Editor's Note: This blog is adapted from a session by Roopasree Ranganna, SVP, Tech and Service Delivery Leader at Synchrony, at thegeekconf mini 2026. Drawing on her work in the financial sector, Roopasree looks at what happens when AI agents move from recommending products to making purchases. She explains how agentic commerce changes the roles of buyers, agents, and sellers; why identity, mandate, and recourse matter; how payment protocols are evolving; where India is taking a different approach; and why trust and the ability to undo a transaction need to be built into agentic commerce from the start.
Commerce Has Had Two Parties. Now There Is a Third.
For centuries, commerce had two parties: a buyer and a seller. Payment infrastructure was then added to facilitate transactions between buyers and sellers. When something goes wrong, there are processes for recovering the money. Now there is a third party: the agent.
We are outsourcing parts of shopping to agents. The commerce flow now introduces another actor: the agent, which can act on the buyer’s behalf within defined permissions. An agent acts on the data and rules available to it. That changes how shopping works. Most commerce systems were built for a human interface; that is the change we are going to see.
Watch the full talk, straight from thegeekconf mini 2026:
What Is Agentic Commerce?
Agentic commerce is delegated authority from a person or business to an agent to select goods and complete payment on their behalf without a human approving each transaction. Today, we control our money. We want to know what we spend, what we buy, and what our bank balance looks like. With agents, the goal is to delegate routine orders.
Think about the orders you place through Swiggy, Zomato, or Zepto. Some of these are routine. You can build agents for those tasks with guardrails.
You set the intent: what do you want to buy?
You set the budget: how much can the agent spend?
The agent is then trusted to act within those boundaries.
Think of the agent as a personal assistant for shopping. That is autonomous execution. The click is no longer a human click. The agent operates within the identity, authorization, and payment controls defined for the transaction. The execution can increasingly be handled by software, with the agent acting within the authority delegated by the user.
Every Agent Needs Identity, Mandate, and Recourse
For this to work, the agent needs to know three things. First, on whose behalf is it acting? If you have multiple agents, each one needs to know who it represents. Second, what is its mandate? That includes the rules and boundaries under which it can act. Third, what happens when you need to revoke the authority or correct what the agent did? That is recourse.
Identity, mandate, and recourse are three important parts of agentic commerce.
The merchant needs to establish who the agent is acting for and what authority has been delegated to it. The legal responsibilities of the user, agent, merchant, and other parties depend on the applicable laws and contractual arrangements. A related concept now being discussed is “Know Your Agent,” which could require agents to have verifiable identities as they participate in regulated transactions. We provide identity documents and information to banks. Agents may need identities of their own as they enter the regulatory space.
AI Tools Are Becoming the Storefront
We use generative AI tools for many things today, and those tools are changing commerce. ChatGPT, Perplexity, Gemini, and Microsoft Copilot are increasingly becoming interfaces for product discovery and, in some cases, purchasing. In the U.S., some AI shopping experiences already support purchases directly within the chat interface. With OpenAI, when you research or compare products, you can get a buy option within the experience. Perplexity works in a similar way. If you search for an answer about a product, the answer can become the storefront.
Google has developed the Universal Commerce Protocol (UCP), which supports agentic shopping workflows and access to product information. Microsoft is doing the same. The face of the storefront is changing. You are not leaving these AI tools to shop somewhere else. You are shopping inside them. That is changing the rules of commerce.
Checkout Is Becoming a Protocol
Some examples are already in play in the US. OpenAI and Stripe allow a user to research a product in ChatGPT and get an option to buy. Perplexity and PayPal have the same type of instant-buy model. Google and Shopify have developed the Universal Commerce Protocol, which enables Shopify merchants to support shopping within the Google Gemini framework. Microsoft Copilot is part of this shift as well.
Checkout is increasingly being enabled through protocols rather than being limited to a traditional checkout page, and the discovery is taking place within the chat.
In many markets, governance, liability, and consumer-protection frameworks are still evolving alongside the technology. A few years ago, shopping inside a chat would have sounded unusual. Now the AI tools are moving faster than the governance around them. Consider a simple example. You ask ChatGPT to order coffee for your office. It can retrieve product information, help the user select a product, initiate checkout, and pass the order to the merchant for payment and fulfillment.
The whole process happens within the interface. You do not leave the chat.
