Convin — internal training

Debugging a bot,
and running a pilot

A live teardown of a configured bot that went wrong in four ways, then the commercial story: how the pilot shrank from ₹1.5 lakh and 10,000 leads to ₹10,000 and 1,000, and the 21-day process that replaced it.

Session05
SpeakerSudesh
TopicDebugging, then the pilot SOP
FormatLive teardown, then a document
Duration~75 min

Contents

One teardown, one business model.

  • 1The teardownFour faults in one configured bot, and their fixes03
  • 2Where the reasoning livesThree hidden icons that explain every decision08
  • 3Human in the loopMessaging and calling as a human agent10
  • 4How the pilot model evolvedFrom ₹1.5 lakh and 10,000 leads to ₹10,000 and 1,00014
  • 5What Sense actually costsLead management fee plus consumption19
  • 6The 21-day SOPKickoff, configuration, UAT, two campaigns, closure21
  • 7Setting expectationsAssist not replace, and the hallucination arithmetic27
  • 8PM lens and prepTakeaways, runbook, questions, reference29
Sections 4 to 6 are the ones an interview will probe.02
1.0 · The teardown

One PM tested her own bot.
Four things went wrong.

The most useful half hour in the series: a configured agent run against a real number, with each fault diagnosed live. Every one of them is a configuration error rather than a bug.

01 · Wrong voice

Agent named Arjun, configured male. A female voice called.

02 · Nothing transcribed

The bot re-introduced itself and never advanced, then hung up.

03 · No visible reasoning

Why this channel, why this message, why the goal is still open.

04 · No call button

Human calling enabled on the campaign, and still no way to dial.

Why this matters more than it looks

Session 4 said a PM who understands the system can diagnose a fault themselves and hand the product team a cause rather than a symptom. This session is what that looks like in practice, and every fault above was found without opening a ticket.

None of the four required engineering. All four were findable in the dashboard.03
1.1 · Fault one

A male agent with a female voice.

The agent was named Arjun, gender set to male, and the prompt referred to him throughout. The call came through in a woman's voice, still introducing itself as Arjun.
The assumption

Setting agent gender to male determines how the bot sounds, so the voice should have followed.

What actually controls it

The voice ID in the TTS configuration, which is a separate field. Gender is metadata used in the prompt; the voice ID is the actual audio. A female voice ID had been selected, so a female voice called.

Two fields, two purposes, no validation between them. Nothing warns you that the persona and the voice disagree.

Check the voice ID, not the gender dropdown, when the voice is wrong.04
1.2 · Voice cloning

Two ways to get a voice that is not on the list.

Settings → voice cloning. Both routes go through Cartesia; they differ in how much audio you feed them and how good the result is.

IVC — instant voice cloning

Five to ten seconds of recording, a name, a description and a language. The clone is ready almost immediately and picks up intonation surprisingly well. This is what the product uses today.

PVC — professional voice cloning

One to two hours of audio uploaded into Cartesia's environment, which they train on. Materially higher quality, because the dataset is far larger.


5–10 s
Audio needed for an instant clone
1–2 hrs
Audio needed for a professional clone
Immediate
Turnaround on an instant clone
Cartesia
Where both are trained

Worth trying on your own voice before an interview. It is the fastest way to understand what the TTS layer is doing.

The gap between IVC and PVC is a dataset-size argument, not a different technology.05
1.3 · Fault two

The bot could not hear her at all.

Call connectsthe bot opens normally
She speaksnothing appears in the transcript
Bot re-introducesit has no input to act on
Loopthe conversation never advances
Disconnectthe call ends having achieved nothing

How it was diagnosed

An empty transcript points at the speech-to-text layer, not the model. If STT works everywhere else and fails on one agent, it is a configuration difference on that agent.

The cause, and the fix

Deepgram Nova 3 had been selected directly rather than Convin STT managed. That specific model may not be integrated to this app domain. Switching to the managed option restores transcription.

Which is a useful default to carry: unless you have a reason to pin a specific model, choose the managed option. It resolves to whichever model is actually wired up for that environment.

Empty transcript means STT. Repeating opener means the bot received nothing.06
2.0 · Where the reasoning lives

Three icons that explain everything.

All three are small information icons, easy to miss, and each one answers a question a PM will otherwise raise as a ticket.

Icon locationAnswersWhat it shows
On a messageWhy this message, on this channelThe full decision context: which channel was chosen and why, why this action was taken, and what came from the knowledge base.
Above lead metricsWhy the goal is still openThe reasoning behind goal achievement, lead qualification status and interest level for this specific lead.
Beside the message timestampWhen the next action firesThe cooldown schedule and the time the next best action is due. Hidden until you hover slightly to the right of the icon.
Session 2 said the reasoning is a first-class field. This is where it actually surfaces.07
2.1 · The cooldown control

The most important control is hidden.

