What does an AI calling agent learn from 19,000 Indian calls?

What we saw. Replies came two to three and a half seconds after the caller stopped. Almost none of it was thinking: it was waiting for the transcript to settle, a half-second aggregation buffer, and the voice starting up.

What changed. The agent starts on the transcriber’s running text the moment the caller stops and confirms it against the final transcript before the reply is spoken; a mismatch discards the reply. Replies now land in one to one and a half seconds, and every turn is timed and stored with the call.

What we saw. A caller paused for a second mid-sentence (“मेरे पास…”) and was answered in the gap. The rest of the sentence became a second question, and the first reply was to half a thought.

What changed. A semantic end-of-turn model judges whether the sentence is finished before the agent speaks. A pause of up to a second and a half is held; one sentence gets one reply.

What we saw. Callers who spoke as the introduction started cancelled it, on about one call in sixteen.

What changed. The introduction is shielded from interruption for its few seconds. Their words are held and answered right after.

What we saw. A callback time was confirmed back as “kal 11 baje” in Latin letters, so the voice read “eleven” in English.

What changed. The confirmation is spoken from the parsed time, in Devanagari, every time.

What we saw. When the caller answered the time question with a clear time, the booking still went through a model round-trip before it was saved.

What changed. A clear time answer to a clear time question is saved directly and confirmed at once; the model is not consulted.

What we saw. A caller said “अभी” to the offer and the agent booked a callback for right now. They had agreed to the idea, not to a time.

What changed. A yes to the pitch is no longer a booking. The agent asks “अभी ठीक रहेगा, या बाद में?” and books only on the answer.

What we saw. A carrier’s “this call may be recorded” value-added announcement was heard as the customer speaking.

What changed. The announcement is recognised as a recording and skipped, like a voicemail greeting.

What we saw. The agent said goodbye and hung up in the same instant the customer gave a time. A real booking was lost to a hang-up already in flight.

What changed. A booking that lands while the close is in progress wins. The close is cancelled and the time is saved.

What we saw. Seven customers went silent right after agreeing to hear more.

What changed. One cause found: the model occasionally returned an empty reply to a complete sentence. An empty reply is now re-run once, and every turn logs what the model said and how long it took, so the rest can be pinned down.

What we saw. One number placed 3,600 dials a day, most of them same-day redials. Within a week carriers flagged it and the answer rate dropped by a third.

What changed. Numbers now live in a pool with warm-up and daily caps. A person is dialled at most twice a week, retries move to the next day, and a refusal is left alone for thirty days.

What we saw. Two customers gave a time and the agent typed the booking into its reply instead of saving it. Two hot leads reached the CRM as plain text.

What changed. Any booking the agent writes out is caught and saved as a real callback. The reply is spoken without the internal note.

What we saw. A complaint was answered with “बढ़िया!” because a word in it looked like agreement.

What changed. Complaint phrasing is checked before the cheerful thank-you line can be spoken.

What we saw. A call-screening assistant kept the line open and the agent waited on it.

What changed. A screener that stalls is closed after a short wait instead of holding a channel.

What we saw. A Pixel phone answered with “X’s assistant. May I ask who’s calling?” and the agent spoke to it for two minutes. The lead was then marked Wrong Number.

What changed. Phrases used by call-screening robots are recognised. The agent gives a one-line reason, ends the call, and the person is retried on another day.

What we saw. Three calls carried a bare “जी.” right after the introduction, because a filler meant for long silences fired during the opening.

What changed. Fillers cannot run while the opening is being spoken.

What we saw. Two analysis providers ran out of quota on the same morning and thirty calls went unanalysed.

What changed. Three providers in a chain for post-call analysis; the calls were re-analysed.

What we saw. Sentences that opened with नहीं or “no” but continued with अभी or कभी भी were read as refusals. Four customers who wanted to talk were hung up on.

What changed. The whole sentence decides, not the first word. Replayed against the old calls: ten of fifteen now continue.

What we saw. A closing line with “aapka” in Latin letters was pronounced as English.

What changed. Every Hindi word is written in Devanagari before it reaches the voice.

What we saw. Four of seven callers stayed silent for seven seconds after a bare introduction.

What changed. The introduction and the first question are one breath, so the caller always has something to answer.

What we saw. Most callers speak the moment they pick up. Their first word cancelled the greeting, so six in ten live callers never heard who was calling.

What changed. Words spoken over the greeting are held, the greeting completes, then the agent answers them.

What we saw. An eleven-second offer line lost three callers mid-sentence.

What changed. The offer is one short sentence with the question at the end.

What we saw. “Twenty five”, “Wednesday 3 baje” and “N बजे के बाद” were parsed wrongly.

What changed. Weekday-plus-clock, spelled-out numbers and after-N-o’clock all book correctly.

What we saw. Replaying 167 recordings showed 28 percent of audible customer replies were never transcribed, mostly one-word answers spoken softly.

What changed. A second decode pass catches short replies the live transcriber missed, and the voice detector starts a tenth of a second earlier.

What we saw. The voice sounded choppy on some lines. The cause was punctuation: long dashes in the script were read as hard stops.

What changed. Dashes become commas before speech.

What we saw. A call reached 418 seconds because background noise kept resetting the silence timer.

What changed. Silence is measured on the customer’s speech only. Noise does not keep a call alive.

What we saw. “तारीख”, evening “8 बजे”, ranges like “5 से 7”, “ten thirty” and “परसों” were mis-booked or asked again.

What changed. The callback parser handles dates, ranges, evening hours and spelled-out times; evening is assumed for small numbers said in the afternoon.

What we saw. A one-word reply that the transcriber dropped made the agent say it could not hear and greet again.

What changed. The agent never re-greets. A dropped reply gets one short re-prompt.

What we saw. Under four concurrent campaigns the analysis provider rate-limited and outcomes came back empty.

What changed. Analysis falls back to a second provider; the 34 affected calls were re-analysed.

What we saw. In about one call in five a television, a second person or street noise triggered a spurious “I can’t hear you”. One live customer was hung up on mid-sentence.

What changed. Speech that is not aimed at the agent is ignored, and the agent never hangs up without a goodbye.

What we saw. The agent occasionally narrated the name of an internal action out loud.

What changed. Anything that looks like an internal instruction is stripped before speech.

What we saw. Campaigns had no pause.

What changed. Pause parks the queue, resume restores it, and in-flight calls finish first.

What we saw. Scripts hard-coded “sir”. Many customers are women.

What changed. The agent uses “जी” unless the customer’s own words make sir or madam right.

What we saw. Nothing stopped a customer from reading out an OTP or PAN.

What changed. The agent stops them gently, on every script, and never asks for any of it.

What we saw. One “no” was sometimes followed by a second pitch.

What changed. One no means no. The agent thanks them and closes.

What we saw. Voicemail greetings received the full monologue.

What changed. Voicemail is recognised and the call ends without leaving a message.

What we saw. Single-word interjections like “हाँ” interrupted the agent mid-sentence.

What changed. An interruption needs at least three words; short acknowledgements are heard but do not cut the agent off.

What we saw. The agent heard its own voice on some lines and answered itself.

What changed. An echo filter drops the agent’s own words coming back down the line.

What we saw. Quiet callers on 8 kHz lines were not detected as speaking at all.

What changed. The voice-activity threshold was tuned on real recordings.

What we saw. Calls over 20 seconds of silence stayed open for minutes.

What changed. Two re-prompts, then a polite close.

What we saw. Greetings were generic.

What changed. Every greeting uses the customer’s first name when there is one, and drops the slot cleanly when there is not.