We tested a phone agent with rounds of real calls. Here is what the calls found.
A voice agent fails in ways a text bot never does: silence, delay and a caller who hangs up.
Maria answers the phone for Arctic Works, a car-services business we are related to. She books appointments in Romanian and switches to English if the caller does. We tested her in iterative rounds of calls. We did not build her.
What was tested
An inbound phone agent for a small car-services business. She can check open slots, book, look up and cancel an appointment, send a WhatsApp confirmation and alert the staff. She speaks Romanian, switches to English and politely declines any other language.
We are the independent party here. The business is one we are related to, but we tested the agent and we did not build it.
How we tested it
We called, listened and tuned. Each round of test calls checked three things:
- How she answers. Wording, length, tone, whether she asks one thing at a time.
- How fast she answers. From the moment the caller hears the phone pick up.
- What she has access to. The knowledge base, the schedule data and the tools.
Then the next round of calls. This was iterative testing, not a formal campaign, so we do not quote a pass rate.
What the calls found
- A slow start. Opening a call looked up the schedule five times separately. One batched lookup cut the time to the first answer from about two seconds to about 0.4.
- Silence while she worked. During a tool call the caller heard nothing, which on a phone sounds like a dropped line. The fix was a rule never to go silent and a short spoken line before each tool call.
- Talking too much. Long answers and too many slots offered at once. The rules became one question per reply, fewer slots and brevity.
- Saying the wrong thing. She told callers "colleagues are busy" during business hours, and did not check the hours before offering slots.
- Ghost calls. Calls shorter than three seconds raised staff notifications. They are filtered out.
The integration surprises
The voice platform's webhooks behaved differently from what we expected: a signature in a different format, a nested payload, the caller's number buried in nested fields and booking parameters read from the wrong place. Each was caught by a call that did not behave and fixed in the integration. The code went through four rounds of review, and one pass over the agent's instructions fixed 20 issues.
What to take to your own team
- Test by listening. Transcripts do not capture silence or delay.
- Test speed as a feature. A caller decides in a second whether to trust the line.
- Test what the agent knows and can reach, not only what it says.
- Use an independent tester. The people who wrote the prompt are the least likely to hear what is wrong with it.
Putting a voice agent in front of your customers?
We call it, listen, time it and check what it can reach, in rounds. The first month is a trial.
