How we work, and what we will tell you before you spend anything.
This page describes a real engagement end to end — what happens in each phase, what you get, what we need from you, and the situations where you should hire someone else.
- 01
It starts with a call
Thirty minutes. You describe what you are trying to do and what you have already tried. We tell you whether it is worth pursuing. A fair number of these end with us saying not yet, and that is a legitimate outcome.
- 02
Assessment before commitment
Two to three weeks looking at what actually happens on your phone lines today — volume, languages, the shape of the conversations, and where calls currently fall through. It ends with a written recommendation and a cost model, not a proposal.
- Call volume and concurrency, measured rather than estimated.
- Which conversations are automatable and which are not.
- Platform and telephony shortlist, with the reasoning shown.
- Cost per minute at your volume, including the parts vendors do not advertise.
- 03
Design before build
The decisions that are expensive to reverse get made here: how the agent escalates, what it does when it cannot help, which languages it handles natively, and how the whole pipeline fits inside a sub-second response budget.
Latency budget · one turn
- Telephony in
- ~80 ms
- Speech-to-text
- ~120 ms
- LLM first token
- ~350 ms
- Text-to-speech
- ~140 ms
- Telephony out
- ~80 ms
- Total
- ~770 ms
Past roughly a second, callers assume the line has dropped and start talking over the agent. Every design decision is made against this budget.
- 04
Built against your real calls
We tune on your actual traffic, not a curated test set. That is the difference between an agent that demos well and one that survives a Monday morning. You see progress weekly, and you can stop at any checkpoint.
- 05
Handed over, properly
You get the source code, an operational runbook, and a working session with whoever will own it. No black box, no dependency on us. If you want us to keep running it, that is a separate decision you make afterwards rather than a condition of starting.
What we need from you
- Access to real call recordings or transcripts — or a week to collect them.
- One person who knows how calls are genuinely handled today, not how the process is documented.
- Whoever administers your phone system.
- A decision-maker who can call the go or no-go at the end of the assessment.
When we are not the right fit
Four situations where we would turn the work down, or tell you to go elsewhere.
You want the lowest cost per minute.
Platform vendors will beat us on price, and they should — they are selling a different thing. If per-minute cost is the deciding factor, buy the platform directly.
You need a chatbot, not a phone agent.
Text and voice look similar and are not. Voice has a latency budget, no undo, and no scrollback. We only work on the voice problem.
You want someone to own it indefinitely.
We build for handover. Retainers exist, but they are optional and they are not what we optimise the build for.
Your call volume is very low.
Below a few dozen calls a day, the engineering cost will not pay back. We will tell you that on the first call rather than after the assessment.
The practice
CentraTalk is a voice AI practice. We work on one problem: making conversational voice agents hold up outside a demo environment.
- Delivery experience
- Fifteen years implementing enterprise cloud platforms — quote-to-cash, sales operations, and business process automation for large, distributed organisations. Rules engines, approval workflows, multi-country rollouts, and integration into established CRM and ERP estates. The work ran end to end: requirements with the business, architecture, build, user acceptance, and go-live support. Voice agents are the same problem wearing a different interface — business logic that has to survive contact with how people actually work.
- Technical depth
- Realtime speech pipelines, telephony integration, escalation design, and per-minute cost modelling for multilingual call traffic.
- We build, not just advise
- We have designed, built, and run voice agents that take real calls — not just advised on them. Most of what we know about latency budgets, barge-in handling, and escalation design came from operating systems in production rather than from reading about them.
Thirty minutes. You will talk to the person doing the work.