The cooldown popup only appears when the pointer is dragged a fraction to the right of the timestamp icon. It took the room several minutes to find it, and the speaker said plainly that this is not the intended experience and will change.

Cooldown scheduledno activity until it expires
Execute manuallyone tap, from the popup
NBA regenerateswith everything that has happened
Action firesan AI call or a WhatsApp message

What it is for

Testing. Rather than waiting hours for a cooldown to expire, execute it and watch the next action fire immediately.

What it does not do

It does not simply resend the last message. Executing the cooldown triggers a fresh NBA, which recalculates and may pick a different channel entirely.

Quoted from the session: “This will be changed eventually, not the ideal UI/UX. Right now we have kept it hidden.” A team that names its own rough edges is easier to work with than one that does not.

A control this useful should not require a pixel hunt to reach.08
3.0 · Human in the loop

Sending as a human, not as the bot.

Switch to human agent, then use the message icon beside the lead's number. Two options appear, alongside a countdown labelled session active.

If a human is typing the message anyway, why is a templated option offered at all?
The reasonable guess

Something to do with the Meta limits on the account or the portfolio.

The actual reason

The 24-hour rolling window from Session 3. Because this customer replied at 16:32 on the 13th, the window runs to 16:32 on the 14th and free-form is allowed. Outside that window, only a template can be sent, by a human or by the bot.

The session-active countdown is the rolling window, made visible per lead.09
3.1 · Fault four

Human calling was on, and there was no call button.

Enabling human call as a channel on the campaign is not sufficient. Three separate things must all be true before the click-to-call icon appears.

01 · Agent channel active

Human call must be switched on in the agent configuration, not only on the campaign. It was inactive here.

02 · User has calling enabled

Settings → user management. Calling is off per user by default, and needs admin access to change.

03 · A number mapped to that user

One of the Exotel numbers procured for human calling is assigned to the individual. Their calls dial out from it.

And a fourth setting that looks unnecessary

The human call channel also asks for a speech-to-text selection, which seems wrong for a human-to-human call. It is not: Convin transcribes human calls too, and feeds that transcript back into the lead's context. When the AI picks the lead up again it knows what the human already discussed.

Permission, assignment and channel state are three different switches.10
3.2 · What the handoff produces

The AI gets the lead back, and knows what happened.

Human dialsfrom their mapped number
Recording startsat call in progress, before pickup
Conversationtranscribed like any other
Dispositionmarked by the human
Ownership returnsNBA resumes on the lead

A detail nobody expects

Recording begins the moment the call is in progress, which is before the customer picks up. Anything said during the ringing period is captured in the transcript. It surprised the room, and it is worth knowing before you speak over a dialling tone.

Why the handoff is not terminal

Once the human ends the call, the lead returns to Convin AI. The next best action is generated with the human conversation in context, so the follow-up references what was actually said.

Which is the substantive difference between a transfer and a handoff. The lead does not leave the system; it comes back better informed.

Session 2 distinguished live transfer from lead handoff. This is the handoff, working.11
4.0 · How the pilot model evolved

It started far too big.

The original motion was a full deployment in miniature: integrate the client's CRM in and out, work around their existing dialer, and prove the concept at scale before anyone had proved it at all.

Jul–AugTwo clients used to validate the concept, unpaid.
Then₹1.5 lakh for a one-month, 10,000-lead pilot.
Now₹10,000 for a two-week, 1,000-lead pilot.

What Sense was at validation

Contextual calling, WhatsApp and an omni-channel loop, with no guardrails at all. The speaker describes even the current product as an MVP. The first two clients were a logistics company recruiting driving partners and a lending business calling tier-2 and tier-3 customers about loans.

The second validation client's name was not clearly audible on the recording.12
4.1 · The ₹1.5 lakh era

Four pilots, one ticket price, one problem.

ClientLeadsTicketNotes
Snabbit10,000₹1,50,000The first paid 10,000-lead pilot. Much of the product was built from their feedback.
Miles Education10,000₹1,50,000Ed-tech. Same shape, same ticket.
Cashify₹1,50,000Ran alongside the others as volume built up.
ABSLI25,000₹1,50,000Insurance. Two and a half times the leads for the same money.
The ticket was never the point

₹1.5 lakh did not make money on a month of work. It existed to prove the concept. Which meant every commercial problem with the model was a timing problem, not a pricing one.

