India Is Building a Registry for the AI Agents That Will Pay on UPI

India is getting ready to let software spend money on people's behalf, and the National Payments Corporation of India is building the thing that decides which software is allowed to.

Reuters reported on 10 September that NPCI is assembling a registry to verify and monitor AI agents transacting on its payment rails, citing three sources involved in the discussions. The registry sits inside a planned Unified Agentic Protocol. Its job is narrow: vet the agents that make payments for a user.

A registry, not a rulebook

Most of the AI governance conversation in India has been about principles. This one is about plumbing, which is a more useful place for it to be.

An agentic payment is a transaction nobody approved in the moment. If an agent reorders groceries when stock runs low, or buys something when a price drops past a threshold, the consent model everyone is used to stops working, because there is no one at the keyboard to authenticate. The question moves from "did the user approve this payment" to "is this agent what it claims to be, and is it allowed to do this". A registry answers the second one by giving every agent an identity the network can check before money moves.

What it covers first

The plan starts small, and deliberately so. The registry will initially vet agents paying over UPI, with cards, bill payments and other methods possible later. Low-value, high-frequency purchases lead, with groceries expected to be among the first applications. More complex instructions come after that, like buying when a discount reaches a set level or investing when a price crosses a threshold.

Reuters reported separately in early September that NPCI was preparing to roll out agentic payments on UPI. So the control layer is arriving next to the capability rather than a year behind it, which is not how this usually goes.

The part nobody has solved

Indexing an agent establishes identity. It does not establish who pays when an agent gets a payment wrong.

One of Reuters' sources said liability for wrong or unauthorised payments will have to be handled through regulation, and regulation moves slower than a directory does. That is worth stating plainly rather than glossing: the identity layer is being built, the accountability layer is not.

Everyone else is doing a version of this

China's Payment & Clearing Association, overseen by the central bank, has already issued guidelines for AI-agent payments covering identification, security and risk controls. Amazon, Google and Microsoft have each built agent registries for agentic commerce.

What differs is where the registry sits. The platforms are registering agents inside their own commerce ecosystems. NPCI would be doing it at the level of a national payment network that already carries most of India's retail digital transactions, which makes it a public rail rather than a private garden.

If you are building agents for the Indian market

Identity is about to become a gating requirement rather than a design preference. An unregistered agent is an agent that cannot transact, and that is a very different constraint to plan around than a governance guideline.

The timing also says something. This is happening while rogue agent behaviour at OpenAI, Meta and Anthropic is under heavy scrutiny. Network operators are not waiting for the technology to settle before deciding agents need papers.

NPCI did not immediately respond to Reuters' request for comment.

Source: India plans AI registry as it looks to roll out agentic payments, sources say (Reuters)