Convin Sense

How Sense works,
and why it looks like this

Twenty minutes, end to end: why the VoiceBot business stalled, what Sense changed, how it is built, what constrains it, and how it is sold today.

SourceFive training sessions
SpanMar 2026
FormatWalkthrough, ~20 min
DetailSession decks, on request

The whole product, in one comparison

60% of a human. Then 80–85%.

Efficiency against a human agent, on the same lead set
Human agent100% The benchmark every deployment is measured against.
VoiceBot alone60% Good enough to sell. Not good enough to renew.
Convin Sense80–85% Good enough that the cost saving survives the top-line hit.

Everything in the next eighteen minutes is an answer to one question: where did those twenty-five points come from?

Session 1 established the gap. Sessions 2 to 5 are how it was closed.02
1 · Why Sense exists

VoiceBot was not a bad product. It was a narrow one.

Three patterns where it reliably paid for itself, found across roughly eight months of pilots.

Lead qualification

50,000 dialled, 25,000 connected, ~5,000 genuinely interested. Human agents only ever speak to the 5,000.

Same top line, 30–40% lower cost.

Surplus leads

A client generates 50,000 leads a month and their agents can only work 25,000. The other 20,000 are never touched.

Pure incremental revenue.

Non-tech-savvy audiences

Farmers, truck drivers, ride-hailing drivers. They do not recognise a bot, so they engage with it.

2min+ calls against 45–50 seconds.

The pattern behind all three

A bot wins wherever the alternative is nobody calling at all, or a human wasting time on a lead that was never going to convert. It loses wherever it is competing directly with a good human agent on the same conversation.

In one case, AgroStar, the bot outperformed human agents on conversion outright.03
1 · Why Sense exists

The ceiling was not the model. It was the hang-up.

Bot dials50% connect, as a human would
First ten secondsthe caller works out it is a bot
20–30% hang upof everyone who answered
Marked not interestedno qualifying answer was given
Never re-contactedhumans skip them too

Why this compounds

Those leads are not merely lost, they are mislabelled. The bot records no interest, so the human agent never calls them either. A silent graveyard forms inside the client's own CRM.

The second, smaller cause

Negotiation and persuasion. On a ₹5,000 or ₹10,000 purchase there is a human nuance the bot does not have, and that is a capability gap rather than a software one.

One of these is fixable in software. The other is not. Sense is what happens when you attack the fixable one properly.

Detection, not intelligence, was the dominant loss.04
1 · Why Sense exists

Clients did the arithmetic and left.

Human-only against a hybrid setup, per 100 leads
Human only10 buy 50 spoken to, 10 purchases. Expensive, and it works.
Bot qualifies, human closes8 buy 20–25 passed through, 8 purchases. Cheaper, and it costs revenue.

40–50%
Cost saved by the hybrid setup
20–30%
Top line lost in exchange
No
Is that a trade a client accepts?

At 90% of a human the argument works: take a 10% revenue hit, save 40% of cost. At 60% it does not, and no amount of discounting fixes it. That is why annual contracts churned at scale, and it is the entire commercial reason Sense exists.

The product worked. The economics did not.05
2 · What Sense changed

Three changes, one idea.

The bot stops being a caller and becomes an agent that owns a lead until a goal is met.

01 · Persistence with context

The hang-up is no longer terminal. Call again, carrying what was said, until the goal is met or the window closes.

02 · Omni-channel

Voice, WhatsApp, SMS, email. If they will not take a call, message them. Same context on every surface.

03 · An intelligence layer

Something decides which channel, at what time, with what content. That decision is the product.


The design brief was literally: replicate how a good salesperson operates. They call, they wait, they message, they try again at a better hour, and they remember the last conversation.

Omni-channel without shared context is just the same message on more surfaces. The context is the differentiator, which is why the full term is contextual omni-channel.

Session 1 called this the intelligence layer. Sessions 2 and 4 show what it actually is.06
2 · What Sense changed

What that looks like on one lead.

Morning

“I'm calling from Physicswallah about your recent website visit. Are you exploring JEE or NEET courses?”

“Yes, but I'm driving. Call me back at 5.”

context
carried →
5pm, generated automatically

“We spoke this morning. Can we talk now about whether the course is right for you?”

No pickup? Retry tomorrow, then the next day, until the window closes.


Under VoiceBot this lead produced a disposition and nothing else. Under Sense it produces an answer, because the callback was scheduled at the time the customer named and opened by referencing the earlier call.

The lead never repeats themselves. That is the whole reason the second call reads as a follow-up rather than a fresh cold call, and it is what converts a retry into a relationship.

The same context moves to WhatsApp if the calls are not being answered.07
2 · What Sense changed

Next Best Action is the product.

Before every call or message, an LLM decides what to do next. Nothing here is rule-based.

Inputs
Prior interactions

everything said so far

System message

campaign intent

Knowledge base

the agent's prompt

NBA

an LLM call, per lead

channel · time · content

Output
The action

an AI call, or a WhatsApp message

Channel reason

why this channel, recorded

Action reason

why now, recorded

The part that makes it defensible

Every action records why that channel and why now, visible per lead in the dashboard. The cooldown between actions is also LLM-set rather than configured. Two leads in the same campaign can be contacted on different channels at different intervals, without anyone configuring that.

