UPI processed 24.51 billion transactions worth ₹29.82 lakh crore in August 2026 alone. That scale is the backdrop to what NPCI (the National Payments Corporation of India, which runs UPI) spent Global Fintech Fest 2026 in Mumbai (9-11 September) talking about: a protocol to let AI agents initiate payments on UPI without a human tapping "approve" on every single one — and, just as pointedly, where NPCI says that shouldn't be allowed to happen.
What NPCI is actually building
The core piece is what's being reported as a Unified Agent Protocol — a registry to identify and authorise AI agents transacting on UPI, so a payment initiated by software has a verifiable operator behind it, the same way a payment initiated by a person has a verifiable account. Earlier reporting (Business Standard, July 2026) had NPCI examining safeguards like per-agent spending limits, explicit scopes, revocation and audit trails; NPCI's own framing at GFF, per Choudhary's remarks, is preserving "interoperability, auditability and settlement finality" as agents enter the picture. The stated first use case is small, repetitive payments — grocery top-ups, recurring bills — not investment decisions or large one-off purchases.
The line: AI can recommend, not authorise
The clearest statement of intent came from Ajay Kumar Choudhary, NPCI's non-executive chairman, at GFF on 10 September: "Decision making and execution must remain separate. AI may recommend, but authentication and final settlement must follow deterministic auditable rules." Read plainly: an AI agent can decide what to buy and draft the payment, but the step that actually moves money is meant to stay a fixed, auditable rule — not a judgement call the AI makes in the moment. That's a direct answer to the failure mode critics had already raised: in the US, agentic-commerce pilots have produced cases of an agent making an unauthorised purchase after being influenced by something it read online. NPCI's position is that the settlement step is exactly where that kind of drift shouldn't be able to reach.
Not hypothetical — Amazon Pay already shipped something
This isn't a whitepaper exercise. Amazon Pay's Smart Wallet, built for agentic UPI payments, launched at the same event — alongside AiNxt, an open-source agentic-AI platform (OS, Code, CLI and Enterprise components) that NPCI itself put out for developers building on this. Whatever the registry ends up looking like in its final form, vendors are already building against the direction NPCI is signalling, not waiting for it to be finalised.
What this means if you're building or buying AI in India
Today, the "Indian payment methods" flag we track on every tool in this catalogue means one thing: you can pay the vendor via UPI or another Indian method. It says nothing about whether the tool itself can autonomously spend on your behalf — that's a separate, much newer capability, and as of this post no tool in our catalogue claims it. If a vendor starts marketing "autonomous UPI payments" as a feature, the useful question isn't "does it work" — demos always work — it's whether the agent is operating inside NPCI's eventual registry with real spending limits and an audit trail, or bypassing UPI's agent identity entirely through a linked card or wallet. The difference is exactly the one Choudhary drew: whether a deterministic, auditable rule is the thing that actually pulls the trigger.
What we don't know yet
NPCI hasn't published a launch date for the Unified Agent Protocol or the agent registry, and the safeguard details (exact spending limits, how revocation works, who's liable if an authorised agent still overspends) weren't settled in what's been reported so far — the July 2026 coverage was explicit that NPCI's own report was "scant on details" at that point. Treat everything above as the direction NPCI has signalled, not a shipped, user-facing feature you can rely on yet — we'll update this post once the protocol actually goes live.