RESEARCH · AUG–OCT 2026

Why People Hang Up on AI Calls in India: 2,053 Call Transcripts Analysed

Across 2,053 outbound AI call transcripts from 325 Indian businesses, 30% of people ended the call within their first two replies. The causes we could see were mostly in the script, not the voice: openers that didn't say who was calling (callers asked "who is this?" twice as often), template placeholders read out loud, and agents with no answer to simple questions. Agents that leaked placeholders lost about twice as many callers early, because a leak marks a script nobody reviewed.

R

Ratnam Sachan, Co-founder · Vyora AI

Published · Last updated

2,053

Call transcripts

325

Businesses

30%

Hung up within 2 replies

15%

Ended in a commitment

Key findings

  1. 01

    When the opening line only checked the caller's name, 11% of callers had to ask who was calling or why. When it named the company, 6% did. The direction was the same on free and paid accounts.

  2. 02

    Naming the company did not, by itself, change how many calls ended in a booked next step. The gap we saw in one month disappeared over two months and when each business was weighted equally.

  3. 03

    Agents on free-trial accounts read a leftover placeholder out loud ("[Company Name]", "{{}}", "Sir/Ma'am") in 9.6% of calls. Those agents lost 39% of callers within two replies, against 18% for trial agents without leaks.

  4. 04

    A placeholder leak is a symptom, not the cause: within the same business, calls with and without a leak ended about the same way. Agents that leak are agents nobody reviewed, and the rest of the script shows it.

  5. 05

    11.7% of trial-account calls ended in a dead-end promise ("main check karke update karungi") because the agent had no answer to a price, slot or order question. Across all accounts it was 5.2%.

  6. 06

    26% of all calls, and 40% of trial-account calls, had the agent mixing Hindi and English. Callers mix freely whatever the agent speaks.

How often does each problem happen?

Free-trial accounts are where agents are first written and tested. Paid accounts run tuned, larger campaigns, and are included in "All calls" but are too few businesses to show on their own.

MeasureAll calls
2,053 · 325 businesses
Free-trial accounts
753 · 315 businesses
Hung up within their first 2 replies30%20%
Ended in a commitment (date, time, visit, demo)15%21%
Agent read a placeholder out loud4.5%9.6%
Agent gave a dead-end "I'll check and update you"5.2%11.7%
Caller had to ask who was calling or why7.4%8%
Caller's first word was a bare "Hello?"11%15%
Agent mixed Hindi and English26%40%

Paid campaigns call larger, colder lists, which is why the all-calls hang-up rate is higher even though those scripts have fewer mistakes.

Does the opening line decide whether people stay on the call?

It decides how much friction there is, not whether the call converts. We grouped every call by its first sentence. When the agent only checked the caller's name ("Am I speaking with Rahul?"), 11% of callers had to ask who was calling or why. When it named the company, 6% did. That held on free and paid accounts alike.

What did not hold was a conversion effect. In one month of trial calls, company-named openers ended in a booked next step far more often. Over two months, and with each business weighted equally, that gap disappeared. Saying who you are removes a reason to hang up; it doesn't create a reason to stay.

First lineCallsCaller asked who/why
Name + company ("Main Priya, Sharma Motors se…")1,4516%
Name check only ("Am I speaking with Rahul?")60211%

A comparison of what happened, not a controlled test: different businesses called different lists. A separate study of about 450,700 inbound AI calls by Upfirst also linked naming the business in the greeting to fewer hang-ups (reported by Devdiscourse, July 2026).

We break down the wording that works, with Hindi and Hinglish scripts for eight use cases, in AI calling opening lines that work in India.

What do callers ask when the agent doesn't say who it is?

They ask, and the call stalls. In 7.4% of all calls the person had to ask who was calling or what the call was about: "kahan se bol rahe ho?", "which company?", "kis cheez ka payment?". The worst cases were payment reminders that stated an amount and asked for a date without saying what the money was for. Callers answered with a question, or with "mera koi paisa baaki nahi". Reminders that named the invoice or product went straight to a date ("pachchis tareekh tak", "kal kar denge").

How often do AI agents read placeholders out loud?

In 9.6% of calls on free-trial accounts, and 4.5% of all calls. These are template leftovers that a person reading the script would skip but a voice agent speaks word for word: "[Company Name]", "[BRAND NAME]", an empty "{{}}", "Good morning/afternoon/evening", "Sir/Ma'am", "bol raha/rahi hoon". Bad contact data caused the same effect. An empty name cell exported from a spreadsheet became "Am I speaking with nan?", and one list had phone numbers in the name column, so the agent read the digits out as the person's name.

Trial agents that leaked lost 39% of callers within their first two replies, against 18% for trial agents that didn't, and the gap held with each business weighted equally. But inside the same business, calls with and without a leak ended about the same way. The leak is a marker: an agent nobody read through before launch tends to have weaker questions, missing facts and a vague opener too. Fixing placeholders is the cheapest check; reading the whole script is the real fix.

Vyora's agent editor and campaign upload now flag these before a call goes out. The full list of mistakes, and how to fix each, is in 9 AI voice agent prompt mistakes.

What happens when the agent doesn't know the answer?

It promises to call back. 11.7% of trial-account calls, and 5.2% of all calls, included a line like "main check karke update karungi" or "I'll have the team follow up" in reply to a straightforward question: package prices at a clinic, available interview slots, why a delivery was late. The voice and language were fine. The agent simply had no facts to work with. These questions are predictable, so the fix is a short FAQ or knowledge base with the 10 questions customers actually ask, and a human transfer for the rest.