STT, LLM and TTS are all bought in. This loop is not.08
3 · How it is built

Three layers, 2.5 seconds, four vendors.

DialerExotel, or the client's own
STTDeepgram
LLMOpenAI
TTSCartesia
Dialerback to the customer
One turn, end to end
Dial
STT
LLM
TTS
0.0s2.5s
Why latency is the lever

A human replies in about a second. At three or four seconds the illusion breaks. Latency, transcription accuracy, response relevance and voice quality compound; get one wrong and the other three cannot rescue the call.

Vendor choices are cost decisions

Cartesia replaced ElevenLabs because the quality delta did not justify the price delta at Convin's volume. TTS is billed per million characters, so prompt length feeds voice cost as well as model cost.

On annual contracts the client's own dialer is used, which removes a telecom migration from the sale.09
3 · How it is built

An agent is assembled bottom up.

Campaign

Leads, engagement days, channel limits, DND hours. The thing you actually run.

Agent

A goal, a personality, channels and tools. Created once, reused across campaigns.

Knowledge basethe prompt
Goalone line
Personalitytone + style
Channelswhat it may use

The goal field is load-bearing

One sentence: “qualify those interested in booking a medical visit.” Over-specify it and it never registers as achieved, so the agent keeps re-contacting the lead. An over-specified goal produces a spammy agent, and spam is what gets an account reported.

Which reorders the work

Meta integration and template configuration are secondary. Handed a platform and told to run a pilot, agent configuration is the first real work.

The knowledge base is the brain; the agent wears it; the campaign runs it.10
3 · How it is built

The prompt is built from the client's own best calls.

Collect recordingshuman agent to customer
Transcribe~20, the working minimum
Extractfillers, flow, objections, tone
Assembleinto eight sections
01Agent persona
02Overall goal
03State machine flow
04Tone, style and language
05Behavioural rules and guardrails
06Task logic
07Objection handling
08Output formatting
Why recordings and not the script

A client's script says what they intend to say. The recordings show what their best agents actually say, including the objection handling and the small acknowledging phrases nobody wrote down. Naturalness is not a model capability you buy; it is an authoring output.

The difference between a good answer and a great one is often one filler clause.11
3 · How it is built

Most competitors still ship a decision tree.

Node-based, and where it breaks

A greeting node, a yes branch, a no branch. Then the customer asks “why are you calling me?” and there is no node for it. The bot jumps somewhere else and has to trace back. It demos beautifully and collapses on the first unscripted question.

One prompt, holding every path

The whole conversation design becomes a single prompt sent to the LLM for every response. An unexpected question is already answerable from inside the prompt, without leaving the flow.


What that costs, on a ten-minute call
~85k
Characters in a large prompt
50+
Turns in a ten-minute call
Every turn
Re-sends the whole prompt
Plus
The conversation so far, growing

Which makes prompt length a cost line, not a style preference. It also degrades answers: over-information is what causes the model to fail on something the prompt already contains.

Trimming a prompt improves margin and answer quality at the same time.12
4 · The constraints

The best channel is the one Convin does not own.

Meta owns the account, approves every template, sets the daily limit, and can switch the whole thing off.

Template category decides both cost and reach
Utility~100% Delivery. Costs about ₹0.12 per message.
Marketing~50% Delivery, ranging 30–70%. Costs about ₹0.95 — seven times more, for half the reach.

The 24-hour window

One reply, even a single dot, unlocks free-form messages for 24 hours. After that, a template again.

The rolling limit

How many unique customers you may open a conversation with per day. A new account starts around 250.

Consent is the lever

Utility is defensible only with terms the customer accepted. Consent capture moves a message from ₹0.95 to ₹0.12.

Reachability, not creative, is the binding constraint on a marketing template.13
4 · The constraints

What a client actually pays.

LineRateUnitNotes
Lead management₹1per lead uploadedCovers the LLM cost of NBAs, metrics and entities. Charged whether or not anything is sent.
Voice₹1per 15-second pulseCharged only on connect. An unanswered call costs nothing.
Utility template₹0.12per messagePaid by the client directly to Meta, not to Convin.
Marketing template₹0.95per messageAlso to Meta. This is where a campaign accrues a real balance.
Two charging models, and it matters

WhatsApp bills on send, whether or not anyone reads it. Voice bills only on connect. Which is why a silent-delivery problem is almost always a lapsed card rather than a broken bot.

A 1,000-lead pilot on utility templates costs the client a few hundred rupees in Meta fees.14
4 · The constraints

Two clients learned this the expensive way.

Loadshare — a limit nobody checked

A brand-new business portfolio starts at roughly 500 conversations a day. A thousand leads were uploaded on day one. The limit was crossed and the entire portfolio was banned, taking down every WhatsApp sender the client owned for ten to fifteen days.

Control that came out of it: read the messaging limit before uploading.

Snabbit — a targeting failure, not a settings one

Bought lead lists, used to recruit home-service workers. Much of the list turned out to be educated professionals, who reported the sender in volume. The account was blocked, appealed successfully, then the same campaign resumed. Within about eight days the whole portfolio was banned.

