UPI's AI bet: why payment data beats fancy models
India's payments chief says AI will drive the next UPI wave. The real lesson for Sri Lankan builders isn't the AI — it's who owns the data and the business model.
AI in digital payments is about to get a lot of attention, and I want to read past the headline before anyone in Colombo copies the wrong lesson. Dilip Asbe, MD and CEO of India's NPCI (the body that runs UPI), told TechCrunch that AI will be heavily involved in the next era of payment growth. You can read the original report here: TechCrunch — Indian payments chief on AI and UPI.
Here's my take: the interesting part isn't that a payments boss likes AI. Everyone says that. The interesting part is where he points it, and the quiet admission that the thing actually blocking competition is a business model, not a model checkpoint.
📊 The numbers that frame the whole story
Before the AI talk, look at the scale Asbe is working with. These are the figures from the report:
| Metric | Figure |
|---|---|
| Current UPI volume | ~750 million transactions/day |
| Stated target | 1 billion+ transactions/day |
| PhonePe + Google Pay combined share | over 80% |
| BHIM UPI share (spun off in 2024) | ~1% |
| Market-share cap deadline | 30% limit by 31 December 2026 |
Two apps own four-fifths of the market. A government-backed app sits at roughly one percent. A regulatory cap is supposed to force that concentration down by the end of 2026, and so far the market hasn't moved much on its own.
Key takeaway: UPI's problem isn't technology. It's that two players dominate and no one has found a way to make a competing app pay for itself. AI is being recruited to fix a business problem.
🔍 Where Asbe actually wants to point AI
Strip out the buzz and the AI use cases he named are specific and, honestly, sensible:
- Fraud and mule detection — finding the accounts that launder scam money.
- Credit distribution — extending small loans to users who have a digital payment footprint but no formal credit history.
- User acquisition — getting more people onto the rails.
- Voice and multilingual onboarding — letting people who can't type English navigate a payment flow by speaking.
None of that is science fiction. NPCI already runs a model called FIMI for resolving user disputes, which the report says serves over a million users for things like cancelling recurring-payment mandates. That's AI doing boring, high-volume customer-support work, which is exactly where it earns its keep.
The voice angle is the honest one. NPCI launched a voice assistant back in 2023 and, by Asbe's own account, adoption hasn't really happened yet. Shipping the feature and getting people to use it are two different battles.
💡 The line every Sri Lankan builder should underline
Buried in the AI talk is the comment I think matters most for anyone building here. Asbe suggested Indian firms can build small, specialised language models trained on India's own rich payment datasets, rather than reaching for a giant general-purpose model.
That reframes the whole "we can't compete on AI" worry:
| The assumption | The reality Asbe points to |
|---|---|
| You need the biggest model to win | A small model tuned on the right data often wins the narrow task |
| The moat is the AI | The moat is the proprietary dataset feeding it |
| Only billion-dollar labs can play | Anyone holding domain-specific data can train something useful |
For a small Sri Lankan team, this is the whole game. You will never out-train a frontier lab. But a fraud model that understands local transaction patterns, or a support bot that handles Sinhala and Tamil payment queries, is something a foreign giant has zero incentive to build well. The data is the asset. The model is just the part you can rent.
🌐 The commercial-model trap, and why it's local too
The other admission in the report: low switching costs mean users could move between apps easily, but new entrants don't invest because there's no clear way to make money. Asbe's framing is that the moment a viable commercial model appears for the ecosystem, newer players will pour money in.
I see the same pattern in Sri Lanka's own QR and instant-payment push. The rails exist. Getting people to tap them is one problem; building an app on top that can actually fund itself is a harder one. A free, well-built payment experience with no obvious revenue is a charity, not a business, and charities run out of runway.
If you're a builder here, the takeaway is to design the money question in from day one:
- Pick a narrow, painful job — reconciliation for small shops, freelancer invoicing, school-fee collection.
- Find the data only you can collect by doing that job well.
- Then decide where AI removes a real cost (support, fraud checks, categorisation).
- Charge for the outcome, not the AI.
If you're already working out the money side, our USD–LKR freelancer earnings calculator and currency converter are the kind of small, sharp, locally-useful tools that win on relevance instead of scale. That's the same bet, smaller.
🚀 What this means for you
If you took one thing from Asbe's comments, don't let it be "add AI to everything." Let it be this:
- AI in payments is mostly plumbing — fraud detection, dispute resolution, multilingual onboarding. Unsexy, high-volume, real.
- The competitive edge is data, not the model. A small model on the right local dataset beats a giant model on generic data.
- A feature without a business model is a hobby. Even India's UPI, at 750 million transactions a day, is stuck on this question.
- Local beats large for narrow jobs. Sinhala and Tamil support, Sri Lankan fraud patterns, local fee flows — no foreign giant will build these for you.
Bottom line: The headline says AI will drive the next era of digital payments. The story underneath says data and a working business model will. Build for the second one and the first one becomes a tool you reach for, not a strategy you hope works.