AI voice conversations · Campaigns · Evidence

Every call answers for itself.

Pragya Voice holds a real conversation in English, Hindi or Telugu, and hands back a decision with the customer's own words attached — the verbatim quote, the timestamp, the recording, and the reason the call was permitted to happen at all.

Start the proof week See the one-week path

It runs under your telecom identity, on your carrier account, against numbers you have consent for. We build and operate the platform. You remain the sender of record — which is the only arrangement that is lawful in India, and the one the software enforces.

The room, minute four

Nobody buys a voice agent because it sounds human.

Every vendor demos well. By minute four the demo is over and three questions are left on the table — and they are not about the voice. Pragya Voice was built backwards from these three.

The fear

“What happens the day it calls the wrong person?”

The answer

It can't. Six checks run in the instant before the carrier is touched — not at upload, not at approval, at dial. A number that became ineligible after you approved the campaign is not dialled.

Every refusal is named, logged and countable. You can show a regulator the calls that didn't happen.

The fear

“When Legal asks what it said, what do I show them?”

The answer

The transcript, turn by turn, with the timing. The recording. The disclosure, spoken aloud before recording started. The opt-out, written the moment it was said — during the call, not in a nightly batch.

Every intent label carries a verbatim quote that is checked back against the transcript before you ever see it.

The fear

“How many months until this is actually running?”

The answer

One week from the first conversation to an evidence pack you can take to your board. No CRM integration, no prompt engineering, no IT ticket.

A spreadsheet and one sentence describing what you want to find out. That is the whole input. The regulatory registration needed to dial customers in India runs alongside it, openly, from day one.

The feature nobody else leads with

It says no. Out loud, in writing, six ways.

Most platforms treat compliance as a setting. Here it is a gate that physically precedes the carrier request, and it returns a machine-readable reason every time it closes. This is the part your risk team will want to read.

GateWhat it stopsReason returned
✕ Contact allow-list A number that isn't on the approved list, however it got into the queue. not_allowlisted
✕ Do-not-call Anyone who opted out — on any earlier call, permanently. It outlives the conversation and the campaign. do_not_call
✕ Recorded consent A contact with no consent on record. A number in an uploaded file is never treated as evidence of consent. no_consent
✕ Calling window Anything outside permitted hours in the customer's local time. If the timezone can't be resolved, it refuses rather than guessing. outside_calling_window
✕ Attempts per person Calling the same person more times in a day than you said you would. attempt_cap_exceeded
✕ Daily volume cap A runaway loop. One misread list cannot become a five-figure phone bill or a regulatory incident. daily_cap_exceeded

Approval is a signature, not a checkbox

No call is authorised by a flag the caller sends. A dial is authorised by a server-issued, single-use token bound to one destination — so nothing between the human saying yes and the phone ringing can substitute a different number.

Approving a campaign freezes it

Recipients, objective, caller identity, pitch, voice, language policy, retry rules and schedule are versioned at the moment you approve. Change any of them and it becomes a new version that needs approving again. What you signed off is what dials.

The actual product

The call is the cost. The evidence is the product.

Twenty seconds after hangup, every conversation resolves into the same structured object — no CRM data entry, no listening back, no “the rep said it went well”.

A decision

One of six labels: high, medium or low interest, not interested, call back, or escalate — with a calibrated confidence figure beside it.

The proof

A verbatim quote from the customer that supports the label, matched back against the stored transcript. If it can't be matched, you're told that too.

The next move

A next action, a follow-up date, and a one-line note written for the human who picks it up — what to say, and what not to repeat.

The context

Executive summary, what the customer actually wants in their own framing, sentiment, openings worth acting on, and objections a human should know before dialling back.

Your dashboard reports the number that matters

Not calls made. Cost per high-intent lead, alongside connect rate, engaged rate, and a funnel that separates attempted, connected, engaged, high-intent and escalated. Cost is accumulated from real usage on every call, not estimated.

Watch it happen, or don't

Live transcript as the call runs, a timeline of every event, and the recording afterwards. Campaigns show what's queued, what's blocked and why, what's been attempted and what's been analysed — as separate numbers, because collapsing them into one percentage is how progress starts lying to you.

Built for Indian conversations

It speaks the way your market actually speaks.

Real conversations here switch language mid-sentence. The agent follows the customer; the transcript keeps every word exactly as it was said.