Product Discovery Will Depend on Data
Another change is happening in how products get discovered. Product discovery has traditionally involved advertising. Traditionally, product discovery has been influenced by search ranking, product relevance, advertising, and other signals. Agentic commerce changes that. Agents look at product data, attributes, discoverability, and metadata. If a merchant makes a mistake in a product attribute, agents can rely heavily on structured product data, attributes, metadata, price, availability, and other machine-readable signals when discovering and comparing products.
That makes structured, machine-readable data important.
Merchants need to make product data readable by machines so agents can discover products.
Real price and real stock also matter. Payment tokens can allow payment credentials to remain protected while the merchant receives the information required to process the transaction. Payment credentials can remain with the payment provider or wallet, while a token or authorized payment instruction is passed through the commerce flow. The merchant remains the book of record for the purchase. The shopper does not have to leave the chat window.
Agents Are Becoming Part of the Shopping Experience
People are changing how they discover products. Catalogs and traditional UIs will continue to have their place, but product suggestions and recommendations are increasingly happening through generative AI tools. For some types of shopping, these tools could become the starting point instead of a browser.
That is how the shopping experience is expected to develop going ahead.
Working in the financial sector means seeing these changes in payment and shopping behavior as part of the daily conversation. The expectation is that by 2030, the majority of transactions could happen through agentic commerce. The speaker also cited a figure that 60% of people want agents that can handle purchases on their behalf.
The reason is straightforward. You set a limit and give the agent a mandate. The agent then takes over the work you do not want to do yourself.
Examples of Protocols and Standards Emerging Around Agentic Commerce
There are several protocols coming together to make this work.
This is an example stack rather than a complete list. The pieces need to be connected to support the payment journey.
MCP provides a standardized way for AI applications to connect with external tools and data sources.
Then there is the Universal Commerce Protocol from Google.
There are protocols for different parts of the journey:
- Discovery
- Identity
- Authorization
- Agent payments
- Checkout
- Settlement
In India, UPI provides a major real-time payment rail that could support agentic commerce.
There is some overlap between these protocols, but together they address different parts of agentic commerce.
The same applies to regular software engineering. You need to know who is entering the system, how they are authorized, whether they can be trusted, and how cryptography and payments work.
Agentic commerce needs the same foundation, with these protocols working together.
India Already Has the Payment Rails
India has a different starting point. In June 2026 alone, UPI processed 22.71 billion transactions worth ₹28.92 trillion.
The speaker also cited a survey showing that 64% of Indian consumers want an AI agent. The use cases do not have to be complex.
You could train an agent to read a vendor contract and tell you when it expires, what the contract terms are, and what it means to exit. You could also set a mandate for the agent to order something for your house every Tuesday.
Set the payment limits. Set the payment gateway. The agent handles the task. The point is to remove routine work.
India Is Extending UPI Instead of Starting From Scratch
India is building a unified agent protocol on UPI. UPI already has a delegation model through UPI Circle. For example, a parent can delegate payment access to a child under 18 while keeping control over the account.
That delegation layer already exists. India can extend the payment network and its existing rails rather than build everything from scratch. The U.S. ecosystem is developing agentic commerce through a mix of card networks, payment providers, merchants, and private-sector protocols.
The speaker described a potential Indian model built around public digital infrastructure such as the ONDC model.
If 60% of people begin creating agents for autonomous purchasing, the impact could extend beyond individual shopping and into the wider economy. UPI already moves ₹28.9 trillion. Wider adoption could change how a portion of existing UPI commerce is initiated and executed.
India and the West Are Taking Different Routes
There is another difference. In India, RBI regulates the banking and payment systems that would underpin agentic commerce, while other aspects of commerce remain subject to their respective regulatory frameworks. In the U.S. and Europe, regulatory responsibility is distributed across different authorities and legal frameworks, depending on the payment, consumer, data, and commerce activity involved. There is no single referee overseeing the entire process. That makes agentic commerce more complex to build across those markets. India already has substantial payment infrastructure that could provide a foundation for agentic commerce.
How Could Agentic Commerce Work in Practice?
At a high level, the Western model involves the user, agent, merchant, payment gateways, and card networks, with different protocols supporting the transaction.
India is looking at a model involving Beckn/ONDC, agent authorization, delegation, and settlement.
UPI is already in play.
The expectation is that users could get an option within a UPI app to create agents because the infrastructure is already available.
The same pattern exists in tools such as Microsoft Copilot and OpenAI, where users can create agents for specific tasks within existing frameworks.
