AI shopping has spent most of its public life in the recommendation stage: find me a product, compare these options, tell me which one is cheaper. The next step is much more consequential. Payment systems are being designed so an AI agent can actually spend money on a person’s behalf.
Reuters reported that India is preparing a framework for agentic payments on UPI, the country’s enormous real-time payment network. The proposed system would let consumers delegate limited funds and rules to an AI agent so it can make low-value purchases without asking for approval every single time.
That sounds like a checkout feature. It is really an authorization problem.
The important question is no longer whether an AI can recommend the right product. It is who gave the AI permission to spend, how much it can spend, and who is responsible when it gets the decision wrong.
Why UPI matters
UPI is not a small experimental payment rail. Reuters reported that it processed 24.51 billion transactions in August 2026 worth roughly $314 billion. Building agentic-payment rules on top of infrastructure at that scale moves the conversation beyond a demo.
The reported framework, called the Unified Agent Protocol, is expected to use existing mechanisms for delegated funds and blocked balances. The design includes rule-based payments, spending limits, identity checks and liability provisions.
That combination tells us what the real challenge is. The payment itself is easy. Payment systems already know how to move money quickly. The hard part is expressing human intent in a way a machine can act on without turning the agent into an unlimited corporate card with a personality.
A spending limit is a trust product
Think about how a normal person delegates money today. A parent gives a child an allowance. A company gives an employee a card with a limit. A procurement team approves certain vendors and blocks others. A subscription renews only under previously agreed rules.
Agentic payments need the digital version of those boundaries.
A consumer might allow an agent to reorder household basics under $50, but require approval for anything larger. A business might let an agent buy shipping supplies from approved vendors, but never change payment details. A travel agent could hold a budget for a trip without having permission to spend outside the itinerary.
Once you look at it that way, the future of agentic commerce starts to look less like “AI shops for you” and more like programmable delegation with a shopping interface on top.
Identity and payment are converging
This also connects to another trend: AI agents increasingly need their own identity.
If an agent can spend money, merchants and payment networks need to know more than “this request came from a browser.” They need to know which agent is acting, who authorized it, what rules apply, and whether the agent is still permitted to transact.
That is why identity, authorization and payment infrastructure are developing together. The same questions that matter in enterprise security—ownership, least privilege, revocation and auditability—start showing up in commerce.
The checkout is the visible part. Governance is the system underneath it.
What this changes for merchants
Most merchants do not need to redesign checkout tomorrow. Agent-initiated payment volume is still early, and infrastructure readiness is not the same thing as consumer adoption.
But businesses should start noticing what machine buyers need from a store.
Product information has to be structured and accurate. Inventory and pricing need to be dependable. Policies need to be clear enough for software to interpret. Payment flows need to distinguish legitimate delegated transactions from fraud. Order history and returns may need to support an agent acting on behalf of the original customer.
That means some of the “AI commerce” work is not glamorous AI work at all. It is the same operational discipline ecommerce teams should already care about: clean data, explicit rules, reliable APIs and good audit trails.
Do not confuse infrastructure with demand
There is an important caveat here.
A national payment network preparing for agentic payments does not prove that consumers want autonomous shopping. It proves that serious institutions believe the possibility is important enough to design for.
Those are different claims.
We should expect a long period where agents are trusted with narrow, repetitive, low-risk purchases before people comfortably delegate expensive or emotional decisions. Reordering detergent is a very different trust problem from buying a laptop, booking a vacation or choosing a financial product.
So I would not build a strategy around the assumption that autonomous shopping is about to replace normal ecommerce. I would pay attention to the pieces becoming real: delegated budgets, machine identity, merchant acceptance and explicit liability.
The next phase of AI commerce is about permission
The first phase of AI shopping was about better answers.
The next phase is about authority.
Can the agent spend? How much? On what? For whom? Under which rules? Can the permission be revoked? What happens when the machine makes a mistake?
Those questions are much less exciting than “AI buys your groceries.” They are also the questions that have to be answered before agentic commerce becomes ordinary infrastructure.
The breakthrough may not be an AI smart enough to shop. It may be a payment system careful enough to tell the AI no.
Sources
Ready to scale your brand?
Let's talk about what growth looks like for your business.
Book a Free Strategy Call