EnglishFull
हिन्दीHindi · Devanagari
తెలుగుTelugu · native script
HinglishCode-mixed
TenglishCode-mixed

Nothing is translated away

Telugu stays in Telugu script. Hindi stays in Devanagari. Words, order, punctuation and casing are stored as they were spoken — because a translated transcript is a paraphrase, and a paraphrase is not evidence.

Your business summary still arrives in English.

Interruptible, like a person

Speak over the agent and it stops inside roughly a third of a second, discards what it was about to say, and remembers only the part you actually heard. That last detail is the one most systems skip, and it's why they repeat themselves after being interrupted.

Both directions

It calls out, and it answers back.

One conversation engine, one evidence record, one set of disclosure and opt-out rules — pointed in both directions. What changes between them is the opening line and which regulations bite. What does not change is that every call still has to answer for itself.

Available now

Outbound — campaigns and single calls

Approval-frozen campaigns over a list you own, single calls placed by a person, and calls triggered from your own tools. Sequential by design: one live conversation at a time, because the alternative is predictive dialling and we are not building that.

This is the path with the regulatory weight on it, which is why every gate on this page exists.

Available now

Inbound — answering the calls you receive

Callers who ring your published number are greeted, qualified and captured in their own words, and hand a human the same evidence object an outbound call produces — the label, the verbatim quote, the recording and the next action.

The opening is written differently, because an inbound caller has already chosen to speak to you: consent and calling windows do not apply. The spoken AI disclosure and the recording notice still do, and still come before anything is recorded.

The two directions carry different obligations, and the platform knows which is which

An outbound commercial call is regulated communication: it needs a registered sender, recorded consent, a permitted calling window and a scrubbed number — so the six gates on this page run before the carrier is touched. An inbound call is a person choosing to reach you. Those gates do not apply, and we do not pretend they do.

What holds in both directions is the part that concerns the person on the line rather than the paperwork: they are told they are speaking to an AI before anything is recorded, and an opt-out is honoured the moment it is spoken.

Onboarding

One week from a spreadsheet to a signed-off decision.

We don't demo. A demo is a recording of somebody else's customer, in somebody else's language, answering somebody else's pitch. We stand the real thing up on your numbers with your pitch, and your own team are the first people it talks to — because in India that is not just the safe order, it is the lawful one.

Day 125 minutes

One conversation, no slides

We ask three things: who you would call, what you need to find out, and what a good answer looks like. That is the configuration. There is no discovery phase, no solution architect, no scoping document.

You bring · one sentence describing the answer you want
Day 1same day, in parallel

The registration clock starts immediately

Dialling customers in India lawfully requires you to be a registered Principal Entity on DLT, with a 140-series calling identity, registered consent records and DND scrubbing through your access provider. It takes weeks, not days.

So it starts on day one, in the open, with us doing the preparation and your counsel doing the approving. Nobody should discover this in month three, and no vendor should quietly route around it.

You start · your own registration, in your own name
Day 2–3build

It stands up on your numbers, and talks to your team first

Your caller identity, your objective, your pitch, your languages. The first conversations go to handsets your own people own and have consented to — the software will not accept a customer list at this stage, and that is a hard limit rather than a policy.

You listen live, read the transcripts as they land, and tell us where the agent is wrong about your business.

You bring · a few colleagues willing to be called
Day 4tuning

Corrected against real transcripts, not guesses

Objections it fumbled, terms it mispronounced, the question it should have asked second instead of fourth. Every change is versioned, and re-approving is a deliberate act rather than a save button.

You give · one round of corrections
Day 560 minutes

You read the evidence pack and decide

Every conversation, its transcript, its recording, its label and the quote behind that label. Turn latency and cost measured on your calls rather than on a benchmark, and the refusal ledger showing what the gates stopped.

If it reads badly, that is a finding about your pitch or your market, and it is worth having either way. We say so rather than reframing it.

You decide · pilot, or walk away owing nothing
Once registeredpilot

Real customers, under your telecom identity

When your registration completes and your counsel signs off, the same configuration points at real prospects — one approved campaign, a real objective, a weekly readout. Until that moment the product refuses Indian customer campaigns outright, and it refuses them in code rather than in a contract clause.

