India’s Unified Payment Interface (UPI) processes over 750 million transactions daily, and the NPCI says AI in payments will be key to reaching a billion. Dilip Asbe, MD and CEO of the National Payments Corporation of India, told TechCrunch at Mumbai Tech Week that AI could drive the next half‑billion users while tightening fraud controls.
Key Takeaways
- NPCI aims for > 1 billion daily UPI transactions by 2027.
- AI will power user onboarding, fraud detection, mule identification, and credit extension.
- Voice‑based onboarding is still early, but NPCI sees it as a future differentiator.
- Regulators plan a 30% market‑share cap for UPI apps, slated for Dec 31 2026.
- Small, domain‑specific language models could give Indian fintech a competitive edge.
Historical Context
UPI was built as a government‑backed, interoperable layer that let banks and fintechs talk to each other without a middleman. From day one, the system was designed to be open‑source, to invite competition, and to keep transaction costs low. That openness attracted a flurry of apps that each added a thin user experience on top of the same backend. The result is a market where dozens of players can launch a payment app with a few weeks of development work, yet almost all of them draw on the same transaction ledger.
Because the underlying protocol is shared, any improvement to the core—whether a speed tweak or a new security check—automatically benefits every participant. This architecture also means that a single AI model, once integrated at the network level, can influence fraud detection and credit scoring across the entire ecosystem. The NPCI’s push for AI therefore uses the very design principle that made UPI a success: a common, extensible platform.
AI in Payments: NPCI’s Vision for UPI’s Next Wave
When Asbe talked about the next half‑billion users, he didn’t just throw out buzzwords. He said AI would be used “very effectively” across the whole stack – from onboarding new citizens to spotting fraudulent mule networks. That’s a bold claim, especially because the Indian market already has a dense ecosystem of apps competing for the same users.
What AI Will Actually Do
According to Asbe, AI will help “find fraud, and find mules” and will also “provide credit to all the users and merchants who have digital footprints.” That’s not just hype; it reflects a concrete plan to embed machine‑learning models into the UPI transaction pipeline. The idea is to let the system flag suspicious patterns in real time, then automatically route the case to a human reviewer if needed.
He also highlighted multilingual voice solutions, saying “we must use AI to look at the voice and multilingual solutions to make onboarding simpler.” While voice assistants haven’t taken off yet – NPCI’s 2023 voice‑assistant system still has low adoption – the company believes that a more accurate model could cut onboarding friction dramatically.
From Voice Assistants to Dispute Bots: The AI Toolkit Growing
NPCI’s own FIMI model, launched last year, already serves over 1 million users who need to cancel mandates or resolve disputes. That’s a solid early win for AI‑driven customer service, and it shows how a narrow‑purpose model can scale quickly. Asbe thinks the next step is to broaden that capability into credit scoring and fraud prevention.
“AI will be used very effectively when we look at the next wave of UPI, and that includes all aspects, including reaching new users. We must use AI effectively to protect our current citizens, to find fraud, and to find mules. AI must also be used to provide credit to all the users and merchants who have digital footprints,” Asbe said.
He added that the voice‑assistant system needs “more accurate” models before it can become a “critical component in the payment ecosystem.” That’s a realistic assessment – the Indian market speaks dozens of languages, and current large‑scale models still stumble on regional dialects.
Regulatory Guardrails and the Competition Question
India’s regulator is planning to cap any single UPI app’s market share at 30% starting Dec 31 2026, unless the deadline gets pushed again. Right now, Walmart‑owned PhonePe and Google Pay together control over 80% of the UPI market. Asbe says that low switching costs keep the field open – if a new app can prove a viable business model, investment will follow.
He noted that both PhonePe and Google have poured “millions” into their apps to secure their positions. That’s why the regulator’s cap could shake up the landscape, forcing newer players to innovate faster. The competition could also spur niche solutions, like AI‑driven credit for small merchants that lack traditional banking relationships.
