The next billion AI users will speak

After watching the debates at the AI Impact Summit, one thing stood out to me.

Most of the world is building English-first AI. India might be building something far more powerful.

The hidden assumption

Most large language models today assume a simple interface: typing, in English.

That's how developers interact with AI. That's how benchmarks are measured. That's how most of the internet was written. So it feels natural that the next wave of AI would be built the same way.

But there is a hidden assumption underneath: that humans interact with machines mainly by typing.

That assumption works in Silicon Valley. It doesn't work in India.

• Hundreds of millions of people are not comfortable typing in English.
• Many prefer speaking to typing, even when they can type.
• Dozens of languages coexist in daily life, often in the same sentence.
• Digital adoption here is mobile-first, and increasingly voice-first.

Which means the interface of AI in India may not be a keyboard at all. It may simply be conversation.

A different race

Today's AI race is largely about bigger models, better benchmarks and faster inference. All of it matters. None of it is where India's biggest opportunity lies.

India's opportunity may lie in voice-first, deeply localised AI, designed for how people actually communicate. A farmer asking about crop prices in Telugu. A shopkeeper reconciling accounts in Hindi and English, mixed. A grandmother booking a train ticket by simply saying where she wants to go.

The next billion AI users will not type prompts. They will speak.

This is not a smaller version of the same product. It is a different product, with different problems to solve.

Why voice is harder than it looks

From the outside, voice AI looks like a solved problem. Speech-to-text, a language model, text-to-speech, done. In practice, each step breaks down in the conditions most Indians actually live in.

• Code-mixing. People switch between Hindi, English and a regional language mid-sentence, sometimes mid-word. Models trained on clean, single-language text struggle with how people really speak.
• Accents and dialects. A single language can sound very different from one district to the next. Accuracy that looks fine on a benchmark collapses in a village a few hundred kilometres away.
• Noise. Markets, traffic, crowded homes. Real speech rarely happens in a quiet room with a good microphone.
• Low-cost devices and patchy networks. Heavy models that need a fast connection and a premium phone will not reach the people who need them most.
• Trust without text. When a user cannot easily read what the machine wrote back, the system has to earn trust through how it speaks, confirms and corrects itself.

Solving these is hard. Which is exactly why it is defensible. A company that cracks them for one language and one use case has built something a general-purpose model cannot easily copy.

Where the value will be created

The interesting business models will not come from voice as a feature. They will come from voice as the front door to services people already need.

• Commerce. Ordering, paying and tracking by speaking, for customers who have never used a shopping app comfortably.
• Agriculture. Advice on weather, prices, pests and government schemes, in the farmer's own dialect, at the moment it is needed.
• Financial services. Explaining a loan, an insurance policy or a savings plan clearly enough that people understand what they are signing.
• Public services. Applying for documents, checking entitlements and raising complaints without needing an intermediary.
• Small business. Bookkeeping, inventory and customer messages handled through conversation rather than forms.

Each of these markets is large. Each has been held back by the same barrier: the interface assumed literacy, language and confidence that most users did not have.

Why it matters beyond India

What gets built for India rarely stays in India. Payments are the obvious example. Systems designed for a billion people with very different levels of literacy, connectivity and income turned out to be remarkably exportable.

Voice-first AI could follow the same path. Most of the world does not live in English, and does not want to type. If India gets this right, the rest of the Global South will be watching closely, and so will everyone selling to it.

There is also a lesson for global companies. Designing for the hardest conditions often produces better products for everyone. A voice assistant that works in a noisy market in Hyderabad will work very well in a quiet kitchen in Munich.

From participant to pace-setter

If this works, India may not just participate in the AI race. It could define an entirely new frontier of it.

The models will keep getting bigger elsewhere. The interface, and the business models built on it, could be invented here. That requires investment in speech data for Indian languages, open infrastructure that startups can build on, and founders willing to solve unglamorous problems for users who have long been an afterthought.

The prize is worth it. The next billion AI users are waiting to be spoken to.