Between the noise of frontier model releases and the ITR filing rush, a quieter Indian move this week deserves more attention. According to a report in Business Standard, NPCI is working on a Unified Agent Protocol so that AI agents can, with the user's permission, make payments over UPI. The plumbing is being drawn now, in consultation with the industry, and it will need RBI clearance before it goes live.
This is not a small feature. It changes who the "user" of a national payments rail is. For eighteen years of digital India, the working assumption has been one human, one device, one intention per transaction. UAP contemplates a different assumption: a piece of software acting on your behalf, initiating value transfers between banks, at speed, at scale, and often when you are asleep.
The rail, not the app, is the state's job
The instinct in newsroom coverage is to ask which AI assistant will pay first, which quick-commerce platform will move first, which bank will onboard first. That framing misses the point. The private sector will produce the agents. But agents cannot talk to a payments system unless someone builds a shared, trusted way of introducing them, verifying them, and revoking them when they misbehave. That is a public infrastructure question, not a product question.
The same reporting notes that Visa is building a Trusted Agent Protocol, Google has an Agent Payments Protocol, OpenAI has an Agentic Commerce Protocol, and Pine Labs has P3P. Each of these is a private schema competing for adoption. India's answer is different in shape: a common protocol, sitting above private agents but below the bank rails, run by a body that already coordinates the industry. That shape difference matters more than the technical details.
Why a state adjacent register is the right answer
An AI agent that moves money on your behalf is, in effect, a narrow private power of attorney executed at machine speed. The three questions the law has always asked about such powers are unchanged. Is the delegate real? What is the scope? What happens when scope is exceeded? UAP appears to want answers to all three: a registry, an authorisation envelope with spending limits, and audit logs that allow a payment to be reconstructed after the fact. None of these can be credibly provided by any one of the competing private schemas, because their commercial incentives run in the other direction.
What this means for tax and compliance
From inside a large tax administration, the second order effects here are more interesting than the first order ones. Consider three.
Attribution of transactions
Every agentic payment is a transaction the user initiated in intention but not in execution. Tax law has thin machinery for that distinction. When your agent buys a cross border subscription, or repeatedly tops up a wallet, the audit trail must show that the human principal, not the agent, is the taxable person. A registry of agents, plus a log of who authorised which agent to spend how much within what window, is exactly the primary evidence a revenue officer will one day want. UAP, incidentally, is building that evidence layer.
Fraud patterns will migrate
Every payments innovation is followed by a fraud innovation. UPI's own history is proof. When agents start executing recurring low value purchases, the fraud will not be brute impersonation of the human. It will be quiet capture of the agent, either by compromising the credential or by prompt level manipulation of what the agent decides to buy. Rules that ask only whether the payment was authorised will not catch this. The real question becomes whether the decision the agent took was one the user would have taken. That is a new class of dispute, and the consumer protection frameworks around UPI will have to grow into it.
The invoice, the payment log, the return
If a routine grocery order is executed by an agent that also holds the user's GSTIN and preferences, the natural next step is that the invoice, the payment log, and the return pre-fill start speaking to each other automatically. Tax administrations everywhere have been talking about pre-filled returns for a decade. Agentic commerce is what finally forces pre-fill to become the default rather than the exception.
What a tax department should be doing this quarter
Two things, quietly.
First, seat someone in the room during UAP design. Not to slow it down, but to make sure the log schema captures the fields a revenue authority will one day need: agent identity, principal identity, spend envelope, timestamp, merchant category, and the human confirmation trace. Retrofitting those fields later is always more expensive than agreeing them now.
Second, begin scenario work on what pre-filled returns look like when the underlying spend is agent initiated. Category assignment, personal versus business use, and the taxpayer's ability to challenge an entry generated by software the taxpayer barely understands. These are not futuristic problems. If UAP moves at UPI speed, they arrive within three assessment cycles.
India tends to build payment rails first and think about their tax and legal downstream later. UPI is itself the case study. There is a narrow, useful window here to do it the other way round.
#UPI #NPCI #AgenticAI #DigitalPayments #IndiaAI #PublicInfrastructure #TaxAdmin #UAP
No comments:
Post a Comment