TL;DR
- Agentic commerce is AI agents searching, comparing and paying for things on someone's behalf, and its rails went live in 2025 with Mastercard Agent Pay, Visa Intelligent Commerce, Google's Agent Payments Protocol and the Agentic Commerce Protocol from OpenAI and Stripe.
- Most of what's written about it is a merchant and checkout story, while the hardest questions sit at the bank: whose money it is, who gave the agent authority, and who answers when it goes wrong.
- A payment token can be scoped to one merchant, one currency and one amount, but it can't tell which of three people on a business account approved the purchase, or whether Friday's payroll can absorb it.
- We believe banks should treat AI agents as a third actor on the account, working through the same governed journeys, permissions and audit trails as customers and staff.
- Small business banking is where this pays off first, because business buying is predictable, multi-user and tied to cash flow, and because 91% of users say they'd consolidate their financial activities into one platform with the right features.
- The bank that makes its accounts, lending and servicing safe for agents keeps primary-bank status; the bank that doesn't becomes the funding source behind someone else's agent.
Agentic commerce arrived on the merchant's terms
Agentic commerce became working infrastructure in 2025. Mastercard launched Agent Pay and Visa launched Visa Intelligent Commerce in April. Google announced its Agent Payments Protocol in September, and OpenAI and Stripe released the Agentic Commerce Protocol the same month. Each gives an AI agent a way to buy with the shopper's permission. By early 2026, ACP was processing live transactions for Etsy and expanding to over 1 million Shopify merchants.
So what is Agentic Commerce? An agent searches, compares and pays on behalf of a person or a business, while the human sets the rules and handles the exceptions. In Stripe's guide to getting started, a customer asks for the cheapest nonstop flight that lands before noon. The agent weighs the options, the customer approves, and the agent completes the purchase without the customer ever seeing a checkout page.
McKinsey projects up to $1 trillion in orchestrated US B2C retail revenue by 2030, and $3 trillion to $5 trillion globally. More conservative estimates put it at $190 billion to $385 billion (Morgan Stanley) and $300 billion to $500 billion (Bain). Still sizable.
Retailers are told to make product data machine-readable, payment providers build tokens, and platforms argue over protocols. The bank turns up, when it turns up at all, as the place the money comes from. We beleive this leaves out the most important part. Whose money is it, who gave the agent the right to spend it, what does the purchase do to the account's cash position, and who answers when it goes wrong? Those are bank questions, and how a bank answers them decides whether it stays its customer's primary bank in a world where this is as common as asking Alexa ti play music.
The token knows the amount, not the business behind it
The merchant side has some elegant controls. ACP completes checkout through Shared Payment Tokens scoped to a single merchant, currency and amount, and agents work inside spending limits and approval thresholds the human sets. A token like that can't be replayed at another store or stretched to a larger sum. For the merchant, that's most of the trust problem solved.
Now look at the same payment from the bank's side of a small business account. Picture a bakery with three people on the account: the owner, a manager who orders flour, and a bookkeeper who pays suppliers. The manager's agent finds a cheaper supplier and reorders. The token is perfectly scoped, and the bank sees a payment. What it can't see, unless the bank has built for it, is which of the three people gave the agent that authority. Nor can it see whether that person may take on a new supplier at all, or that payroll goes out on Friday and the balance won't cover both.
Business buying is where agents fit most naturally. Recurring purchases, subscription management, business-to-business procurement and inventory replenishment are predictable, standardized transactions that suit automation, which is a fair description of how a small business spends its week. Several users, delegated authority and a tight cash cycle are the normal state of a business account, and none of them is in the token.
The networks secure the payment and stop at the account
Speaknto people in this field and you'll here a common objection that this is already handled. Visa and Mastercard run agent programs, payment protocols pass a permissioned token so the shopper's card details are never exposed, and the merchant usually remains the merchant of record. Why should an issuer do more than authorize the transaction?
Because the payment is the last step of a longer chain, and the networks see only that step. Before it, the agent needs to know what balance is available, which user it's acting for and what that user may commit to. After it comes the dispute when the flour arrives damp, the refund, the reconciliation, and perhaps a short-term credit line when a large order lands before the bakery's own invoices are paid. Those are onboarding, servicing and lending journeys, and they run inside the bank.
McKinsey argues that agentic commerce can spread faster than web or mobile commerce did, because agents can travel the digital paths to purchase that people already use, and that it calls for re-architecting identity management. Carry that into banking and the exposure is clear as daylight: an agent will use whatever path a bank opens, including journeys designed on the assumption that a human is the one tapping. Our advice to banks are that agentic payments need a control layer on the bank's side. The networks secure the money in motion. The bank still has to govern the account it moves from.
Agents belong on the account as a third actor
For most of banking's history, the frontline had two actors: customers and employees. AI agents are the third, and each one needs a clear answer to what it's entitled to do, under what authority and within what limits.
The tempting shortcut is a separate channel for agents: an API on the side, with its own rules, its own logs and its own team. That leaves the bank with two versions of who can do what. When an agent's payment goes wrong, the fraud team, the contact center and the relationship manager each look at a different record.
