NPCI chairman Ajay Kumar Choudhary confirmed at Global Fintech Fest in Mumbai on 10 September 2026 that the organisation is examining protocols to identify and authorise digital agents inside UPI, preserving interoperability, auditability and settlement finality (MediaNama). The design constraint matters more than the announcement: an AI agent may recommend a payment, but authentication and final settlement must follow deterministic, auditable rules. This lands on rails that processed a record 24.51 billion transactions worth Rs 29.82 lakh crore in August 2026 alone (PTI via The Hindu BusinessLine).
TL;DR — Last verified: September 13, 2026
- NPCI is working on protocols to register, identify and authorise AI agents on UPI; Reuters reported the framework is called the Unified Agent Protocol (UAP).
- The architecture separates intent, authorisation and settlement. An agent can read intent; identity, mandate, limits and consent checks stay with the authoriser.
- First use cases are deliberately low-value and repetitive: groceries and subscription renewals, then conditional purchases such as buying when a price falls below a set level.
- An agent registry is planned so networks can verify and monitor which agents are transacting.
- Liability, per-transaction caps and whether any developer can register an agent are all still open. No UPI-specific AI framework exists in rule today.
- Amazon Pay, Pine Labs and Visa shipped private agentic products at GFF 2026 - none of them is the NPCI protocol.
What are the new UPI rules for AI agents?
There is no published rulebook yet - what exists is a stated architecture. Choudhary said decision making and execution must remain separate, and that AI may recommend while authentication and final settlement follow deterministic auditable rules (MediaNama). In practice an agent never becomes the payer - it is a requester whose request carries a verifiable identity and mandate, and the bank or PSP still runs the checks.
He also drew two lines around it. Interoperability, he said, must remain non-negotiable, and he warned that dependence on a single national or commercial stack could over time reduce sovereign choices.
What has NPCI actually built so far?
Reuters, citing three people familiar with the plans, reported on 1 September 2026 that NPCI built a framework called the Unified Agent Protocol letting AI agents execute UPI payments without a human approving each one, with a launch expected around Global Fintech Fest (Reuters, via syndication). Business Standard had reported similar work in July 2026. The GFF keynote was the first time an NPCI board member described it publicly.
A registry sits alongside the protocol: NPCI is working on a mechanism to verify and monitor agents that pay via UPI, expected to form part of the planned agentic protocol (New Indian Express). Card networks converge on the same idea from the other side: Visa, Mastercard and Ant International have begun a shared Know-Your-Agent framework to identify agents across networks, wallets and marketplaces.
NPCI does not have to invent the plumbing. Two existing UPI features map onto delegated spending, as Reuters noted: UPI Circle (a primary holder delegates payment authority to a secondary party, including an agent) and Reserve Pay (funds blocked for multiple debits, capped at about Rs 10,000 for up to 90 days (Reuters, via Yahoo Finance). An agent spending inside a pre-blocked, pre-limited pot is a smaller change than one with open access to a bank account.
Which payments will agents be allowed to make first?
| Stage | Example | Why first |
|---|---|---|
| Low-value repeats | Grocery top-ups, subscription renewals | Small amounts, predictable merchants, easy to reverse |
| Rule-bound purchases | Buy when a price crosses a user-set threshold | Requires the user to define limits up front |
| Broader autonomy | Wider merchant categories and higher values | Needs liability rules and audit tooling first |
Under the model being discussed, the user sets rules and limits once, and qualifying payments then proceed without per-transaction approval (New Indian Express). If you already run assistants that book or buy on your behalf, the shift is that your mandate becomes a machine-readable object rather than a habit. Our comparison of personal AI agents in 2026 covers which tools are close to that level of task execution, and the A2A protocol and shared-memory guide explains how agents pass state between each other, which is the layer an audit trail has to capture.
What launched at Global Fintech Fest 2026, and what did not?
Several agentic payment products went live in Mumbai, and none of them is UAP. Amazon Pay showed a consent-based Smart Wallet for agentic UPI payments. Pine Labs announced P3P, where an agent completes a UPI payment after a single upfront authorisation, deployed with L&T Finance on the PLANET app for flight booking inside user-defined spending limits (MediaNama). Visa ran live agentic commerce demonstrations with Federal Bank, IndusInd, RBL, Razorpay, Paytm Payments Services, Cashfree and Pine Labs, built on tokenisation and its Payment Passkey (The Hindu BusinessLine).
The distinction matters: a private wallet-level mandate binds you to one provider; a network protocol with a registry makes an agent portable across apps.
What is still undecided?
- Liability when an agent acts within its mandate but against the user's interest. Choudhary framed the challenge as keeping autonomy bounded, accountable and auditable, and said a liability mechanism is needed as responsibility shifts between users, agents, banks and merchants (Economic Times).
- Whether agent registration opens to any developer or only to regulated entities.
- Actual per-transaction and cumulative limits, which have not been published.
- Regulatory status. No framework specific to AI agents governs UPI yet. CERT-In has proposed human-in-the-loop controls above defined financial thresholds in its Digital Threat Report 2025-26, and the RBI FREE-AI committee's recommendations are not binding.
Identity assurance is the open technical question: delegated authority only works if the mandate cannot be forged or reused - the same evidentiary problem as our piece on digital signature misuse in India.
We measured demand for this vocabulary ourselves (n=5 queries per geography, two geographies): — "npci ai agents upi", "upi ai payments", "agentic payments india", "ai agents upi" and "npci unified agent protocol" — DataForSEO returned zero matching keywords in both the India and US databases (measured 13 September 2026). The agent-payments vocabulary has no measured search demand yet in either market - readers search upi new rules, not the jargon. For pricing AI tooling for Indian users, our Claude India pricing and payments guide is the more practical read.
FAQ
Q: Can AI agents pay on UPI right now?
A: Not under a common NPCI protocol. Amazon Pay's Smart Wallet and Pine Labs' P3P are private implementations, not a UPI-wide rule.
Q: Will I have to approve every payment my agent makes?
A: Under the model reported so far, no: set rules and limits once, and qualifying payments proceed without per-transaction approval.
Q: What is the Unified Agent Protocol?
A: The name Reuters reported for NPCI's framework to register, identify and authorise AI agents on UPI, with a registry so agents can be verified. NPCI has described the design intent but not published the specification.
Q: Who is liable if an agent spends wrongly?
A: Undecided. NPCI's chairman says a liability mechanism is needed as responsibility shifts between users, agents, banks and merchants.
Q: How large is UPI, for context?
A: 24.51 billion transactions worth Rs 29.82 lakh crore in August 2026, up 22% in volume and 20% in value YoY, per NPCI data reported by PTI (PIB).
Corrections log
No corrections yet. If a figure or attribution here is wrong, we will note the change and the date in this section.
This article was produced with AI assistance and reviewed against primary sources before publication. Every figure is linked to its source. See how we work for our verification and disclosure process.
