Ancila Didi works in three homes across south Bangalore. She earns about twenty-five thousand rupees a month. She has a smartphone, a prepaid plan she recharges in small amounts, and a WhatsApp account that is, practically speaking, the full extent of the internet that works for her.
I watched her the other morning in my kitchen. She had received a voice note from a woman in the next apartment complex asking if she was available on Tuesdays. Ancila played it once. Played it again. Then held the phone to her ear a third time, lips moving slightly, rehearsing what she would say back. She recorded her reply, listened to it, deleted it, recorded again. The whole exchange took about five minutes for what was, functionally, a yes.
She cannot read recipes, so I have tried teaching her to use YouTube through voice search. It works sometimes. Mostly it doesn't. The results come back in the wrong language, or the interface demands a tap she is not sure about, or the connection drops. She has learned to accept UPI payments, which is real progress. But she is terrified of losing money through her phone. This fear is not irrational. She knows people it has happened to.
Last month, I built a retro-style writing tool and my personal website in about twenty-four hours. I am not an engineer. I described what I wanted to one of the new coding assistants and iterated until it worked. The quality surprised me. The ease of it surprised me more. I keep building things now just to see if the tools will let me, and they keep saying yes.
This is the contradiction I sit inside every day. The same month I build working software by talking to a screen, Ancila is rehearsing a five-minute voice note to say she is free on Tuesday.
The distance between us is not access. She has a phone. She has the internet. She has UPI. The distance is that every system I use was designed for someone who already knows what to ask, how to ask it, and what to do if something goes wrong. Ancila has questions. Real ones. Whether there are better-paying homes in her area. What rate do others charge for the same work? Whether her daughter should learn tailoring or typing. How to open a savings account and connect it to UPI. She does not lack ambition. She lacks a safe, private way to ask.
I read both essays that dominated the conversation these past few weeks. Matt Shumer's, warning that these tools will replace most white-collar jobs within years. Will Manidis's, arguing that most of what we are calling progress is really consumption dressed as output. Both were sharp. Both were worth reading. And both assumed the same user. Literate. English-fluent. Laptop-open. Confident enough to type a question and iterate on the answer. Confident enough to fail twelve times on a prompt and call it learning.
I have been on the other side of this assumption before. Years ago, I was CMO at Babajob, a platform connecting India's blue-collar and gray-collar workers to employment. I spent years learning how people earning less than 250$ a month actually use a phone. What they trust. What scares them. What they need but will never search for because they do not know the right words. Later, I worked alongside Facebook's Free Basics team as they tried to bring the internet to India's next billion. That project failed, and for some good reasons. But the underlying observation was correct: the Western technology establishment was designing for a user who did not exist in most of the world. The diagnosis was right. The prescription was wrong. And the assumptions that sank Free Basics a decade ago are now baked into the foundation of a far more consequential technology.
India is approaching a billion smartphone users. Of those, roughly 300 to 400 million are what you might call constrained: illiterate or semi-literate, operating in languages the major models handle poorly or not at all, sharing devices, topping up data in ten-rupee increments, living in places where connectivity drops mid-sentence. They have been trained, by years of spam calls and phishing links, to distrust anything that asks them to tap. Last year, Indians made over 130 billion transactions through UPI, moving more than two trillion dollars. The infrastructure for digital trust exists here. The interface does not.
And this is not only an Indian problem. Across sub-Saharan Africa, Southeast Asia, and Latin America, the pattern is the same: phones everywhere, functional access nowhere. The GSMA estimates that nearly four billion people globally are unconnected or under-connected, and 88 percent of them already live within range of a mobile broadband network. The barrier is not infrastructure. It is usability, literacy, and trust. India is the proving ground. The market is the world.
The incentive structure inside the major labs is not mysterious. You ship models. You benchmark against other models. You optimize for the user who will pay twenty dollars a month, post about it, build an agent workflow and screenshot the results. That user gives you signal, revenue, and a story for your board. There is nothing wrong with this. It is how technology companies work. But it has produced a specific sorting mechanism. These tools reward fluency. They reward confidence. They reward a willingness to fail and try again. They quietly filter out anyone who hesitates. And hesitation, in most of the world, is not a deficiency. It is rational. If you have two hundred rupees of data left and one wrong tap might subscribe you to something you cannot cancel, you do not experiment. You retreat.
I know the obvious rebuttal. Google Assistant speaks Hindi. Alexa speaks Tamil. ChatGPT has a voice mode. Voice is solved.
It is not solved. What exists is voice as an input layer on top of a visual, text-based system. You speak, and it shows you a screen. It opens a list of results. It routes you into a menu. For Ancila, the screen is the problem. She does not need a better way into the interface. She needs the interface to never require reading, tapping, or navigating at all. What is missing is not voice recognition. It is a fully voice-native interaction layer where the system can answer a question, walk her through a transaction, show her a short video instead of a recipe she cannot read, and confirm before anything costs money. Where the entire exchange can happen without a single tap if she is not sure where to tap.
That is a fundamentally different design surface than adding a microphone icon to a search bar. And it is the reason nobody has built it yet. The metrics the industry optimizes for, benchmarks, English-language performance, developer adoption, twenty-dollar subscriptions, structurally deprioritize this work. The business model does not look like the current one. Ancila is not paying twenty dollars a month. The revenue is transactional, or embedded in services, or subsidized by the employers and platforms that need her in the digital economy. It requires a different kind of imagination. But it is not a small market. It is the largest untapped market in technology.
Here is what I want the people building these systems to consider. Not as a moral argument. As a market one. The next defensible platform will not be the one with the highest benchmark score. It will be the one that earns trust where trust is expensive. Dialect-tolerant, not just multilingual. Designed so that recovery from a mistake is obvious, immediate, and carries no shame. Able to work on a shared device without exposing one user's questions to another. Resilient on a patchy connection. Smart enough to know when it is being used by someone who is scared, and to slow down instead of showing off.
These are hard engineering problems. That is precisely why they are a moat. Anyone can fine-tune a model on legal briefs for Manhattan associates. Building a system that Ancila trusts enough to ask about her wages requires understanding what fear costs, what shame prevents, what silence hides. The company that solves this does not just capture India. It captures every market where users are cautious, connectivity is uneven, literacy is partial, and confidence is earned not assumed. That is most of the planet.
India is not a localization problem. It is a design philosophy. The company that treats it as one will spend the next decade wondering why its product only works for people who already knew how to use it.
Maybe this weekend I will sit down and try to build something. A rough, voice-first assistant in Hindi and Kannada that assumes nothing about literacy and treats trust as the primary design constraint. The foundation models are good enough now that one non-engineer can sketch a prototype in a couple of days. But a prototype is not a platform. The real work is the trust infrastructure, the cultural design, the local data, the distribution into the lives of people who have never typed a search query. That takes serious investment from the people who control the foundation.
What I really hope is that they make it. Billions of people are holding phones full of questions they have no safe way to ask. That is not a feature request. That is a platform waiting to be built.