No configuration setting would have prevented this.

Which is why the guardrails are the product

Message caps, engagement days, number rotation, DND hours. Recovery is asymmetric: a first block is usually appealable, a repeat offence takes the whole portfolio. That asymmetry should change how hard you push after any warning.

Targeting is a deliverability control, not only a conversion one.15
5 · How it is sold

They made their own pilot smaller on purpose.

The offer, before and after
Was₹1.5 lakh 10,000 leads over a month. Snabbit, Miles Education, Cashify, and ABSLI at 25,000 on the same ticket.
Now₹10,000 1,000 leads over two weeks.

Why smaller is better evidence

Proof arrived in the first fortnight anyway. Beyond that, a bigger lead set adds variance rather than proof: a different lead set behaves differently for reasons unrelated to the product.

The token is a filter, not a price

₹10,000 was never revenue. It exists so companies with no intention of buying do not consume weeks of a PM's time.

Less money per pilot, far shorter time to proof, and more pilots closed. Pricing used as qualification.

Deals now close in days to a week, against a month-plus before.16
5 · How it is sold

A pilot is an experiment, not a demo.

Day 0Internal kickoffsales and PM
Day 1Client kickoffdecision maker present
Bot configurationrecordings, eight pillar
UAT100 leads, two audits
Campaign 1450 leads
Campaign 2450 leads

Closure deck against the benchmarks, then handback to sales. 100 + 450 + 450 makes the thousand.


The move that makes it rigorous

Before anything runs, capture what the client's own human agents achieve: connectivity, interest identification, conversion. A pilot returning 10x proves nothing if their agents were returning more. Matching the human benchmark is already a win, because Sense does it at a fraction of the cost of five or six agents.

Which turns the closing conversation from “was the bot good?” into “here is your number, and here is ours, on the same leads.”

Client testing loops do not close on their own. Internal testing first, then UAT.17
5 · How it is sold

Where that has got to.

80–85%
Of a human agent, against 60% for VoiceBot
7–8
Annual contracts live
25+
Active pilots running in parallel
~100x
Return demonstrated on a strong pilot

The commercial shape

Roughly three pilots running for every contract signed. Consumption pricing means a client who stops finding value stops spending, so the decline shows up in usage months before it shows up in a contract.

And the honest framing

Sense qualifies leads; it does not close them. A team of ten becomes five people plus Sense, not zero people. Said at kickoff that sets a bar the pilot can clear. Said at closure it sounds like an excuse.

Which is the same conclusion Session 1 reached from the business side, arrived at again from the mechanics.

Nine or ten annual deals were signed in the two to three months after Sense launched.18
Close

Four things the mechanics tell you.

  • 01
    The moat is orchestration, not voiceEvery model in the stack is bought in. What cannot be bought is the NBA loop, the Eight Pillar authoring method, and eighteen months of turn-taking work.
  • 02
    Constraints are the productMessage caps, engagement days, number rotation, DND hours. The guardrails are what make an autonomous agent safe enough to sell to an enterprise.
  • 03
    The dangerous failures are silentA lapsed card, an unregistered certificate, a filled rolling window, an over-specified goal. None of them throw an error at the operator; messages simply stop.
  • 04
    Cost per outcome is a PM numberVendor tier, template category, prompt length and window length all move margin. Under consumption pricing that belongs to product, not to finance.
Analysis layered on the sessions, not claims made in them.19
Close

What I would want to confirm in week one.

Points where the five recordings disagree, or where I could not tell from the material. Bringing the list seemed more useful than picking an answer.

Free-form message rateSession 3 gave ₹0.50; Session 5 gave ₹0.05. A tenfold gap on the message type sent most often inside the window.
Marketing template rate₹0.90, then ₹0.95, then ₹0.94–0.95 across three sessions. Converging, but worth pinning to a rate card.
LLM providersDescribed once as open-source models, and later as OpenAI, Gemini, Claude or closed-source. Not the same claim.
The new dialerA dialer purpose-built for voice AI was named as being trialled to replace Exotel. I could not make out the name.

Happy to go deeper on any section. There is a full deck behind each of the five sessions.

Thank you.20
Backup · Reference

Every number, in one page.

VoiceBot efficiency60%of a human
Sense efficiency80–85%of a human
First-10-second drop20–30%of connected callers
Latency budget~2.5 sone full turn
Voice billing₹1 / pulse15 s, on connect
Lead management₹1 / leadNBA and metrics
Utility template₹0.12 · ~100%cost, delivery
Marketing template₹0.95 · ~50%range 30–70%
Free-form window24 hfrom last customer msg
Starting WhatsApp limit250 / daya new account
Transcripts for a prompt~20the client's best calls
Pilot today₹10,0001,000 leads, two weeks
Pilot before₹1.5 lakh10,000 leads, one month
Pilot split100+450+450UAT, two campaigns
Business7–8 · 25+contracts, pilots
Strong pilot return~100x₹10k in, ~₹4L out
Figures as given across the five sessions. Approximations preserved.21