Agent to Agent Workflows

Voice agent

Phone is still the highest-volume channel in most service organisations and the one least touched by automation — because the things that make speech hard are exactly the things that make a service call a service call: interruption, background noise, accents, and part numbers read out fast.

  • A spoken conversation, not a phone tree
  • Handles interruption instead of restarting
  • Built for technical vocabulary and part numbers
  • Transfers to a person with the context intact

Watch the full tour

Verify your work email once and every research paper, case study and product tour on ascendo.ai opens, right here, without leaving the page.

  • Takes about 30 seconds
  • Work email only, no spam

Chat deflection never touched the phone queue

Most support organisations have spent several years improving chat and self-service, and still have a phone queue that looks much as it did before. The customers who call tend to be the ones with the most urgent problems.

Speech is harder in the specific ways service is hard. People interrupt, correct themselves mid-sentence, and read a part number with the letters run together over the noise of the machine beside them. And a phone tree that asks for the problem in one word then routes wrong is worse than waiting.

Three cards: years of chat and self-service improvement leaving the phone queue unchanged, the specific ways service speech is hard, and a phone tree that routes on a one-word guess.
The callers left in that queue are the ones with the most urgent problems, which is the worst place to have made no progress.

What a spoken conversation has to survive

None of these are edge cases on a service call. They are the normal conditions.

  1. Step 1

    Handle interruption

    People talk over the system, change their mind halfway through a sentence, and correct a detail they already gave. A conversation that restarts every time is not a conversation.

  2. Step 2

    Get the technical detail right

    Service calls are full of model numbers, part codes and error strings, spoken quickly and in noise. Everything downstream is built on them.

  3. Step 3

    Reason on the same knowledge

    Voice is a surface, not a separate product. What it can answer comes from the same manuals and resolved cases as every other channel.

  4. Step 4

    Hand over cleanly

    When it reaches its limit the caller goes to a person who already has the transcript and what has been established.

Four stages a spoken service conversation has to survive: handling interruption and self-correction, capturing model and part numbers accurately, reasoning on the same knowledge as other channels, and transferring cleanly.
Interruption and a part code read over machine noise both arrive before the system has answered anything.

Three things that decide whether callers tolerate it

Customers form a judgement about a voice system in roughly four seconds.

A conversation, not a phone tree

A tree makes the caller translate their problem into categories somebody else defined; a conversation lets them describe it and does the translating itself. People can tell which one they are in almost immediately.

Built for how service calls actually sound

Technical vocabulary, part numbers, background noise and accents are the normal case here. A voice system that only performs in quiet, general-purpose speech does not survive a plant floor.

The handover decides the trust

Every deployment eventually meets a call it cannot handle, and how that moment goes is what the customer remembers about the whole system.

A call transcript showing a caller interrupting and correcting themselves, a part number read over machine noise and confirmed back, and a transfer where the agent already holds what was established.
The caller corrects themselves mid-sentence and reads a part code over a running machine, and neither restarts the conversation.

Why the usual approaches fall short

The phone queue has resisted every one of these.

An IVR phone tree

Asks the caller to classify a problem they have not diagnosed, then routes on that guess.

General-purpose voice AI

Handles everyday speech well and has never encountered your part numbering.

Chat deflection alone

Helps the customers willing to type. The phone queue stays exactly where it was.

More phone agents

Works, scales linearly with cost, and does nothing about the calls that never needed a person.

Two columns comparing an IVR phone tree, general-purpose voice AI, chat deflection alone and hiring more phone agents with a voice agent built for service calls.
General-purpose voice AI handles everyday speech well and has never encountered your part numbering, which is the part of the call that matters.

What things are called

The terms that come up when a voice deployment is being evaluated.

Voice agent
A spoken conversational surface over the same knowledge every other channel uses.
Barge-in
The caller speaking over the system. Handling it is most of what makes a conversation feel like one.
Warm transfer
Passing a call to a person along with the transcript and what has already been established.
Containment
The share of calls resolved without a person — only meaningful if the caller actually got their answer.
Omnichannel
The channels you interact with customers through: email, phone, Slack or Teams, website, training portal, customer portal and knowledge portal.

Frequently Asked Questions

It transfers to a person, with the transcript and what has already been established attached. That handover is the part worth examining in any evaluation, because every voice deployment reaches it eventually and it is what the customer will remember.

Support Channels

Test it on your real calls

Voice claims are easy to make in a quiet room. Send us recordings from your actual queue, with the accents and the part numbers, and we will show you how it handles them.

Talk to us