For centuries, commerce was built around two parties: a buyer and a seller. The payment rails connecting them were designed around a simple assumption: a human makes the purchasing decision, authorizes the payment, and can take responsibility when something goes wrong.
AI agents are challenging that model.
The next evolution of commerce is not simply about better recommendations or smarter chatbots. It is about giving software the authority to discover products, make decisions, complete purchases, and initiate payments on a person's behalf.
That creates a fundamentally different system: buyer + agent + seller.
And the difficult question is no longer whether AI can shop for us. It is whether we can trust it to spend our money.
From Human Clicks to Delegated Authority
Agentic commerce can be understood as delegated authority. A person or organization gives an AI agent permission to select goods and complete payments within predefined boundaries without requiring human approval for every individual transaction.
Think about everyday purchases. You might regularly order groceries, household supplies, coffee, or other routine products. Instead of repeatedly opening an application, searching for the same items, comparing options, and completing checkout, you could give an agent a mandate.
The mandate could include what the agent is allowed to purchase, how much it can spend, which merchants or categories it can use, how frequently it can purchase, and when its authority should expire.
The agent then operates autonomously within those boundaries.
This changes the meaning of a payment transaction. A human is no longer necessarily clicking the final "Buy" button. The system is establishing trust between the user, the agent, the merchant, and the payment infrastructure.
Identity, Mandate and Recourse
For agentic commerce to work safely, three concepts become critical: identity, mandate, and recourse.
First, the merchant needs to know who the agent represents. An agent may be acting on behalf of an individual, an employee, or an organization. This makes agent identity increasingly important.
Second is the mandate. What exactly has the user authorized the agent to do?
Giving an agent permission to "shop for me" is very different from allowing it to spend an unlimited amount. The agent needs explicit boundaries.
Third is recourse. If an agent makes an incorrect purchase, users need a mechanism to revoke its authority, correct the decision, dispute the transaction, or recover the funds.
This is where agentic commerce starts looking less like a traditional shopping experience and more like a regulated authorization system.
The Storefront Is Moving Into the AI Interface
One of the biggest changes is happening at the storefront itself.
Traditionally, a customer visits a website or mobile application, searches through a catalog, selects a product, adds it to a cart, and checks out.
AI is changing that flow.
Large AI platforms are increasingly becoming discovery interfaces for products. Instead of leaving an AI conversation to visit a merchant's website, a user can discover products, compare them, and potentially complete the purchase within the same interface.
This creates a new model:
Discovery happens inside the conversation.
The AI interface effectively becomes the storefront.
The ecosystem already includes developments involving platforms such as OpenAI, Perplexity, Google, Microsoft, Stripe, PayPal, and Shopify. The important shift is not simply that AI can recommend a product. It is that the entire journey, from discovery to authorization to payment, can increasingly happen through machine-to-machine interactions.
Checkout Is Becoming a Protocol
For years, checkout was primarily a user interface.
A customer added products to a cart, entered payment information, selected an address, and clicked a button.
Agentic commerce changes that abstraction.
Checkout is increasingly becoming a protocol rather than simply a page.
An agent does not need to interact with a traditional checkout page in the same way a human does. It needs structured information about the product, price, inventory, authorization requirements, payment mechanisms, fulfillment, and transaction status.
This is why protocols are becoming such an important part of agentic commerce.
The emerging stack can involve different layers for discovery, identity, authorization, payments, checkout, and settlement. Protocols such as MCP and emerging commerce and payment protocols are pieces of this larger architecture.
The exact technology stack will continue to evolve, but the architectural requirement is clear: AI agents need standardized ways to communicate with merchants and payment networks.
Product Data Becomes More Important Than Advertising
There is another major implication for merchants.
Traditional commerce has heavily relied on advertising and human discovery. A product can gain visibility because it appears prominently in search results, advertisements, recommendations, or sponsored placements.
Agents operate differently.
An agent evaluates machine-readable information.
That means product attributes, pricing, availability, metadata, specifications, policies, and structured catalog information become critical.
If a product has incorrect or incomplete attributes, an agent may simply exclude it.
Humans are relatively forgiving. A person may see an imperfect product description and still investigate further.
An agent is much more literal.
If the data says the product does not meet the user's requirements, the agent can move on to another product.
This means merchants increasingly need to think about agent discoverability, not just search engine discoverability.
The quality of product data becomes part of the product itself.
Real-Time Price and Inventory Matter
Agents also need reliable information.
Imagine giving an AI agent permission to purchase coffee for an office every Tuesday. The agent needs to know whether the product is available, what the current price is, whether the merchant can fulfill it, and whether the purchase falls within the user's mandate.