Do Indian callers reply in Hindi, English or Hinglish?

Hinglish, whatever the agent speaks. On free-trial accounts Hindi agents ran 61% of calls and English 20%, with Telugu, Marathi and Tamil making up most of the rest. Across all accounts English led at 58%, because larger paid campaigns ran in English. Callers mixed freely either way: "Ma'am EMI kal bhijwa dunga", "Budget hoga tees hazaar aur starting date October se". 26% of all calls, and 40% of trial calls, had code-mixed agent speech too.

The calls that went wrong were language mismatches, not mixing: a Marathi greeting on a Hindi agent, a Bengali opening followed by Hindi questions, a Telugu prompt typed in Roman letters. Several callers asked to switch mid-call ("Can we talk in Hindi?", "Telugu language"). Agents that offered a language choice up front kept them. For how this works across eight languages, see AI voice agents in Hindi, Hinglish and Indian languages.

Why do so many calls start with "Hello?"

11% of callers' first words were a bare "Hello?" (15% on trial accounts). That usually means the agent started its intro before the person had the phone to their ear, or spoke over the pickup, so the opening line was lost and had to be repeated. Two related patterns showed up less often but recurred: "Are you still there?" firing a few seconds into a pause, and the agent repeating the same sentence twice in a row. A short pause before the intro and a sensible silence timeout fix most of it.

What did the best calls have in common?

  • ✓

    They said who and why in one breath. Name, company, reason, then a question.

  • ✓

    They asked one thing per turn. A real-estate qualifier went self-use or investment, then 1 or 2 BHK, then a visit day. "Naam, city, state aur preferred language" in one question confused people.

  • ✓

    They read details back. An order confirmation that read the delivery address back got a clean "yes" and closed in under a minute.

  • ✓

    They handled the exits. A wrong number got "sorry, I won't disturb you again". A busy caller got "kab free honge?" instead of the pitch.

  • ✓

    They matched the caller's language. Either the agent asked at the start or it switched the moment the caller did.

"We expected the voice to be the thing people judged. It wasn't. People forgave an accent; they didn't forgive '[Company Name]' or a call that never said why it was ringing them. Almost every fix here is in the script, not the model."

Ratnam Sachan, Co-founder, Vyora AI

Methodology

Period. 9 August to 8 October 2026.

Population. Outbound calls on Vyora with a transcript of at least three conversational turns: 2,053 calls from 325 businesses, 753 of them from free-trial accounts. Vyora's internal accounts are excluded. Trial accounts include first-draft agents and test calls to colleagues. Answer rates for campaigns are in our Q3 2026 campaign report.

Definitions. "Hung up within 2 replies" means the person called ended the call after speaking no more than twice. "Commitment" means the agent proposed a time, date, visit, demo or booking and the person replied with a specific time or day; it is matched by keyword, so read it as directional. "Placeholder leak" means the agent said bracketed text, an empty or malformed variable, a slash alternative like "Sir/Ma'am", or "nan" for a name.

Anonymisation. Only aggregates are published. Quoted phrases are short, common expressions with no names, numbers or brands. A figure is shown only when it covers at least 30 calls and 5 businesses, and no single business makes up more than half of it.

Robustness. Every comparison was checked across two 30-day halves, with each business weighted equally, and within businesses where possible. We report an effect only if it survived those checks. A first pass on 30 days of trial calls showed company-named openers converting far better; it did not survive, so we report no conversion effect.

Limitations. Correlation, not causation: businesses chose their own scripts and lists. Speech-to-text can mishear, especially in Tamil, Telugu and Bengali.

How to cite this research

Vyora AI (2026). Why People Hang Up on AI Calls in India: 2,053 Call Transcripts Analysed. https://www.vyora.ai/research/why-people-hang-up-ai-calls-india

Frequently asked questions

Why do people hang up on AI voice agent calls?

In our analysis of 2,053 outbound AI call transcripts from 325 Indian businesses, 30% of people ended the call within their first two replies. The avoidable causes we could see in transcripts were confusion about who was calling, template placeholders read out loud, and dead-end answers when the agent lacked facts. Agents with placeholder leaks lost about twice as many callers early, though that reflects an unreviewed script overall rather than the leak alone.

Should an AI agent say the company name in the first line?

Yes. Callers asked "who is this?" in 11% of calls when the opener only checked their name, against 6% when it named the company. It did not, on its own, raise the share of calls ending in a booking, so treat it as removing friction rather than a conversion trick.

Do Indian customers prefer Hindi or English on AI calls?

Both, mixed. On free-trial accounts 61% of calls used Hindi agents and 20% English. Across all accounts English was 58% because larger paid campaigns ran in English. Whatever the agent spoke, callers mixed Hindi and English, and 26% of all calls had the agent mixing too.

How was this data collected?

We reviewed the transcripts of outbound calls placed on Vyora between 9 August and 8 October 2026, keeping calls with at least three conversational turns, from free and paid accounts. Vyora's internal accounts were excluded. Only aggregate figures are published; no names, numbers, brands or recordings.

Does telling people it is an AI make them hang up?

Our transcripts do not show disclosure driving hang-ups. Very few callers asked whether they were speaking to a machine; far more asked who was calling. A separate study by Upfirst of about 450,700 inbound calls found that disclosing AI was linked to roughly 20% fewer hang-ups.

Build an agent that says the right thing first

50 free credits, no card. The editor flags placeholders before your first call.

Related

We use cookies to improve your experience. By continuing to use this site, you agree to our Privacy Policy.