NPCI's FIMI Dispute Model Now Serves One Million Users. The Broader AI Rollout Is Still Waiting.

Since Asbe outlined NPCI's AI ambitions at Mumbai Tech Week 2026 last month, one concrete deployment figure has come into sharper focus: FIMI, the small language model NPCI launched to handle user disputes, is now serving over one million users, according to TechCrunch's interview with Asbe. The system lets users cancel payment mandates and resolve transaction issues without human agents, and Asbe said it is scaling fast.

It is also a narrow use case. Dispute resolution is not the broader vision of AI-driven credit, fraud detection, and new-user onboarding that Asbe described as the next wave.

What Is Built vs. What Is Pitched

Asbe told TechCrunch that AI will be used to reach the next half-billion UPI users, detect fraud, identify money mules, extend credit to merchants and users with digital footprints, and power voice and multilingual onboarding. Those are goals, not deliverables. NPCI demonstrated agentic commerce capabilities with Razorpay last year, but Asbe confirmed there has been no wider rollout of those features as of this interview.

Voice payments are in a similar holding pattern. NPCI launched a voice assistant-based interactive payment system in 2023. Three years later, Asbe acknowledged adoption has NOT taken off. His explanation: voice models need to be more accurate before the use case becomes viable at scale. Voice remains stuck in early-days territory, despite being one of the most-discussed AI interfaces for reaching India's less digitally literate population.

The Small Language Model Argument

Asbe made a pointed claim about where India's competitive advantage in AI finance could come from. Rather than competing on general-purpose large language models dominated by U.S. and Chinese players, he argued Indian banks, fintechs, and payment networks should build small language models trained on India's payments data — models that are, in his words, "sharp, specific, and as deterministic as possible."

His reasoning is sound on its face. NPCI sits on transaction data from 750 million daily UPI payments across hundreds of millions of users. That data set, if properly governed and made accessible to Indian financial institutions, would give domestically built models an edge in fraud detection and credit underwriting that no foreign general-purpose model can easily replicate.

The counterargument: financial data of that scale and sensitivity raises real questions about privacy, consent, and who controls the models trained on it. Asbe acknowledged that for AI-powered finance to work, users need robust protections and a clear record of what instructions and consent they gave to any AI agent acting on their behalf. That framework does NOT yet exist in deployable form. He described it as a requirement for adoption, not a current reality.

Market Concentration Is the Unresolved Problem AI Doesn't Fix

One structural issue sits underneath all of this. PhonePe, owned by Walmart, and Google Pay together hold over 80% of UPI transaction volume, according to TechCrunch's reporting. NPCI has long said it wants healthy competition between UPI apps. It has not achieved it.

AI-driven onboarding and voice interfaces are theoretically tools to bring in the next wave of users who are currently unserved. But if those users arrive and continue routing through PhonePe and Google Pay, the concentration problem deepens rather than resolves. Asbe did not address how NPCI intends to use AI-era growth to shift that market share distribution.

Where the U.S. Comparison Cuts Both Ways

In the U.S., Coinbase and Robinhood have already rolled out AI agents that execute trades on users' behalf, and OpenAI has enabled users to load personal financial data into ChatGPT for advice, according to TechCrunch. India's regulatory posture is more cautious. Asbe explicitly tied AI-powered agentic finance to the existence of sufficient regulatory safeguards first.

That caution has a cost: slower deployment. It also has a benefit: fewer users burned by AI agents making bad financial decisions with no accountability framework in place. Whether NPCI and India's central bank can move fast enough to close the gap while keeping that protection layer intact remains an open question.

FIMI's one-million-user milestone gives NPCI a proof of concept. Whether it becomes the template for the broader AI buildout, or a footnote while voice and agentic payments continue to wait, depends on regulatory framework development that, as of June 28, 2026, has no announced timeline.