Same ticket regardless of lead volume, which is itself a signal the price was nominal.13
4.2 · Why it broke

Value arrived in two weeks. The contract ran for a month.

The cycle a 10,000-lead pilot actually ran
Get recordingsor just a call script
Configureno eight-pillar yet
Internal testingboth teams audit every call
Deploy to 100–200then audit again, by hand
Repeatuntil someone calls it done

1–2 wks
To demonstrate the concept, on 1,000–2,000 leads
1 month
The contracted pilot duration
By hand
Auditing was done by the PM, not a data team
Not linear
Results on 1,000 leads do not carry to 3,000
The part that is easy to miss

Bigger pilots are not more convincing. A different lead set behaves differently for reasons that have nothing to do with the product: where the leads came from, how old they are, what mindset they are in. Running three times as many leads adds duration and variance rather than proof.

The long tail after the proof point is pure cost to both sides.14
4.3 · The two-week pilot

₹10,000 and 1,000 leads.

Fifteen times less money and ten times fewer leads, deliberately. The token is not revenue; it is a filter.

Why ₹10,000 rather than free

To maintain sanity. A free pilot attracts companies with no intention of buying, and each one costs a PM several weeks. A small cheque filters for seriousness without pretending to be a revenue line.

Why 1,000 leads rather than 10,000

Because 1,000 is enough to show ROI, and everything past that adds elapsed time and variance without adding evidence.


What the client gets at the end
Leads uploaded1,000
Cost incurredfixed plus consumption
Revenue returnedmeasured against their own data
ROIthe number the decision turns on
Two weeks is a target rather than a guarantee; the point is the urgency it creates.15
5.0 · What Sense costs

A per-lead fee, plus what you consume.

Lead management fee — new in this session

₹1 per lead uploaded. It covers the LLM cost of generating next best actions and computing metrics and entities for that lead. A hundred leads is a hundred rupees.

Consumption

Charged on what actually goes out: voice minutes, and messages by category. During a pilot the client pays only the fixed token; consumption is used to compute the ROI figure.


Lead management₹1per lead uploaded
Voice₹1per 15-second pulse, on connect only
Utility template₹0.12per message, paid to Meta
Marketing template₹0.94–0.95per message, paid to Meta
Free-form messagesee notebilled by Convin, covers the LLM

The dashboard shows a voice figure of 1.1025; the working number is 1.1.

The free-form rate is contradicted across sessions. See the next page before quoting it.16
5.1 · One rate, three versions

The free-form message rate is unresolved.

It matters more than it looks. Free-form messages are the ones sent inside the 24-hour window, which is where most of the conversation happens.

Session 3

“The cost is point five, charged from our platform.”

Reads as ₹0.50 per message.

Session 5

“If it's a free-form message I incur five paise.”

Reads as ₹0.05 per message.

Why a tenfold gap is worth resolving

On a thousand-lead campaign with a handful of free-form messages per engaged lead, the difference between five paise and fifty paise is the difference between a rounding error and a real line in the ROI calculation you are about to present to a client.

Marketing has also moved across sessions: ₹0.90, then ₹0.95, now ₹0.94–0.95.17
6.0 · The 21-day SOP

A strict process, and the timeline is the point.

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

Closure deck, then handback to sales for the annual contract.


Timelines are not advisory

Any delay must be raised immediately in the Convin Pilot group. If you own a pilot you own its outcome, and slippage is the default failure mode.

Honest about adherence

Asked whether 1,000-lead pilots actually finish in fourteen days, the answer was no. The deadline still works, because it creates urgency on both sides.

Note the arithmetic: 100 leads for UAT plus 450 and 450 makes the thousand. The pilot is designed so the first hundred are a rehearsal.

Two weeks is the promise; twenty-one days is the process it maps onto.18
6.1 · The two kickoffs

Align internally, then align the decision maker.

Internal kickoff — sales and the assigned PM

What is this client, what is the use case, and how many leads do they have per month. Low monthly volume caps the revenue, which is the moment to ask whether a second use case exists. Human-agent involvement is agreed here, before anyone promises anything.

Client kickoff — the decision maker must attend

Whatever the internal team agreed is worthless if the person signing has a different picture. The DM needs to hear that Sense qualifies rather than closes, and that human agents remain part of the loop.


Three things to leave the client kickoff with
20 call recordings

The minimum for eight-pillar generation. Their best calls, same use case.

A WABA contact

Someone with portfolio admin, free for a fifteen-minute call to do the Meta integration.

Benchmark data

What their human agents currently achieve. Without it you cannot prove anything.

Keep the DM engaged throughout: share good call recordings in the group as they happen.19
6.2 · Benchmark data

ROI on its own proves nothing.