The better design decision is to have agents act through the same governed journeys, the same permissions and the same audit trail as the person who delegated to it. When the bakery manager's agent initiates a payment, it hits the entitlement check the manager would hit in the app. If the business requires a second approval above a threshold, the agent waits for it too. And the record shows who delegated, which policy applied, which model version acted and what happened.
That's how we built the AI-native Banking OS. It sits above the core and the bank's existing systems rather than replacing them, so customers, employees and AI agents work from the same context, the same policies and the same execution layer, which is what we mean when we talk about the Unified Frontline. Account opening, credit decisioning, payment initiation and KYC checks are reusable capabilities that any workflow or agent can call, and every action carries a record of the policy applied, the actor, the model version and the outcome. Autonomy is staged: an agent starts by preparing and recommending, moves to acting with explicit human approval, and earns the right to act alone within guardrails, one domain at a time and revocably. Control built in this way speeds delivery up, which is why the bank that governs AI best will ship it fastest.
Small business banking is where agentic commerce makes sense first
Companies with fewer than 200 employees manage an average of 42 different SaaS applications (BetterCloud), and 91% of users say they'd consolidate their financial activities into a single platform if it offered the right features. A decade of unbundling is turning into rebundling.
Someone will do the rebundling. The embedded finance market was valued at $148 billion in 2025 and projected to pass $1.7 trillion by 2034 (Credence Research), and 74% of businesses plan to launch an embedded finance offering. Accounting platforms, ERP providers and marketplaces already sit in a small business's daily work. If the agent that runs its purchasing lives there, the bakery's takings keep landing in the same bank account while the agent deciding where that money goes next works for someone else. With 46% of banking customers open to switching or multibanking (Financial Brand), primary-bank status can't be assumed.
Banks do hold an advantage the platforms can't copy quickly. While 72% of consumers use AI tools, only 24% feel comfortable using them for financial decisions (Bain). That gap is the bank's opening. An agent that watches cash flow, flags the Friday payroll gap, proposes the credit line and pays the supplier, all inside rules the owner set, is a service only the account holder's bank can offer with the full picture. Real-time cash visibility is the first thing such an agent needs, and modernizing cash management is the groundwork.
What to build before the agents takenover
- Give delegation a home. Decide, per business account, which user can authorize an agent, for which actions and up to what amount, and keep it where every channel reads it.
- Expose capabilities to agents. Balance checks, payment initiation and credit decisions should be callable by an agent through governed endpoints, never scraped from an app built for thumbs.
- Make agent actions leave human-grade evidence. Every action should carry who delegated, the policy applied and the outcome, so a dispute an agent caused is resolved like any other. Our guide to reviewing agentic AI architecture before it acts covers the checks, and an AI governance framework for banking sets out who owns them.
- Start with one loop. Supplier payments for small business customers, with dual approval, at the stage where the agent prepares and a person approves. Earn autonomy from there, the way banks roll out Digital Banking and Agentic Banking on the Banking OS: one priority first, then the next.
The protocols will keep multiplying and changing names. The question each bank has to answer will stay the same: when an agent asks to spend a customer's money, does it go through your rules or around them?
FAQ
What is agentic commerce?
Agentic commerce is buying and selling in which AI agents search for products or services, compare options, make purchase decisions and complete payments on behalf of a person or a business. The human sets preferences and limits, such as price caps, delivery times or approved payment methods, and steps in on exceptions. Unlike a chatbot or a recommendation engine, the agent completes the transaction instead of only suggesting it.
Is agentic commerce real?
Yes, though it's early. The Agentic Commerce Protocol from OpenAI and Stripe processes live transactions for Etsy, and Perplexity and Amazon Alexa Plus already complete purchases for consumers. A 2026 IBM Institute for Business Value study found that 45% of consumers already use AI for at least part of the buying journey. Gartner has noted that vendors use the term loosely, so check what any offer labeled agentic actually does.
What companies are using agentic commerce?
Amazon and Yahoo are among the enterprises adopting it, and Alibaba, Tencent and ByteDance have built AI shopping apps, with the Qwen chatbot letting users complete transactions inside its interface. On the infrastructure side, Visa, Mastercard, Google, OpenAI and Stripe have launched agent payment programs or protocols. Google and Shopify co-developed the Universal Commerce Protocol, announced in January 2026 with Etsy, Target and Wayfair among its founding members.
Is agentic commerce the next big thing?
The forecasts agree it's large and disagree on how large. McKinsey projects $3 trillion to $5 trillion globally by 2030, while Morgan Stanley estimates $190 billion to $385 billion and Bain $300 billion to $500 billion. For banks, the sharper question is whether they take part in agent-led buying or only fund it.
What does agentic commerce mean for banks?
A bank needs to know which agent is acting, for which customer or user, under what authority and within what limits, before it lets money move. For small business accounts with several users and tight cash flow, that means agents should act through the same governed journeys, permissions and audit trails as customers and staff. Banks that build this keep their place in their customers' financial decisions; banks that don't risk becoming the funding source behind someone else's agent.
