Vinod Khosla says the Indian contact center will die. He is right about the business. He has the wrong cause of death.

“IT and BPO services will disappear, almost certainly within the next five years.” Vinod Khosla, February 2026.

The reaction in India went outrage, denial, then a dozen panels on reskilling. Nobody asked the plainer question. What does a contact center actually produce, and is anyone still buying it?

Take one business where the contact center is the entire sales engine. Used cars.

How a used car gets sold in India

A seller types the car into an app and gets an instant estimate. Within minutes a tele-caller rings, in Hindi, Tamil, Telugu, Marathi, Kannada, Bengali, or whichever of twelve-plus languages the seller picked up in. The telecaller has one job: book an inspection. An evaluator then drives to the car, runs the checklist, and quotes a real price. The seller accepts, haggles, or vanishes. The car that is bought gets refurbished and sold to a buyer who arrived through the same funnel, in reverse.

Every step is a phone call. The inspection is the expensive one. Here is what the numbers look like.

  • India will trade about 6 million used cars this year, worth roughly Rs 4 lakh crore. For the first time, 1.4 used cars change hands for every new one sold. (CRISIL)
  • Four in five of those sales still happen outside any organized platform. (Redseer)
  • The largest platform runs over 1,00,000 inspections a month. Six in ten inspected cars differ from what the seller typed. Four in ten final prices land below the online estimate.
  • That same platform sold about 2,00,000 cars in FY24. Set against 12 lakh inspections a year, that is roughly one purchase for every six evaluator visits.
  • FY25 losses: Rs 543 crore at one platform, Rs 424 crore at another. Procurement is over 80% of cost. Marketing at one of them was Rs 125 crore.
  • A fully loaded telecaller costs Rs 22,000 to 28,000 a month. A connected call costs Rs 7 to 13. Attrition runs 35 to 45% a year, so the ear on the line is usually a new one.
  • When one platform put AI agents on its discarded seller leads, 12% of them came back and converted. Twelve percent of the “no intent” pile had intent. The humans on the line could not hear it.

Five evaluator visits in six end without a car. Who decided those five sellers had intent, and on what evidence?

So how will they increase revenue?

There are three levers. Buy more leads, which means more marketing against a Rs 125 crore line that is already not working. Hire more evaluators, which scales cost in a straight line with volume. Or sort the leads you already have so the evaluator goes to the one in six who will sell.

Only the third lever changes the ratio. Everything else just buys more of the same ratio. So the question is not how to get more calls. It is what data you use to sort the calls you already have.

If your conversion is one in six and your competitor’s is one in four, which of you is still around in five years?

They are looking at the wrong data

Open any contact center dashboard. Connect rate. Average handle time. Disposition code. CSAT. Lately a sentiment score computed from a transcript. Every one of these measures the words or the clock. Not one of them measures the seller.

Three facts make the transcript the wrong place to look.

One. The call is in twelve languages, often two inside one sentence. Speech-to-text for Indic languages is uneven, and code-switched speech is where it is worst. Whatever the transcript gets right is already a translation of a translation, and the model then scores the translation.

Two. Even a perfect transcript discards the signal. Intent is not in the words. It is autonomic. A rise in fundamental frequency is the most consistently reported vocal marker of stress across more than fifty studies, driven by the sympathetic nervous system, not by vocabulary. Hesitant speech shows longer response latencies, more pauses, rising terminal intonation and unstable intensity. None of that survives speech-to-text. “Dekh ke batata hoon” transcribes identically from a man who will sell on Saturday and a man who will never sell.

Three. Sales are seasonal. From Navratri to Diwali the call-back queue swells and the evaluators do not. Every platform sorts that queue by date entered and disposition code. Intent is not a date field. The seller who is three days from a decision and the seller who is three months from one sit in the same column.

If the audio is deleted after the call, with what, exactly, did you measure intent?

You have ten thousand festival call-backs and two hundred evaluators. Which calls do you dial first, and what evidence did you use to pick them?

The contact center that dies, and the one that does not

Khosla is right that the contact center as a mouth will die. A script read aloud by a human in twelve languages is exactly the thing a model now does cheaper. He is wrong that nothing is left. Someone still has to decide which of the six sellers gets the evaluator, and the evidence for that decision is already sitting in the audio the industry throws away.

That decision is deterministic signal processing on the voice itself. Same physics for Marathi, Tamil and Bengali. Calibrated on the population of calls it sees. Running on-premise, with no language model guessing what the seller meant. It will not tell you what the seller said. It will tell you how much of him was in it.

The industry built mouths. We are building the ears.

The seller told you everything in the first forty seconds. You kept the words and deleted the rest. Why?