The platform is carrier-agnostic — Exotel for India, Twilio internationally — so this is a swap rather than a rebuild.

You keep · every recording, transcript and decision, exportable
India · who is allowed to call

We will not place calls to your customers on your behalf.

That sentence costs us the easiest demo in this market, and we mean it. Under TRAI's TCCCPR framework a commercial call belongs to a registered Principal Entity, dialled through registered telemarketing resources, against consent recorded in your name. A vendor who offers to skip that by putting your pitch on their own carrier account is handing you an unsolicited-commercial-communication exposure and keeping the convenience.

You are the sender of record

  • The Principal Entity registered on DLT, in your own name
  • The 140-series calling identity the recipient sees
  • The consent — held by you, evidenced by you, revocable to you
  • The carrier account the traffic runs on
  • The controller of your contacts' personal data under the DPDP Act
  • The counsel who signs off before customer dialling begins

We are the platform underneath it

  • The conversation engine, the gates, and the evidence record
  • DND and consent scrubbing executed on your records, before dial
  • The spoken AI disclosure, before recording starts, every call
  • The opt-out, written during the call and honoured permanently
  • A processor acting on your instructions, under contract
  • The refusal ledger you hand a regulator if they ask

The product enforces this rather than promising it

A campaign containing Indian numbers will not run until your readiness is recorded in the system. Until then the only door open is an operator test mode capped at a handful of numbers you own yourself — and it structurally cannot carry a customer list. That gate was built before this page existed, for our own protection as much as yours.

The 2025 amendment to TCCCPR also brings AI-generated voice under artificial-voice and auto-dialler rules, with notification duties attached. That is precisely why the disclosure on this platform is spoken aloud before recording begins and cannot be switched off.

This describes how we structure engagements. It is not legal advice, and it does not replace your own counsel's review — which is a step in the onboarding path above rather than an afterthought.

Design targets

What it is engineered to do.

Stated as targets, because a number that hasn't been measured on your list is a number about somebody else's.

1.2s Turn latency, measured Time from the customer finishing to the agent speaking, on live Indian PSTN. Target p50 ≤ 800 ms.
300ms Interruption to silence And the agent's memory is truncated to what was actually heard.
A fraction Of what the same dial costs a person Carrier, model and analysis are metered end to end on every call and driven to the lowest viable cost per conversation. Quoted on your volumes and languages, not from a rate card.
Zero Calls outside the list The acceptance criterion is zero, not "low". A single one is a failed release.

Every figure above is instrumented rather than assumed. Cost in particular is accumulated from real usage on each call and reconciled against your own provider's billing — where the two disagree, both are shown, because a spend figure you cannot audit is a spend figure you should not trust.

Scope, stated plainly

What it deliberately does not do.

A short list here is worth more than a long feature list above it. Every item below was a decision, not an omission — and if one of them is a requirement for you, the honest answer today is no.

Calling your customers on our carrier account, under our identity, or as your telemarketer of record
Running an Indian customer campaign before your DLT registration is recorded in the system
Predictive, power or parallel dialling — calls run one at a time
More than one live conversation at a time
Warm transfer to a human mid-call
Automated voicemail drops
CRM synchronisation
Payment collection or anything touching card data
Recording before the AI disclosure has been spoken
Treating an uploaded phone number as proof of consent
Suppressing an opt-out after the call instead of during it
Dialling a number your own records say not to
Reporting one blended progress percentage
The offer

One week to a decision you can defend.

You bring one sentence describing what you need to find out, and a few colleagues willing to be called. We stand the agent up under your identity, on your numbers, in your languages — and start your registration the same day, so the path to real customers is open rather than improvised. On day five you read the evidence and decide.

Start the proof week See who is allowed to call
Pragya Voice™ · PragyaMetrics
AI calling is regulated. Deployments are configured to the transparency, consent, calling-window and record-keeping obligations of the destination market — EU AI Act Article 50, TCPA and FCC 24-17, and TRAI TCCCPR for India.
PragyaMetrics supplies and operates the platform. The client is the Principal Entity and sender of record for all commercial communication, holds the consent, and is the data fiduciary for their own contacts under the DPDP Act 2023. We do not act as a registered telemarketer and do not originate commercial calls on a client's behalf. Nothing here is legal advice.
Transcript shown is illustrative. Performance figures are design targets unless marked measured.