IOI
An AI that books your flight, fills the confusing airline form and pays for it while you make chai sounds like the future. On 15 October, it became a promise. Elon Musk said that if Grok’s AI agent makes a mistake with a purchase or leaks your card, SpaceX “will make you whole”. A few weeks earlier in Mumbai, India’s payments authority was asking a harder question: before a machine is allowed to spend your money on UPI, who checks that it is allowed to, and who pays when it gets it wrong?
Those two moments sit at opposite ends of the same shift. Payments are moving from instructions we type to intent we describe. How that shift is governed will decide whether agentic commerce earns trust or burns it.
The promise: “SpaceX will make you whole”
It started with a happy user. An X account shared that Grok had booked a flight for them after they sent it a photo of their credit card, calling the airline’s website “way too confusing”. Musk quoted the post with a line that has since been viewed more than 2.6 million times:
As Business Insider pointed out, that is unusual. Grok, Meta and OpenAI broadly take the same line in their published terms: users are responsible for what their AI agents do on their behalf. None of them had offered an explicit pledge to reimburse mistakes. Musk did not spell out the scope, though. It is not clear whether the promise covers an unauthorised purchase, a wrong booking, fraud by a third-party merchant or knock-on losses like a missed flight.
The risks are not theoretical. The same report describes a founder whose personal-finance agent posted his bank balances into a company Slack channel after confusing two similarly named chats, an agent that ordered 20 pounds of chicken wings during a grocery run, and another that accepted a low offer on a keyboard and shared the seller’s home address with the buyer.
India’s answer: the rails must stay boring
India has a bigger stake in this than almost anyone. UPI processed 24.51 billion transactions worth ₹29.82 lakh crore in August 2026 alone, according to MediaNama. Put AI agents on rails that busy and a small error rate becomes a very large number of unhappy families.
That is why the most important line on agentic payments this year may have come from Ajay Kumar Choudhary, non-executive chairman of the National Payments Corporation of India (NPCI), at the Global Fintech Fest in Mumbai on 10 September:
“Decision making and execution must remain separate. AI may recommend, but authentication and final settlement must follow deterministic auditable rules.”
Ajay Kumar Choudhary, Chairman, NPCI, at Global Fintech Fest 2026
In NPCI’s own framing, AI is adaptive and its output is probabilistic, while payment infrastructure has to stay deterministic, auditable and final. So the architecture splits into three parts. An agent may read your intent. Authorisation checks identity, mandate, limits and consent. Settlement stays rules-based and legally final. “The objective,” Choudhary said, “is not unlimited machine autonomy. It is to be bounded and accountable.”
What NPCI is building, and why it paused
Reuters reported that NPCI is building a registry to verify and monitor AI agents that transact on UPI, as part of a planned Unified Agentic Protocol (UAP). According to sources quoted by BusinessLine, the protocol would build on two features many Indians already use:
- UPI Circle, which lets a primary account holder delegate payment authority to a secondary user, which could be an AI agent.
- Reserve Pay, which lets customers block funds for multiple debits. Banks currently cap such blocks at ₹10,000 for up to 90 days, a limit that may be revisited for agents.
The first use cases would be small and frequent, such as groceries, with spending limits, audit trails, identity checks and a liability framework built in. Then came the brake. Business Standard reported that the UAP launch, planned for the Global Fintech Fest, was put on hold the night before while NPCI works through user protection, what happens when agents go rogue, and governance, before seeking clearance from the Reserve Bank of India. No new launch date has been announced.
Industry is not waiting. Pine Labs has rolled out its own agentic protocol, Amazon Pay launched a consent-based Smart Wallet for agentic UPI payments at the fest, and global players are building their own standards, from Visa’s Trusted Agent Protocol to Mastercard’s Agent Pay and Google’s Agent Payments Protocol (AP2).
Two models of trust
Put the two stories side by side and you get two very different ideas of how to make people comfortable letting software spend their money.
- The guarantee model. Let the agent act freely and have a deep-pocketed company stand behind it. It is fast and it feels generous, but it depends on one company’s goodwill and on terms that have not been written down yet.
- The rails model. Register the agent, cap what it can spend, keep a human-approved mandate and log every step, so trust lives in the system and not in a single promise. It is slower, but it scales to a billion people.
India has seen this movie before. UPI won because the rules were public and interoperable, so any bank, app or kirana store could plug in. Choudhary made the same point about agents: interoperability “must remain non-negotiable”, so agents can work across institutions “without creating new walled gardens”.
What it means for you right now
- Don’t hand an agent your full card. Sending a chatbot a photo of your credit card gives it far more than one purchase. Prefer virtual cards, spending caps or UPI mandates with limits.
- Read the “Allow Always” setting twice. Several of the reported mishaps came from users who thought the agent would still ask before acting.
- Know who is liable. Until a company publishes clear reimbursement terms, assume the risk is yours.
- Builders, design for the rails. Indian startups building shopping, travel or bill-pay agents should plan for registration, mandates and audit trails now, because that is where NPCI and RBI are clearly heading.
The bottom line
Musk’s promise shows how quickly AI agents are moving from demos to daily errands. NPCI’s pause shows that when hundreds of millions of people depend on a payment system, speed can’t come before trust. The agents are coming either way. The real question is whether they arrive on rails that everyone can rely on, and India has a real chance to write that rulebook for the world.
Sources: Elon Musk on X, Business Insider, Reuters, BusinessLine, Business Standard, MediaNama and NPCI. Topic suggested by Anjali. Featured photo: iMin Technology / Pexels.
