Voice, Language, and the Unfinished Half of AI

Most of the world's business does not happen in text boxes.

It happens in phone calls, voice notes, market stalls, showrooms and service counters — in Thai, Urdu, Arabic, Spanish, Bahasa, and a few hundred other languages. If AI is going to run real businesses, it has to work there, not just in an English chat window.

Why voice is hard, and why that's good news

A voice conversation is unforgiving. Latency is visible. Interruptions are normal. Accents, background noise and half-finished sentences are the default, not the edge case. A system that survives real phone calls has crossed a bar most demos never approach.

Hard problems are good news for builders: the difficulty is the moat.

Language is not a feature checkbox

I've watched software exclude brilliant founders for fifteen years simply because it assumed English. The spreadsheet-famous statistic is that a few languages cover most software; the street-level reality is that most entrepreneurs are outside them.

Multilingual AI changes this only if it's built in from the architecture up — the agent reasons, speaks and writes natively in the customer's language, rather than passing everything through a translation veneer. That's an infrastructure decision, not a settings toggle.

The compounding step

Here's what I find genuinely exciting: a voice agent that handles a business's calls doesn't just save labor. It hears the market — every objection, every request, every price signal — and turns what used to evaporate into proprietary business intelligence.

Text AI conquered the demo. Voice and language are where AI meets the real economy. That's the half I'm building for.