Skip to content
Voxoth

Four things, and then it answers the phone.

Describe the job. Draw the parts that matter. Break it in private. Put it on the line. Most of an afternoon, and none of it needs anybody technical.

01

Say what it should do, in English.

Type “book appointments for my dental clinic, and ask new patients for their date of birth” and you get a working conversation back — with new and existing patients already split onto separate paths, because a question only one of them can answer cannot sit on a step both of them have to pass through.

When you want to change something specific, the conversation is a diagram: steps you can drag, and lines between them for “if they say this, go there”. A step has to collect what it asks for before the call moves on, so “always check whether they are a new patient” is enforced by the conversation rather than hoped for in a prompt.

02

Break it in private first.

Write down what a caller might say and what the agent has to do about it. Every one of those runs through the real conversation — the same steps, the same documents, the same rules — with only the microphone and the speaker stood in for. Change the wording, run them all again, and see what you broke before a customer does.

When you go live with a change, you can start it on one call in a hundred and watch how those calls went before giving it the rest. Putting it back is one number set to zero, and it applies to the next call rather than the next deploy.

03

Put it on a real phone line.

A recorded call

Agent

यह कॉल गुणवत्ता और प्रशिक्षण के लिए रिकॉर्ड की जा रही है।

Spoken before the greeting, as fixed text — not left to the model to remember.

Agent

नमस्ते, मैं अस्योर क्लिनिक से रिया हूँ। मैं आपकी क्या मदद कर सकती हूँ?

Caller

क्या आप इंग्लिश में बोल सकते हैं

Can you speak in English?

Agent

Sure, I can speak in English. How can I help you with Assure Clinic today?

2320 ms · The caller switched language mid-call and the agent followed.

Caller

आई वांटेड टू बुक एन अपॉइंटमेंट फॉर टुमारो

English words, Devanagari script — how Hinglish comes back from transcription.

Agent

I can help you book an appointment. Could you please share your preferred location and phone number?

2066 ms

Inbound call to a Mumbai number, 28 July 2026 · 81 seconds, 19 turns. Verbatim from the stored transcript, with four opening turns of background noise cut and the excerpt stopped before the caller gave their name. Latencies are the ones recorded for these turns.

It answers your number and it dials out from it. The caller can interrupt it and it stops talking, the way a person would. Somebody who rang last week does not have to explain themselves again, and only the agent they spoke to remembers them — your collections agent never reads your support agent’s notes.

The timings on this call are worse than the ones we run now. It was recorded before a fix that took roughly eight hundred milliseconds out of every turn, and we have left it as it was rather than swapping in a better-looking call. Today it answers in about a second and a half. We want it under one and a quarter, and it is not there yet.

When it is out of its depth it rings your colleague, tells them who is waiting and why, and only then moves the caller across. If nobody picks up, the caller is told so in words rather than left holding a silent line.

04

Then read back every call it took.

When somebody tells you the agent got it wrong, you can go and read what it actually said. That is the whole point of keeping all of this.

The recording

Every call, playable against the transcript so you can find the moment.

The transcript

What was actually said, with card and Aadhaar numbers already removed.

A summary

Written after the call ends, so a slow moment costs the caller nothing.

How it went

Whether the caller sounded happy about it, counted across all of them.

What it looked up

Every time it checked a document or called one of your systems.

What it learned

The handful of things worth remembering next time that person rings.

Bring us a call you keep missing.

Describe what usually happens on it and we will build that agent and let you talk to it.