Outdated information creates a completely different risk profile when software is authorized to spend automatically.
Machine-readable catalogs therefore need to expose accurate product information, real-time pricing, inventory, and transaction conditions.
The better the data, the better the agent's decisions.
India Has a Different Starting Point
The evolution of agentic commerce looks particularly interesting in India because the country already has extensive digital payment infrastructure.
The discussion highlights UPI as a foundation for this transition and points to delegated payment models such as UPI Circle as an important conceptual building block.
Instead of creating an entirely new payment ecosystem, the opportunity is to extend existing UPI infrastructure to support delegated authority for AI agents.
The transcript cites 22.7 billion UPI transactions and ₹28.92 trillion in transaction value through June 2026, illustrating the scale of India's existing digital payment infrastructure.
That foundation could make agentic commerce significantly easier to integrate into the Indian ecosystem.
Public Rails vs. Private Rails
The emerging approaches in India and Western markets are not identical.
In the US, payment ecosystems involve private networks and organizations such as Visa, Mastercard, American Express, and PayPal.
India has a strong public digital payment infrastructure through UPI and an ecosystem where merchant participation can be coordinated through public platforms.
The discussion also points toward ONDC as an important component of India's approach because merchant catalogs can become more machine-readable and discoverable.
This creates an interesting possibility: a future UPI-based agent could potentially identify a product, evaluate it against the user's mandate, authorize the payment, and complete the transaction without requiring a human to interact with every step.
The Biggest Problem Is Not Intelligence. It Is Trust.
AI models are getting better at reasoning.
Protocols are being developed.
Payment infrastructure already exists.
Merchants are experimenting.
But one problem remains difficult: trust.
Would you allow an AI agent to spend ₹5,000 without asking you?
What about ₹500?
What about ₹50?
The amount is only one part of the question.
Users need to know exactly what the agent is allowed to do, how its decisions are made, what information it uses, and what happens when it makes a mistake.
The transition to digital payments offers a useful comparison. When UPI was introduced, many users were initially cautious. Over time, familiarity, infrastructure, security mechanisms, regulation, and repeated successful transactions helped establish trust.
Agentic commerce may follow a similar path.
The technology can arrive before users are comfortable with it.
That means trust cannot be treated as a feature that gets added later. It has to be part of the architecture.
Build the Undo Before the Buy
One of the most important engineering principles from the discussion is simple:
Build the undo before you build the buy.
When an AI agent has the ability to spend money autonomously, the ability to reverse, cancel, revoke, dispute, or recover a transaction is just as important as the ability to initiate it.
Before allowing an agent to execute purchases, engineers should answer:
What happens if the agent buys the wrong product?
What happens if the price changes?
What happens if inventory information is stale?
What happens if the agent exceeds its mandate?
What happens if the user wants to revoke its authority?
What happens when the merchant disputes the transaction?
A system that can execute transactions but cannot provide effective recovery is incomplete.
Guardrails Should Be Designed Before the Model
It is easy to focus on the intelligence layer.
Developers naturally want to build agents that reason better, use more tools, and automate more tasks.
But agentic commerce requires a different priority.
The guardrails need to be designed first.
An agent should have clearly defined identity, mandate, budget, scope, expiration, revocation, recourse, and observability.
Only after these boundaries are established should the system focus on increasing autonomous execution.
B2B Could Be the Bigger Opportunity
Much of the current attention around agentic commerce focuses on consumer experiences.
That makes sense. An AI that automatically orders groceries or recommends products is easy to demonstrate.
But the larger opportunity may exist in B2B procurement.
Organizations make significantly more complex and repetitive purchasing decisions. Procurement involves vendors, contracts, approval policies, budgets, inventory requirements, compliance, and multiple stakeholders.
These are exactly the kinds of structured processes where autonomous agents could have significant impact.
A B2B purchasing agent could potentially monitor inventory, evaluate approved suppliers, compare products, verify contractual conditions, and execute purchases within predefined authorization limits.
The technology therefore has an opportunity to move beyond consumer shopping assistants into autonomous procurement systems.
Regulation Becomes Part of the Architecture
Commerce involving AI agents creates a new regulatory challenge because there are now multiple actors involved in a transaction.
The buyer delegates authority.
The agent executes the action.
The merchant fulfills the order.
The payment network moves the funds.
A financial institution may provide the payment account.
A regulator oversees the system.
In India, the role of the RBI becomes particularly important because payment activity is already highly regulated.
Agentic commerce cannot simply be treated as another AI feature. It sits at the intersection of AI, payments, identity, authorization, consumer protection, cybersecurity, and regulation.
That means engineering teams building these systems need to think beyond model performance.
What Developers Should Take Away
The most important lesson is that agentic commerce is not primarily a model problem.