A pilot that returns 10x sounds like a success. It is not, if their own agents were returning more. The comparison, not the number, is what closes the deal.

The three benchmarks to capture before anything runs
Connectivity50% If their agents reach 500 of 1,000 leads, that is the bar Sense has to clear.
Interest identification25% How many of those connected are correctly identified as interested.
Conversiontheir number The final number, human-only, to compare the pilot against at closure.

Matching the human benchmark is already a win, because Sense does the same job at a fraction of the cost of five or six agents. Beating it is upside. Falling well short means the client should hire instead, and you should know that before they do.

Percentages are the worked example from the session, not a client's real figures.20
6.3 · Configuration and testing

Test internally. Not with the client.

Transcribethe 20 recordings
Eight pillargenerated from them
Configure the botagent, channels, campaign
Test internallycolleagues, not the client

Why client testing is a trap

Once the client is testing, the loop does not close. It can run a month or two. There is always another imperfection, and the client is not wrong to see them.

Why they care so much

The calls go out under their brand, from a Truecaller-registered number carrying their name. A bad call is their reputational problem, and it can end up on LinkedIn or Instagram. For a large brand that is a real risk, so they will push for perfect.

The honest position to hold: there is always a delta, and it cannot be perfect. The pilot is judged on business outcome, not on whether every call was flawless.

Internal testing is the single biggest schedule protection in the SOP.21
6.4 · UAT on 100 leads

Two audits, running in parallel.

The first hundred leads of the thousand are a rehearsal. They are audited twice, by two teams looking for different things.

Data Labs — quality
  • STT, TTS and LLM failures
  • The responses actually generated
  • Latency, and the bot going silent
  • Entity extraction accuracy
  • Interest-level accuracy
The PM — conversation

You configured the agent, so you know how the conversation was meant to go. You are auditing whether it went that way: whether the flow held, whether objections were handled, whether the bot stayed inside its guardrails.

What a good UAT buys you

If the UAT result is strong, the friction disappears for the rest of the pilot. The client hands over the remaining leads without argument. If it is weak, every subsequent campaign is negotiated.

In the ₹1.5 lakh era this audit was done by the PM alone. Data Labs is the newer half.22
6.5 · The two campaigns, then closure

450, then 450, then a deck.

Campaign 1450 leads, tracked live
Incorporatefeedback taken mid-flight
Campaign 2450 leads, the final run
Collateconnectivity, conversion, ROI
Presentto the DM and stakeholders

What goes in the closure deck

Leads uploaded, connectivity achieved, conversion achieved, cost incurred, and the ROI that falls out of them, set against the benchmarks captured at kickoff.

Then the PM hands back

Sales re-enters and negotiates the annual contract from the pilot numbers. The PM's job ends at a defensible result, not at a signature.

Feedback during campaign one is incorporated while it runs rather than deferred, which is what makes two campaigns better than one long one.

The annual contract itself is a separate conversation, not covered in this session.23
7.0 · Setting expectations

Assist, not replace. Say it early.

A client sees one demo of contextual omni-channel outreach and concludes it will replace their agents. They are not being unreasonable; the demo is genuinely impressive. But the expectation has to be corrected before the pilot, not after it.

The framing the session uses
5 roles Sense can absorb
5 roles that remain human
qualification, first contact, follow-up closing, negotiation, judgement
The line worth memorising

Every AI currently can assist a human being, not replace one. Said at kickoff it sets a bar the pilot can clear. Said at closure it sounds like an excuse.

The proportions above illustrate the argument; they are not a measured split.24
7.1 · The other expectation

Some call will go wrong. That is arithmetic.

During internal testing clients will report that a word was misheard, that the pacing was uneven, that the bot did not stop when it should have, that it hallucinated at minute eight. All of that is real, and some of it is unavoidable.

Why it compounds
One API calla 20–25% chance of going wrong
Sense chains manyNBA, response, validation, more
Across a whole callstep five, step seven, minute eight
Probability risessomething will land badly
Where the conversation has to land

Not on call quality, on business sense. Ten thousand rupees in and seven lakh out is roughly seventy times the investment, and that argument survives a handful of imperfect calls. The quality conversation is real and it belongs to the roadmap, not to the pilot decision.

The 20–25% figure is the speaker's rule of thumb for a single call, not a measured error rate for the product.

Be careful quoting the percentage. Quote the reasoning.25
8.0 · PM lens

Five things this session tells you.

