News: Avesta Labs is exhibiting at Retail Show Australia 2026, 22–24 September at the MCEC in Melbourne. Stand 489. Meet us there →

Talent Carriage

Case study · Voice AI · HR shared services · India

The voice agent that writes the ticket while you talk.

Talent Carriage runs HR for more than 50 companies. When an employee at one of them has a problem, they ring the support line. We built the voice co-worker that answers that line at any hour, in Hindi, English or both at once, works out what is actually wrong, and has the ticket written before the caller hangs up.

Client
Talent Carriage
Channels
Phone and WhatsApp
Languages
Hindi, English, Hinglish
Status
Live since August 2026
Book an AI KickoffHear a real call2 min

Change the industry. The job is the same.

  • Payroll bureaus
  • Managed IT services
  • Facilities management
  • Member associations
  • Franchise support desks

If your service desk only has a record of the queries someone remembered to write down, then your reporting is not a picture of demand. It is a picture of who had a quiet afternoon.

The queries did not vanish. They were never written down.

Before

6steps

Each week, one person at Talent Carriage carried support duty on top of their own job. On a quiet day, a few calls. On a busy day, the whole day. And every call that was answered still had to be typed into a list afterwards.

  1. 1Answer the phone, if you are at your desk
  2. 2Understand the problem
  3. 3Finish the call
  4. 4Remember to open the list
  5. 5Type the query in
  6. 6Work out whose job it is

Four ways a query died: the FAQ answered in a minute and never logged, the call taken mid-task and logged later, the missed call meant to get a call back, and the after-hours call nobody heard.

After

2steps

The line is answered on the first ring, at any hour. The query is written down while the employee is still on the phone, and it arrives with the team that owns it.

  1. 1Ring the same number as always
  2. 2Say what is wrong
  • Finish the call
  • Remember to open the list
  • Type the query in
  • Work out whose job it is

The record is a by-product of the conversation instead of a task that comes after it.

Talent Carriage did not arrive with a failed project or a shortlist of vendors. They arrived with a sentence: every query logged, every time.

The ask was smaller than the problem

They asked for a way to capture queries. What they needed was for capture to stop being a separate act of remembering, which is a different thing to build.

We mapped their day before we designed anything

Discovery sessions on how support actually ran, then a user story map: every step of the existing process, next to what the new one would do instead.

They could see it before they bought it

The map meant Talent Carriage recognised their own problem in the plan, and knew what they were getting, before a line of code existed.

We split it in two, and only sold the first half

Phase one captures every query and routes it. Phase two resolves them from each client's policies. Phase one had to earn phase two.

Nobody at Talent Carriage was careless. There was simply more arriving each day than one person on duty could hold.

What was at stake

Talent Carriage does not sell software to its clients. It sells the service around it, and the support line is where that service is either delivered or not. An employee whose query was never written down does not shrug and forget. They ring again, and then again, and then they tell their own HR team that the provider is not answering. That conversation happens inside the client company, where Talent Carriage cannot hear it, and it happens shortly before a renewal.

So every query had to be:

  • Captured, not remembered

    Written down during the conversation, by the thing having the conversation, so no step depends on someone's memory at 6pm.

  • Understood before it is logged

    A ticket that says 'portal issue' is a second conversation waiting to happen. It has to arrive complete enough to act on.

  • Sent to the right team first time

    Fifty companies feed eight specialist teams. A query in the wrong queue is a query that waits while somebody forwards it.

  • Escapable

    Anyone who does not want to talk to an AI must be able to reach a person, and the query is logged on the way out either way.

How we worked

  1. 01Before any build

    Sat in on how support actually ran

    Discovery sessions with the people who carried support duty, on what a real day looked like, not what the process document said it looked like.

  2. 02Then the map

    Drew every step, old and new

    A user story map putting the existing process beside the proposed one, step for step. Talent Carriage signed off on a picture, not a promise.

  3. 03Build, then rebuild

    Threw the first version away

    The first agent followed a fixed order and could not cope with real callers. We rebuilt it as one that decides what to ask next. See below.

  4. 04End of August 2026

    Their own team tested it, then a staged rollout

    Talent Carriage staff tried to break it alongside our own testing. Then it went live for all clients, with Talent Carriage briefing each company separately over the following weeks.

The hard part

Logging a ticket is easy. Getting someone to stay on the line is not.

An employee with a payroll problem has no patience for a machine that mishears them, talks over them, or asks for something they said thirty seconds ago. If the call feels like a form, they hang up and ring a person, and Talent Carriage is back where it started.

What broke

The first version asked questions in a fixed order

It worked out the category, then asked that category's questions, then logged the query. The questions were flexible. The order was not. Real callers open with a detail before they open with a subject, and something that sounds like a leave question turns out to be about payroll. Once it had settled on a category it could not go back, so it asked for things it had already been told, or filed the query under the wrong team.

