When AI starts spending your money

Akshara (0:04)

In today's episode, we'll do one deep dive and cover another story for the day. First, we'll talk about when AI starts spending your money, and then we'll talk about how pure your silver really is.

Welcome back to The Daily Brief by Zerodha, where we cut through the noise to help you understand what's actually happening in the most important stories from business and markets. I'm your host Akshara. Today is Wednesday, 9th September. Coming to the first story.

Ask an AI to find you a good pair of shoes today, and it can already do most of the annoying work. It can search across websites, compare prices, and narrow down the options. But eventually, you still have to take over. You open the merchant's app or website, add the shoes to your cart, and approve the payment. Now, companies are already trying to remove that final handover. Razorpay and the National Payments Corporation of India, or NPCIA, have piloted purchases inside ChatGPT and Cloud using UPI tools such as UPI Circle and DeservePay. PineLabs has also built its own protocol on top of UPI, and Mastercard and Visa are building similar systems for card payments, while Google has proposed a broader standard called AP2. Now, these are separate products, pilots, and protocols built by individual companies or partnerships. So India may now try to create a common framework across the UPI ecosystem. Now, Reuters reported that NPCI is working on a Unified Agent Protocol or UAP. This could give banks, merchants, payment apps, and AI agents a shared way to identify an agent, define what it may spend, record what it did, and decide who is responsible when something goes wrong. Until now, software mostly helped move money after you had chosen what to buy. But agentic payments ask you to do both. Make the choice and complete the payment. Before we get to AI, consider what happens during an ordinary payment. The system must know where the money should go and whether you approved sending it, and with cash, both are obvious. If you buy something for 500 rupees and hand the seller a 500 rupee note, the money changes hands because you physically gave it to them. And no bank or payment network needs to separately verify your permission. But digital payments work differently. The instruction to move money and the proof that you approved it are two separate things.

One instruction tells the system who should receive the money and how much, and another check establishes that the account holder approved it. Now, a check does this using your signature. In a card payment, the merchant sends a request through the payment network to the bank that issued your card, and the bank checks the transaction and either approves or rejects it. Internet banking lets you send instructions directly from your bank account. IMPS made these transfers instant and available around the clock. Now, these systems didn't neatly replace one another. Each removed a different piece of friction. And then, UPI compressed much of the process into a few tabs. So, UPI launched in 2016 and gave banks and payment apps a common system. So, instead of entering someone's account number and IFSC code, you could scan a QR code or enter a UPI ID. So, the QR code or UPI ID tells the app who should receive the money. You enter the amount, choose your bank account, and approve the transaction. UPI then carries that instruction between the apps and banks involved. But the bank still needs to establish that you authorize the payment. And that is what the UPI PIN helps with. So, the PIN does not unlock money stored inside your phone. It tells your bank that you approved that particular instruction.

From the user's point of view, a digital payment therefore involves two basic questions. Where should the money go? And did the person controlling the account agree to send it?

UPI made answering those questions remarkably easy. In August 2026 alone, Indians used it for 24.51 billion transactions worth Rs 29.82 lakh crore. But the person generally still had to approve each purchase. But that began to change with mandates.

So UPI AutoPay allows you to approve payments to a known merchant in advance. And there's nothing particularly intelligent about telling software to pay Netflix Rs 499 every month. The merchant is known, the purpose is known, and the role is largely fixed. So the software is just executing your decision and not really making one for you. Now let's change the instruction. Tell the software to keep you subscribed to whichever streaming service offers Formula One at the lowest price. The payment is suddenly the easy part. The software must figure out which service has the rights, compare the plans, and decide which one offers the best deal. And it may even need to cancel one subscription before starting another. Now, large language models or LLMs have made it easier for software to work with the vague way people naturally speak. So you can ask for a decent morning flight to a LL that isn't too expensive without defining exactly what decent morning or too expensive means. But that doesn't mean the model will interpret those words exactly as you intended. It simply means the software can now attempt to work with them. Give the model access to other tools and it can do more. A search tool can pull live flights, a calendar can tell the agent when you need to arrive, a browser can fill in the booking details, add a payment tool and the agent can carry the task through to check out. But it cannot simply take your card or bank account and spend whenever it wants. The payment system needs a way to limit that authority. So card-based payment systems are trying to solve this using tokens and records of user consent. Suppose you ask an agent to book a flight to Delhi for no more than 10,000 rupees. You may first authorize it to use a particular card within defined limits, and the payment provider doesn't need to give the agent your actual card number. It can create a separate digital credential linked to that card, and this substitute credential is called a token. Now, the token may be restricted to a particular merchant, amount, or period, and the agent can use it after finding a flight that fits your instructions, while the payment system checks whether the transaction falls within the authority you granted. Now, the token protects the card details, but other parts of the system still need to record what you ask the agent to do.