  • 01
    They shrank their own pilotFrom ₹1.5 lakh and 10,000 leads to ₹10,000 and 1,000. Less revenue per pilot, far shorter time to proof, and more pilots closed. A rare example of deliberately making the deal smaller.
  • 02
    Benchmarks turn a demo into an experimentCapturing human-agent connectivity, identification and conversion before anything runs is what converts an ROI number into evidence. Without it the pilot proves nothing.
  • 03
    The token price is a filter, not a price₹10,000 exists to screen out companies that will never buy. Pricing used as qualification rather than as revenue.
  • 04
    Every fault in the teardown was configurationWrong voice, dead STT, missing call button, invisible reasoning. None needed engineering, all needed someone who knew where to look.
  • 05
    The product admits its rough edgesA hidden cooldown control, an eight-pillar step still outside the dashboard, fields that do nothing. Being told this in training is a good sign about the team.
PM framing. Analysis layered on the session, not claims made in it.26
8.1 · Pilot runbook

What to have, and when.

WhenWhatWhy it matters
Before the client kickoffUse case, monthly lead volume, human-agent planAgreed with sales at internal kickoff
At the client kickoffDecision maker on the callNon-negotiable. Alignment without them does not hold
From the client20 call recordings, a WABA contact, benchmark dataAll three, or the pilot cannot be judged
Before any client testingInternal testing completeClient testing loops do not close on their own
At 100 leadsData Labs quality audit, PM conversation auditRun in parallel, then incorporate
At 550 leadsCampaign 1 reviewed, feedback appliedMid-flight, not deferred to the end
At 1,000 leadsClosure deck against the benchmarksPresent to the DM, then hand back to sales
Raise any slippage immediately. Silent delay is how a two-week pilot becomes two months.27
8.2 · Interview prep

Questions an interviewer will actually ask.

Commercial“Why would you shrink your own pilot?”₹1.5 lakh and 10,000 leads took a month and proved nothing extra. ₹10,000 and 1,000 leads proves the same thing in two weeks and closes faster.14 · 15
Rigour“How do you know a pilot succeeded?”Against benchmarks captured at kickoff: their human connectivity, identification and conversion. ROI alone is not evidence.20
Pricing“What does a client actually pay?”₹1 per lead as a management fee, plus consumption: ₹1 per 15-second pulse on connect, and per-message rates by template category.16
Debugging“The bot is not responding. Where do you start?”Empty transcript means STT. Check whether a specific model was pinned instead of the managed option. Then the three reasoning icons.06 · 07
Judgement“A client wants to keep testing. What do you do?”Move them to UAT on 100 leads with a real audit. Client testing loops do not close, and the calls carry their brand, so the anxiety is legitimate.21 · 22
Honesty“Will this replace our agents?”No. It absorbs roughly half the roles in a team and the rest stay human. Say it at kickoff, where it sets a bar, not at closure where it sounds like an excuse.24
The commercial question is the one to lead with. It shows judgement rather than recall.28
8.3 · Reference card

The numbers, in one place.

Pilot token₹10,000two weeks, 1,000 leads
Previous pilot₹1,50,000one month, 10,000 leads
Lead management₹1 / leadcovers NBA and metrics
Voice₹1 / pulse15 s, on connect
Utility template₹0.12per message, to Meta
Marketing template₹0.94–0.95per message, to Meta
Free-formunresolved₹0.05 or ₹0.50; verify
Recordings needed20minimum, eight pillar
UAT size100 leadsaudited twice
Campaign split450 + 450completes the thousand
SOP length21 daysbehind a two-week promise
Instant voice clone5–10 saudio needed

Named pilots

Loadshare and a lending client for unpaid validation.  ·  Snabbit, the first paid 10,000-lead pilot.  ·  Miles Education, 10,000 leads.  ·  Cashify.  ·  ABSLI, 25,000 leads at the same ₹1.5 lakh ticket.

All figures from Session 5. Approximations preserved as given.29
8.4 · Before you quote any of this

What still needs confirming.

Carried across all five sessions. Each is a place where the recordings disagree, or where the audio was not clear enough to be sure.

Free-form rateSession 3 said point five; Session 5 said five paise. A tenfold gap on the message type sent most often.
Marketing rate₹0.90 in Session 2, ₹0.95 in Session 3, ₹0.94–0.95 in Session 5. Converging, but ask for the rate card.
LLM providersSession 2 described open-source models; Session 4 described OpenAI, Gemini, Claude or closed-source.
Sign-in methodSession 2 said Google or Microsoft; Session 3 said Microsoft only.
The new dialerNamed in Session 4 as being trialled to replace Exotel. The name was not audible.
Second validation clientNamed alongside Loadshare in Session 5. A lending business; the name was not clear.
Six open items across five sessions. Bringing the list is a better move than picking answers.30