What we did

We rebuilt it to decide what to ask next

Not a script with branches, but an agent that holds the whole conversation in view and works out the next question the way a trained support person would. It can change its mind about what the query is halfway through, because callers do.

If we had shipped the first version, employees would have asked for a person within a fortnight. The calls would have gone back to the duty roster, and the only people who would have noticed early are the specialist teams reading the logs.

One real call

Two minutes, start to finish.

This is a real call to the support line, played in full and unedited. Nothing is staged, nothing is re-enacted and nothing is cut. Every false start, every pause, and the moment the caller stops to go and check, are all exactly as they happened.

Voice call · logged as one query

0:00 / 2:01

  • It knows who is ringing from the number, and which of the 50 companies they work for.

  • A real opening. False starts, no subject, no system named.

  • It does not guess. It asks, and it offers examples so the question is easy to answer.

  • Then five and a half seconds of nothing while they go and look. This is the hardest thing on the page to build and the easiest to miss.

  • Reads it back, so a misunderstanding surfaces now rather than in the ticket.

  • The caller cuts in over the question. The agent takes the answer and carries on.

  • Still missing one thing it needs, so it asks for that and nothing else.

  • One last confirmation of the whole problem before anything is written down.

  • The ticket exists before the call ends. Nobody has to remember anything.

What the team received

Written by the agent, during the call.

Channel
Voice
Query type
Data Management & Tools
Priority
Medium
Status
Closed

Description

Raj reports that the inbox tabs for new messages, sent items, archive and inbox are not displaying in the HR app when he logs in. The same tabs are visible to other employees. Raj is unable to see any tab content in his own interface.

Reconstructed from the call, in the format the real tickets use. The ticket for this query was never exported, so this is not its verbatim wording.

What the voice agent does now

The phone line was the build. WhatsApp came second and runs on the same understanding underneath, so everything below is true of a call and of a chat.

The numbers so far

Five weeks in, and every figure is measured.

Talent Carriage went live at the end of August 2026 and briefed each client company separately over the weeks that followed, so this is an early and deliberately small picture. It is drawn from their own live dashboard on 10 September 2026.

40 of 41

tickets needed no correction

In 40 of 41 queries, the person who picked the ticket up did not flag the summary or the team the agent chose as wrong. They started work from what the agent wrote.

77 / 23

call it, versus type it

Roughly three in four employees still ring. The rest use WhatsApp, and before this there was no way for them to raise anything in writing at all.

41

queries on record since launch

Not one of them depended on somebody remembering to open a list after the call had ended. That is the entire point of the project, and it is the number that carries it.

There is no reliable before figure to compare any of this to, and that absence is the reason the project exists. The old list was missing precisely the queries it was meant to capture, so nobody knows what the real volume used to be.

Most AI projects are at their best on the day they launch.

Two things push against that here, and both of them are running now rather than planned.

The model was chosen for one specific thing

We tested cheaper and faster options and rejected them. They could not reliably work out the right next question to ask, which is the only job that matters here. A model that asks the wrong follow-up produces a ticket the team has to chase, which is the problem we were hired to remove.

Which model answers is a setting rather than a rebuild, so moving to a better one later does not mean starting again.

  • Every call is read by the person who resolves it

    The recording, the transcript and the agent's summary sit on the ticket. Anyone who finds the summary wrong flags it in one field, and that field is the accuracy figure above. The people best placed to catch a problem are the people already reading every log.

  • Real calls feed the next version

    Employees turned out to talk to it in ways testing never produced. Those calls are where the tuning since launch has come from, rather than from a script we wrote in advance.

What's next

Talent Carriage's first request after launch was the second phase.

The agent logs queries. It does not answer them. Phase two would let it resolve the common ones from each client company's own policies, so the specialist teams see fewer repeat questions. It is not built, so this page claims nothing for it.

Behind the scenes

What it is made of.

The call itself
Plivo carries the phone line, LiveKit carries the live conversation. Each call runs as its own session.
Hearing and speaking
Sarvam turns speech into text and text back into speech, across Hindi, English and the mix of both.
Working out what to ask
Anthropic's Claude holds the conversation and decides the next question.
WhatsApp
Meta's WhatsApp Cloud API, running on the same core as the calls.
One shared core
A ticket raised by phone and a ticket raised by chat are the same object. The team behind it cannot tell which came from where, and does not need to.
Where it runs
Tickets, recordings and transcripts stay in Talent Carriage's own cloud storage and database, in the Central India region.

Start with the line that rings when nobody is free to answer it.

A short workshop, then one co-worker live on a single channel. You finish with something answering real calls, the record of what it heard, and a clear view of whether to take it further.

Book an AI Kickoff

Or email hello@avestalabs.ai