It is a systems problem.
The agent needs good reasoning, but reasoning alone is insufficient.
It needs high-quality product data, verifiable identity, explicit authorization, payment controls, observability, dispute mechanisms, and regulatory discipline.
The winning system will not necessarily be the one with the smartest model.
It will be the one that successfully combines agent reasoning + product data quality + delegated identity + payment control + observability + regulatory discipline into a reliable end-to-end system.
The Next Commerce Interface May Not Look Like an App
The biggest shift may ultimately be invisible to users.
People will still buy coffee.
They will still order groceries.
Companies will still procure equipment and services.
The difference is that users may increasingly describe what they want instead of navigating through interfaces to make every individual decision.
The browser, mobile application, catalog, cart, and checkout page may remain important, but they will no longer be the only interface to commerce.
The AI agent becomes an intermediary.
And once software can act on our behalf, the most important question is no longer "Can the agent buy this?"
It becomes:
"Can we trust the agent to buy this within the boundaries we gave it, and can we recover when it gets something wrong?"
That is the real engineering challenge behind agentic commerce.
The technology is already moving. The infrastructure is being assembled. The experiments are happening.
The next phase will be about making autonomous transactions trustworthy enough for people and organizations to actually use them.
Frequently Asked Questions
What is agentic commerce?
Agentic commerce is a model where a person or organization delegates authority to an AI agent to discover products, make purchasing decisions, and complete payments on their behalf without requiring human approval for every transaction.
How is agentic commerce different from traditional e-commerce?
Traditional e-commerce generally involves a human browsing products, adding items to a cart, and completing checkout. In agentic commerce, the AI agent can perform these activities autonomously within predefined rules and spending limits.
What are the three key requirements for an AI agent making purchases?
The three important concepts are identity, mandate, and recourse. Identity establishes who the agent is acting for, mandate defines what the agent is permitted to do, and recourse provides mechanisms to revoke, correct, or recover from an incorrect action.
Why is trust important in agentic commerce?
Users need confidence that an agent will operate within its authorization, use accurate information, and provide ways to recover when something goes wrong. Trust therefore needs to be designed into the architecture rather than added after deployment.
Will AI agents replace traditional shopping apps?
Not necessarily. Traditional catalogs and shopping interfaces will continue to have a role, but product discovery and purchasing could increasingly happen inside AI interfaces. Users may interact with an AI agent instead of manually navigating multiple shopping applications.
Why is product data becoming so important?
AI agents depend heavily on structured and accurate product information. Incorrect attributes, pricing, inventory information, or metadata can cause an agent to skip a product. This makes machine-readable product data increasingly important for agent discoverability.
What does “checkout is a protocol” mean?
In traditional commerce, checkout is primarily a user interface or page. In agentic commerce, checkout can become a machine-to-machine protocol through which an agent communicates purchasing and payment information without requiring a human to interact with every step.
How could agentic commerce work with UPI in India?
The discussion describes extending existing UPI infrastructure to support delegated payment models for agents. Since UPI already provides a large digital payment ecosystem, agentic commerce could potentially build on existing payment rails.
Is agentic commerce already being implemented?
Yes. Companies and technology providers are experimenting with AI-powered shopping, payment integrations, merchant catalogs, and protocols. However, trust, regulation, authorization, dispute handling, and interoperability are still evolving.
What happens if an AI agent makes the wrong purchase?
The system needs mechanisms for cancellation, revocation, disputes, recovery, and correction. This is why the principle "build the undo before the buy" is particularly important when designing autonomous payment systems.
Is B2B a major opportunity for agentic commerce?
Yes. B2B procurement involves repetitive purchasing activities, vendors, contracts, budgets, inventory, approval rules, and multiple stakeholders. These structured workflows could provide significant opportunities for autonomous purchasing agents.
What should developers prioritize when building shopping agents?
Developers should consider data quality, identity, authorization, spending limits, guardrails, observability, recovery mechanisms, and regulatory requirements alongside the AI model. A capable model without these controls can still produce an unreliable payment system.
What will determine the winners in agentic commerce?
The strongest systems will not necessarily have the most capable AI model. Success will depend on combining agent reasoning, high-quality product data, delegated identity, payment controls, observability, and regulatory discipline into a reliable end-to-end system.
Watch the video here!
Top comments (1)
The shift from “buy” to delegated authority is the real inflection point. Once agents can initiate transactions, identity, scoped authorization, observability, and rollback become first-class engineering concerns. From a product engineering perspective, this is where the real complexity lies: building agents that are not just capable, but predictable and controllable in production. The “build the undo before the buy” principle is one we strongly agree with at GeekyAnts.