Product Design · April 2026

TelvoxDesigning a Single Command Centre for Teams Running AI Voice Agents

  • TelvoxClient
  • Product DesignDiscipline
  • Whole productScope
  • April 2026Shipped
View live product
Telvox brand mark over the platform's overview dashboard in dark and light themes.

Tl;DR

Telvox puts AI voice agents on the phone. I designed the whole thing.

Agent builder, call auditing, campaigns, billing. 18 screens, one product.

Problem

But.. who runs the agent once it picks up?

Configs in one tool. Call logs in another. Leads in a spreadsheet. Billing somewhere else entirely. Operators were doing the joins in their head.

every arrow here
was a different tab

Mapping the operator journey from agent creation through to invoicing.

My role

Build it, watch it, trust it, bill for it.

End to end — research, IA, every screen, the design system underneath it. All of it was mine.

Overview dashboard: twelve live metrics, weekly consumption, billing panel and open tickets.

twelve metrics, one screen — the answer to “is anything on fire?”

Process

Config is easy. Predicting behaviour isn’t.

So the voice, the prompt and the live status sit together — with a Talk to the Agent button right there.

hear it before you point
a phone number at it

Agent Model panel with assigned voice, system instructions, live status and a Talk to the Agent button.

Trust

You can’t trust what you can’t replay.

Every call is recorded and transcribed. Being able to check is what lets you stop checking.

Call Details modal with audio playback and a turn-by-turn transcript.

play it, read it, done

The decision I’d defend

A prompt is code. So I versioned it.

One careless sentence changes how the agent speaks to every caller. Now every save is a version, and a bad prompt is a five-second undo, not a bad day.

one click back
to yesterday

Instructions tab with an inline prompt editor beside a version history panel offering one-click restore.

Building an agent

An agent is four decisions, so each one got its own surface.

What it knows, which model runs it, how it sounds, and what it can reach. Splitting them meant none of the four had to be understood before the others.

The knowledge base attached to an agent, with documents it can draw on.

documents attach to the agent, so “what it knows” is a place you can look

Model configuration for an agent: provider, language, greeting and overlap limit.
The voice library, filterable by language, gender and provider.

filter by language, gender and provider — picking a voice is browsing, not configuring

Tools management: each endpoint with its method and what it returns.

every endpoint with its method and what it returns

Then it runs

Once it is live, the questions change from how to what happened.

Call logs, conversation totals and outbound lists are the screens people return to daily, so the numbers sit above the table rather than inside it.

An agent's call log with start, end and duration per call.

duration per call on the row — the number people scan for first

Calls and conversations: totals across completed, incoming and outgoing, over the call table.
Outbound leads with connection state and last-updated per number.

connection state per number, so a failed batch is visible without opening it

Running the business on it

Tickets, clients and appointments — the operations side.

Usage against plan is the number that decides renewals, so it reads at a glance rather than living in a billing export.

Support tickets raised against the AI agents, with priority and assignee.
Client organisations: active agents, usage against plan and account manager.

usage against plan on the row, because that is the renewal conversation

The appointments calendar, synced and bookable from the agent.

Impact

And the money has to add up.

Voice minutes cost real money, so revenue, cost, credit limits and risk are first-class screens — not an export.

  • 18screens, designed end to end
  • 2,420calls in the telephony log
  • 7voices, four providers
  • 1place to run all of it
Credit management dashboard with credit limits, outstanding balances, risk scores and account standing.

What this taught me

Designing for AI is mostly designing for the moment it gets something wrong.

Nobody remembered the clever screens. They remembered the transcript, the version history, the overdue flag.