BHIM’s Role in a Competitive Market
In 2024 NPCI spun off the BHIM UPI app to make it more competitive. Even though BHIM’s transaction volume has risen, its market share hovers around 1%. Asbe clarified that NPCI isn’t chasing a specific share for BHIM; instead, the goal is to offer a “sovereign and secure” alternative to the big players.
Why Small Language Models Matter for Indian FinTech
Asbe believes that India’s fintech ecosystem can build “small language models which are sharp, specific, and as deterministic as possible.” The richness of Indian payment data gives local firms a unique advantage – they can train models on transaction histories, merchant categories, and multilingual text that global giants simply don’t have.
He warned that models will differentiate themselves based on the data sets they can access. That’s a call to action for banks, fintechs, and even startups to share anonymized datasets under a regulated framework, so that home‑grown models can compete with imported ones.
- AI‑driven fraud detection could reduce false positives by up to 20% (if models are well‑tuned).
- Voice onboarding could cut new‑user acquisition time from weeks to days.
- Small, deterministic language models could enable faster credit decisions for underserved merchants.
International Context: AI in Finance Beyond India
Across the Atlantic, firms like Coinbase and Robinhood already let AI agents trade on users’ behalf, and OpenAI lets you load personal account data into ChatGPT for financial advice. NPCI showed demos of “agentic commerce” with Razorpay last year, but it hasn’t rolled those capabilities out widely yet. Asbe says India can follow suit, provided there are “strong regulations and a framework” that protect users and mitigate risk.
He stressed that any AI system must be able to audit user consent – if something goes wrong, the system should reference the original instructions and permissions. That’s a sensible safeguard, especially given the rapid pace of AI integration in payment flows.
What This Means For Developers and Builders
For engineers working on payment platforms, the takeaway is clear: AI isn’t a future add‑on, it’s becoming a core infrastructure piece. You’ll need to design pipelines that can ingest transaction streams, apply real‑time anomaly detection, and expose explainable outputs for compliance teams. If you’re building voice interfaces, expect to train models on regional language corpora and to iterate quickly as accuracy improves.
Startups should watch the regulator’s market‑share cap closely. A tighter cap could open slots for specialized apps that focus on credit scoring, AI‑driven dispute resolution, or niche verticals like agritech payments. Those who can prove a sustainable business model will likely attract the “millions” of investment that Asbe says are already flowing to the big players.
Concrete Scenarios for Builders
- Fraud‑first microservice. Create a stateless service that subscribes to the UPI event bus, scores each transaction with a lightweight model, and returns a risk flag. Pair the flag with a confidence score so that downstream compliance can prioritize alerts.
- Voice‑onboarding flow. Deploy a speech‑to‑text engine tuned on Indian dialects, then feed the transcript into a intent‑classifier that extracts user intent, KYC fields, and consent markers. The classifier should be able to reject ambiguous inputs and ask follow‑up questions in the same language.
- Deterministic credit scorer. Build a model that consumes only a narrow set of signals – transaction volume, merchant category, and repayment history – and outputs a binary decision. By limiting feature scope, the model stays explainable and can be audited against regulatory expectations.
Looking Ahead: AI’s Role in India’s Payment Future
India’s push to cross the 1 billion daily transaction threshold will hinge on how quickly AI can be woven into every layer of the UPI ecosystem. If voice models become reliable and small language models prove deterministic, we might see a wave of AI‑powered fintechs that democratize credit and protect users at scale.
Will the regulator’s cap successfully diversify the market, or will the incumbents simply absorb smaller players? The answer will shape how AI is deployed across India’s payments landscape for years to come.
Key Questions Remaining
- How will data‑privacy regulations evolve to allow the sharing of anonymized transaction logs without compromising user trust?
- What governance structures will be needed to audit the decisions of deterministic language models, especially when they influence credit outcomes?
- Can the voice‑assistant pipeline achieve the latency targets required for a smooth onboarding experience, given the diversity of Indian languages?
- Will the 30% market‑share ceiling spur collaboration between smaller fintechs, or will it simply reinforce the dominance of the existing giants?
- How will the industry balance the speed of AI innovation with the need for transparent, explainable models that regulators can inspect?
Sources: TechCrunch, original report
