Super.money's AI Agents: A Blueprint for Indian Builders

Most conversations about AI agents in India are still hypothetical: what an agent could do for your business once you build one. Super.money, the Flipkart-owned payments app, has made it real. Starting in September, it began rolling out AI agents that shop on Flipkart and buy gold automatically on a user's behalf, with CEO Prakash Sikaria telling Bloomberg the rollout would reach its full base of roughly 20 million monthly users within about two months, putting full coverage around mid-November. Whatever you think of letting software spend your money, the design choices behind this launch are worth studying, because they're a clean example of how to let an agent act with real autonomy without losing control of it.

What Super.money actually shipped

Two capabilities, not a general assistant, per reporting from The Paypers and NewsBytes:

• Agentic shopping on Flipkart. The agent can shop for a user on Flipkart's platform, a narrow, single-destination capability rather than an open-ended "buy me anything, anywhere" mandate.
• Automatic gold purchases at a set price. A user names a target price for gold. When the market hits it, the agent executes the purchase without the user approving that specific transaction in the moment.

The roadmap extends this to shopping on other retailers beyond Flipkart, and eventually to bill payments and investments. For now, the agents are free to use. Sikaria has said the company expects to explore subscription-based monetization as adoption grows, and that agentic use cases could eventually make up 30 to 40% of company revenue within three to four years. About a fifth of this year's engineering budget is reportedly going into building these consumer agents, which tells you how seriously Flipkart is betting on this direction rather than treating it as a side feature.

The real design decision: pre-set authority, not open-ended autonomy

It would be easy to describe the gold feature as "the agent decides when to buy gold for you." That's not quite right, and the distinction matters for anyone designing a similar feature. The user sets the trigger condition in advance: a specific price. The agent's job is narrow and mechanical: watch the market and execute the instruction the user already gave, the moment it's true. The agent isn't deciding whether buying gold is a good idea, or picking the price, or choosing the amount. It's executing a pre-authorized instruction against a condition the user defined.

This is a principle worth holding onto for any AI feature you build: humans stay in the loop, and AI prepares or executes work within boundaries a person set, rather than making the judgment call itself. Super.money's gold feature is a working example of what that looks like in a live financial product, not just a governance slide. The agent has real autonomy, it acts without a human clicking "confirm" at the moment of execution, but that autonomy is bounded by a condition the human chose beforehand and can presumably change or cancel.

Why starting narrow matters more than the tech

Super.money didn't launch a general financial AI assistant that can do anything with your money. It launched two specific, bounded jobs: shop on one platform, and execute a pre-set purchase when a price condition is met. That's consistent with the pattern we've pointed to before when covering other agent launches, including Salesforce's lineup of single-purpose named agents: a narrow, well-defined job is something you can actually test against real cases before trusting it with real money or real customers.

If Super.money had launched "an AI agent that manages your finances," the surface area for things to go wrong would be enormous, and so would the trust problem. Two specific, explainable jobs are something a user can reason about: "this agent buys gold for me when it hits my price, nothing else." That's the right bar for any AI feature aimed at a general, non-technical audience: a sentence they can understand and trust.

What this means if you're building or buying a consumer-facing agent

If your business is considering something similar, in fintech, e-commerce, travel or any other consumer category, a few questions are worth answering before you write a line of code:

• What's the one job, and what's explicitly out of scope? Super.money's answer: shop on Flipkart, or execute a pre-set gold purchase. Nothing else. Write your own answer in one sentence before building anything.
• Where does the user set the boundary, and where does the agent act within it? The user picks the price. The agent watches and executes. Keep that split clear in your own design: the human decides the rule, the agent follows it.
• Can the user see and cancel the standing instruction at any time? An agent that can execute a financial transaction without a human confirming it in the moment needs an equally easy way for that human to revoke the instruction before it's a problem, not after.
• What happens when the agent is wrong, or the market moves in an unexpected way? Decide your error-handling and refund or reversal process before launch, not after your first support ticket.
• Who does the agent hand off to when a request falls outside its scope? Super.money's agents handle a narrow set of actions; anything else presumably routes to the regular app or a human, and that handoff needs to be designed, not assumed.

Where the caution has to sit

Letting an agent execute financial transactions, even narrowly scoped ones, raises questions Indian businesses can't skip: consent under the DPDP Act for whatever data the agent uses to act, clear audit trails for every transaction an agent executes without a real-time human click, and a straightforward way for a user to pause or kill the standing instruction. None of that is visible in the press coverage of this launch, which is exactly why it's the first thing to build if you're following this pattern rather than just admiring it.

The bottom line

Super.money's launch is useful less for the specific features, shopping and gold, than for the design pattern underneath them: narrow jobs, pre-set boundaries the user controls, and real autonomy only within those boundaries. That pattern applies whether you're building a consumer fintech feature or an internal agent for your own operations. If you're thinking about where to start building AI agents for your business, our AI Automation & Agents Bootcamp walks through exactly this kind of scoping, from picking the one job an agent should own to defining the boundaries that keep it trustworthy.