Skip to content
revot
revot

Voice AI4 min readBy Revot

Before a voice agent answers your phone: what to validate

A voice agent is a chain of services pretending to be one conversation. Here is what to test — latency, interruptions, consent, language, fallback, handoff and cost — before it speaks to a real caller.

In short

  • Measure latency across the whole conversational turn, stage by stage, on a real phone line.
  • Test interruptions, consent and recording notices, and each language your callers use.
  • Decide in advance what happens when a service is slow or down, and when a person takes over.
  • Know your cost per minute at realistic volume before you scale.

A voice agent sounds like one thing: a voice on the line that answers questions. Underneath, it is a chain. Telephony carries the call. A speech-to-text service turns audio into words. A language model decides what to say, sometimes calling tools along the way. A text-to-speech service turns the reply back into audio. Every link in that chain adds delay, can fail, and costs money per minute.

That is why a voice agent that works in a demo can still disappoint on a real phone line. The demo tests whether the chain can work. Production tests whether it works for your callers, on your lines, at your volume, when something upstream is having a bad day. These are the things we validate before a voice agent answers a real call.

1. Latency across the whole turn

The number that matters to a caller is the gap between when they stop speaking and when they hear a reply. It is not the speed of any single service. That gap includes detecting that the caller has finished, transcribing the final words, the model’s reasoning and any tool calls, synthesising speech, and getting the first audio back down the line.

So measure the turn, and measure each stage inside it. If you only know the total, you cannot tell whether the fix is a faster model, streaming the reply sooner, a shorter prompt or a different speech provider. And measure it on real telephony, not a browser test, because the phone network adds its own delay.

2. Interruptions

People interrupt. They correct themselves, answer before the question is finished, or say “sorry, no — the other one.” An agent that keeps talking over a caller feels broken, however good its answers are. Test barge-in deliberately: interrupt mid-sentence, interrupt with a correction, interrupt with background noise. Check that the agent stops, listens, and does not lose track of what was already agreed.

If calls are recorded or transcribed, callers usually need to be told, and the rules vary by place and sector. Decide what the agent says at the start of a call, what is stored, for how long, and who can access it. Then check the implementation matches: that recordings and transcripts land where you expect, are retained as long as you intend, and can be deleted when they should be.

4. Language performance

A speech pipeline that is excellent in one language can be noticeably weaker in another — in transcription accuracy, in how natural the synthesised voice sounds, and in how well the model handles names, numbers and local phrasing. If your callers use more than one language, or switch mid-call, test each one with realistic speech: accents, background noise, and the specific vocabulary your callers actually use.

5. Fallback when a service is slow or down

Every external service will, at some point, be slow or unavailable. The question is what the caller experiences when that happens. Set timeouts on each stage. Decide what the agent says if the model does not respond in time. Decide when to route the call to a person or a voicemail rather than leaving silence on the line. Then test those paths on purpose, by simulating the failure, rather than discovering them in production.

6. A clean handoff to a person

A voice agent should know its limits: the caller asks for a person, the request is outside its scope, the caller is upset, or the agent is not confident. When it hands over, the person who picks up should not have to start from scratch. Pass the context — who called, what they wanted, what has already been said or done — into the system your team uses. A handoff that loses context turns one call into two.

7. Cost per minute

Telephony, transcription, model usage and speech synthesis are typically each billed by usage. Add them up per minute of conversation, at the volume you expect, including the calls that go badly and run long. Long prompts, verbose replies and unnecessary tool calls all add cost as well as delay. Knowing the cost per minute early tells you which design choices are affordable at scale.

Log everything you would need to debug a bad call

When a caller complains, you need to be able to reconstruct what happened: the transcript, what the model decided and why, which tools it called, how long each stage took, and whether a fallback fired. Structured logs with a correlation ID per call make that possible — within whatever consent and retention rules you have set.

None of this guarantees a perfect call; nothing will. But it moves the failures you will inevitably have from “discovered by a customer” to “found in testing, with a plan for each.” That is the difference between a demo and a system you can put on your phone line.

Next step

Start with the process that costs you the most.

Seven short questions about how the work happens today. No sales deck — the first conversation starts from your answers.