Most AI Products Are Built for the Wrong Half of the World

5 MINS

The current AI product wave is largely a San Francisco conversation. Slack-like UI, English-first prompts, $20/month subscriptions, infinite bandwidth assumed. It works for a specific user. That user is not most of the world.

I've spent the last year taking an AI product into India, Africa, and South America. Here's what nobody tells you in the launch posts.

The model is the easy part

Everyone obsesses over which foundation model to use. After a year of shipping, I've concluded the model choice matters far less than:

The first 10 seconds of onboarding — does the user understand what this thing is *for*?

The cost-per-active-user math — at $0.02 per query, with users sending 30 queries a day, you're spending $18 a month on a user paying you nothing yet

The fallback behaviour — what happens on slow networks, low-end devices, and timeouts A weaker model with a tight UX and a strong cost ceiling beats a frontier model with a leaky funnel. Every single time.

English-first is not language-first

The teams I see struggle most are the ones who treat localisation as a translation problem. It's not. It's a mental-model problem.

When a Marathi-speaking farmer types into a chat box, they don't write Marathi the way Google Translate writes Marathi. They write a code-mixed shorthand. They expect short answers. They abandon the session if the first reply takes more than three seconds. None of that is captured in an i18n file.

Voice-first inputs outperform text in most non-English markets

Short responses beat long ones, even when the user "asked for detail"

Latency budget is a product spec , not an infra concern

The retention question is brutal

In emerging markets, novelty retention is real but short. We see users return once a day for the first three days, then drop off cliff-style on day seven if we haven't given them something to come back to.

The fix isn't push notifications. The fix is embedding the AI in a workflow they already do every day. For an agriculture user, that's the morning weather check. For a small-business owner, that's the evening day-end summary. For a student, that's the homework hour.

If your AI product can't ride on a habit that already exists, you are inventing both the user behaviour and the product at the same time. That's two miracles, not one.

What makes me optimistic

The AI products that will matter in five years are not the ones with the best benchmarks today. They're the ones that figured out how to be cheap, fast, and useful in the worst conditions — because those constraints force good design.

A model that works on a 2G connection, in mixed-language, on a Rs 8,000 phone, with sub-second responses, is a model that will work everywhere else by default. Building for the median user of the planet is not charity. It's a competitive moat.

The next 10 million

I've shipped the first 1M users on this product. The next 10M won't come from doing the same thing 10x harder. They'll come from picking a single retention behaviour and being unreasonably good at it.

That's the only AI product strategy that's ever survived contact with a real cohort.

The Growth Loop That Survives a Pivot Is the One Worth Building

Sai skipped presentations and built real AI products.

Sai Gole was part of the March 2026 cohort at Curious PM, alongside 17 other talented participants.