rang for, each with a test.
When typed tests are enough and which faults only show up in audio.
What a good handover to a person looks like, and how agents get stuck in loops.
How a timed trace of each call shows where a slow or wrong reply came from.
The two ways to build a voice agent and what each means for testing.
Wrong passes, partial answers, persuasive callers and spelled letters.
Thirty terms you will meet when building or testing a phone voice agent.