AI engineering · HR shared services · India
Talent CarriageThe voice agent that writes the ticket while you talk.
Talent Carriage runs HR for more than 50 companies. Every employee query used to arrive on one support line, and most of it depended on whoever answered remembering to write it down. We built a 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.
At a glance
- Status
- Live since August 2026
- Channels
- Phone and WhatsApp
- Languages
- Hindi, English, and both in one sentence
- Where it runs
- Tickets, recordings and transcripts in Talent Carriage’s own cloud, Central India region
What changed
What changed
Before
A query survived only if someone answered, understood it, remembered it and typed it in.
Now
The employee rings the same number as always and says what is wrong.
The rest happens while they talk.
- Tickets needing no correction, and counting
- 70 of 71
- Client companies on one line
- 50+
- Specialist teams it routes to
- 8
- Callers who choose voice over WhatsApp
- 3 in 4
The ask
The queries did not vanish. They were never written down.
Six steps stood between an employee’s problem and a ticket: answer the phone, if you are at your desk; understand the problem; finish the call; remember to open the list; type the query in; work out whose job it is. Four things broke that chain. A frequently asked question got answered and never logged. A call was taken mid task. A missed call waited for a callback. An after hours call went unheard.
After: ring the same number as always, and say what is wrong.
What was at stake
Captured, not remembered
Written during the conversation, not dependent on someone’s memory at six in the evening.
Understood before it is logged
Complete enough to act on, without a second conversation to find out what was meant.
Sent to the right team first time
Fifty companies feed eight specialist teams. The routing has to be right on the first pass.
Escapable
Anyone who refuses to talk to a machine reaches a person, and the query is logged either way.
The method
How we worked
01
Before any build, we sat in on how support actually ran, in discovery sessions with the people who took the calls.
02
Then the map. Every step, old and new, drawn out in a user story map, so the first release was the thinnest slice that did the whole job.
03
Build, then rebuild. We threw the first version away and rebuilt it as an adaptive agent.
04
End of August 2026: internal testing by Talent Carriage staff, then a staged rollout.
The hard part
Logging a ticket is easy. Getting someone to stay on the line is not. Employees with a problem have no patience for a machine that mishears, talks over them, or asks for something they already said. Poor performance sends callers straight back to people.
What broke
The first version asked questions in a fixed order. It categorised first, then asked the questions for that category. Real callers open with the details before the subject, and conversations change direction. Once it had categorised, it could not reconsider, so it routed wrongly and repeated questions.
What we did
We rebuilt it as an agent that decides what to ask next, holding the whole conversation in view the way a trained support person does. It stops and listens when you cut in, and ignores a cough or a laugh. It knows the difference between a finished sentence and someone gathering their thoughts. It follows Hindi, English, or both inside one sentence.
What we built
What the voice agent does now
- Answers the line at any hour, on the same number
- Runs identically over WhatsApp
- Logs the query and picks the specialist team
- Tells both sides on WhatsApp when a ticket opens and when its status changes
- Chases tickets: escalated at 36 hours, breached at 48
- Hands over to a person on request, and logs the query either way
- Gives unknown numbers a straight answer: contact your own HR team
- Keeps the call with the ticket: recording, transcript and summary attached
What it changed
The number that matters
The measure of this agent is not how many calls it takes. It is whether the person who picks up the ticket has to fix it. Since launch, 70 of the first 71 tickets have needed no correction, and the count is still running: the specialist accepted the agent’s summary and team assignment as written. Roughly three in four callers choose voice; the rest use WhatsApp.
There is no reliable before figure to compare any of this to, and that absence is the reason the project exists. Every query is now on record, which was never true before.
Behind the scenes
- The call itself
- Plivo for the phone line, LiveKit for the live conversation
- Hearing and speaking
- Sarvam for speech to text and text to speech, Hindi, English and mixed
- Working out what to ask
- Anthropic’s Claude, holding the conversation and deciding the next question
- Meta’s WhatsApp Cloud API on the same core
- One shared core
- Phone and chat tickets are identical objects
- What is next
- Talent Carriage’s first request after launch was phase two: the agent resolving common queries from each client’s policies, so the specialist teams stop answering the same question twice. Not yet built.
Built to last
Why it keeps working
Most AI projects are at their best on the day they launch. This one is measured on every call.
- 01
The model is a setting
We tested cheaper and faster options and rejected them, because they could not reliably work out the right next question. Model selection is a setting, not a rebuild, so that choice can change the week a better option lands.
- 02
Every call is read by the person who resolves it
Recording, transcript and summary sit on the ticket. The person flags a summary error in one field, and that field is the accuracy measurement. Real calls have fed tuning since launch.
Beyond this client
The job transfers
The same problem turns up well beyond this one client. These are the teams it fits.
Change the industry. The job is the same.
- Payroll bureaus
- Managed IT services
- Facilities management
Anyone with one line and many callers
- Member associations
- Franchise support desks
- Any shared services desk
Next case study
view.com.au · Product development
Rebuilt in under four months, live to millions the whole time
Facing something similar?