Merchants already participating in relevant open networks can expose their catalog and inventory information through machine-readable interfaces.
Depending on the mandate and implementation, the agent could complete the transaction within predefined limits, with or without additional human confirmation.
That is the direction agentic commerce in India is taking. Work is already underway, and these options could start appearing within about a year.
Trust Is the Missing Layer
Protocols are being developed. Payment infrastructure exists. Merchants are experimenting. But one problem remains: trust. How many people trust an agent to shop for them? When UPI first arrived, many people did not trust it either. The trust layer is still being built.
Companies moved quickly to build models, agents, and protocols because they did not want to miss the opportunity. In that rush, building trust with the people who would use these systems did not receive the same attention.
That is the problem that still needs to be solved. The speaker cited several figures from the market. PayPal found that nearly half of merchants were actively tracking visits or transactions from AI agents, yet only about one in five had at least 80% of their product catalog available in structured, machine-readable formats, revealing a significant gap between merchants’ awareness of agentic commerce and their operational readiness. There also needs to be a dispute layer, which brings governing bodies such as RBI into the picture. These numbers point to the same problem: adoption is moving, but trust, disputes, and transaction errors still need to be addressed.
Build the Trust Layer Before the Payment Layer
When building agents, the trust layer needs to come before the payment layer. The agent needs boundaries that people can trust.
There are four things to take away from this.
1. Your Data Determines What the Agent Sees
Your data is the foundation, and the model is only as good as the data it has. The data needs to match the purpose for which the agent is being built. Agents are less forgiving than humans because they depend on the data available to them. A person looking at a product may notice that something seems wrong and investigate. An agent may see an incorrect attribute and skip the product. Merchants need accurate product data for agents to discover and select their products.
2. Design the Mandate, Not Just the Model
The mandate and guardrails should be defined before optimizing the model. You can build a strong model, but if the system does not serve the business purpose, the project could fail.
3. Build the Recovery Process Before Purchase
In payments and banking, anything that affects a user's money needs a way to recover from an error. You need to know how to undo a transaction. A model can work well and still result in a failed project if there is no way to recover the money after something goes wrong. Then comes the question of accountability. Who is responsible? The bank? The agent? The merchant? The payment gateway? The user who gave the mandate?
That creates at least five parties whose responsibilities may need to be defined: the user, agent, merchant, payment provider, and bank or issuer.
That is why the undo needs to be designed before the buy. When building an agent, define what happens when a transaction fails or needs to be reversed.
The Bigger Opportunity May Be B2B
Much of the attention around agentic commerce is focused on consumer-facing experiences. They are easier to demonstrate. A consumer agent that finds and buys a product makes for a clear demo. But the larger opportunity may be B2B.
A large volume of purchases happens in the B2B sector, and there is a gap in the market because many agents are not built for B2B procurement. Much of the current work focuses on business-to-consumer (B2C) commerce. For someone thinking about a new venture or a new business model, B2B is an area worth looking at.
The Market Is Still Experimenting
Across the ecosystem, companies and organizations are building and testing agentic commerce, including:
- Protocol creators
- Payment rails and networks
- Merchant adopters
- Executors
- Referees and regulators
- B2B procurement
They are experimenting, enabling purchases, and adjusting their systems. The market needs to continue testing these systems to understand what agentic commerce should replace, where it works, and where it does not. The situation echoes an earlier challenge faced by digital payments: technology can become available before users fully trust the new way of transacting. Trust was low, but the technology was available. Trust developed over time. This time, there is an opportunity to build trust, undo mechanisms, regulatory controls, and guardrails alongside the technology.
The Best Agent Is Not Just the Best Model
Teams need to bring together:
- Agent reasoning
- Product data quality
- Delegated identity
- Payment control
- Observability
- Regulatory discipline
When you build an agent, the model is only one part of the system. You also need to ask:
- Who is the end user?
- How does the agent affect that user?
- Where does the governing body come in?
- What are we not doing correctly?
- How will this affect the Indian or global economy?
The financial sector has a high level of regulation, so it does not adopt every new AI system as soon as it appears. In financial services, AI systems typically operate within model-risk management, validation, audit, compliance, and regulatory oversight processes before and after deployment. That does not mean there is no experimentation.
Agentic commerce capabilities are already being deployed in production in some markets, with more experimentation underway. The direction is taking shape. The next stage is to build agents with the payment, identity, governance, recovery, and trust questions in mind.
