top of page

The Foundations of Agentic Commerce

13 hours ago
6 min read
"Agentic commerce does not replace the payments stack. It runs on top of it and creates new dynamics – and pressure – on every layer."

Most payments conferences this year have an agentic commerce panel. The demos are impressive: an AI agent compares flights, picks a fare and pays while the customer sits in a meeting. What we don't see of course, is the messy plumbing – or the number of times the agent prompts the user for login credentials or authentication to complete a task. When software, not a person, presses "buy," assumptions our industry has relied on for decades get tested. Who is actually paying? Did they really agree to this purchase? Who absorbs the loss if the agent gets it wrong?

The good news for payments professionals is that the existing stack still matters – pre-authorization, fraud/ID verification, ATO checks and settlement are still happening. Card networks, wallets, tokenization, fraud engines and dispute rules all still apply.

What changes is that each layer now has to account for a new party in the transaction: the agent.

Below are six foundations any agentic commerce ecosystem will need. Read them as a checklist. If you run payments for a platform, or build for people who do, each one is a gatepost you will likely face in the new world of Agentic Commerce.

Six foundations of agentic commerce: agent identity, delegated authority, credentials, merchant readiness, risk and disputes, and rails and settlement.

The six foundations at a glance · one question each

Each foundation adds the agent to a layer payments already has; the sections below take them in turn.

1. Agent identity: know your agent

The headline: Merchants must tell trusted agents from bad bots, and someone has to vouch for each agent.

Before a merchant accepts an order from an agent, it needs to know which agent this is, who built it and whom it is acting for.

This turns a decade of fraud thinking on its head. Merchant risk stacks were built to detect and block bots. Agentic commerce asks them to welcome some bots and keep blocking others. The emerging answer is "know your agent" (KYA): verify the agent and its developer, bind it to a verified person or business, and make that binding checkable at the moment of purchase.

The card networks are already building this layer. Visa's Trusted Agent Protocol helps merchants recognise trusted agents and tell them apart from malicious bots. Mastercard's Agent Pay requires agents to be registered and verified before they can transact.

For many Asian markets, there is a twist. In markets where wallets already hold verified identity for hundreds of millions of users, the wallet may become the natural anchor for agent identity. Whoever holds the verified account is well placed to vouch for the agent acting on it.

2. Delegated authority: consent that can be proven

The headline: Every agent purchase needs signed proof of what the customer actually allowed.

Identity says who the agent is. Delegated authority says what it is allowed to do, and the industry needs a record of that permission that holds up in a dispute.

Today consent is implied by the act of checkout: a person clicked "pay." With agents, the person may have given an instruction hours or weeks earlier, such as "buy the flight if it drops below US$400." That instruction has to be captured, scoped and signed, with limits on amount, merchant type, timing and whether the person must confirm the final basket.

Google's Agent Payments Protocol (AP2) gives one model. An intent mandate records what the user asked for, a cart mandate approves the final basket, and a payment mandate tells the network an agent executed the order. For unattended purchases, a detailed intent mandate sets the rules upfront so the agent can act later without the user present.

The key distinction for payments teams is human-present versus human-not-present. Regulators already care about it. India's rules on recurring card mandates and Europe's strong customer authentication both decide when a person must step back in, and agent flows will be tested against them.

3. Credentials: never hand the agent the card

The headline: Agents get scoped, revocable tokens, never the card or account itself.

Agents should never hold a raw card number or wallet password. They need credentials that are scoped, revocable and traceable back to the agent that used them.

Tokenization, which the industry built for mobile wallets and card-on-file, is the obvious starting point. Mastercard's Agentic Tokens extend its existing tokenization so that one token represents a cardholder's card, limited to one agent and one set of rules, and the cardholder can revoke that agent's access through the issuer's app. Processors are wrapping these network tokens too; Stripe, for example, supports both Visa and Mastercard agentic tokens through its Shared Payment Tokens.

The principle carries beyond cards. A wallet balance, a bank account linked by open banking, or a stablecoin wallet each needs its own version of a scoped, revocable, agent-bound credential. For providers in Asia, where account-to-account and wallet payments dominate many markets, this is where much of the build work will sit.

4. Merchant readiness: checkout becomes an API

The headline: Checkout is becoming an API, so plan to support more than one protocol.

An agent cannot use a checkout page designed for human eyes. Merchants need to expose their catalog, prices, stock, delivery options and policies in a form machines can read and act on.

Two open standards are competing to define that interface. OpenAI and Stripe launched the Agentic Commerce Protocol (ACP) in September 2025 to power checkout inside ChatGPT. Google and Shopify followed with the Universal Commerce Protocol (UCP) in January 2026, which pairs with AP2 for payment. ACP leans toward a platform-run checkout; UCP lets merchants publish their own profile and keep more control.

For payments directors, the practical lesson is to plan for more than one protocol. Large merchants are already supporting both. The questions that matter are who stays merchant of record, who owns the customer relationship, and what data the agent platform gets to keep.

5. Risk and disputes: who pays when the agent gets it wrong

The headline: Who pays for an agent's mistake is the biggest open question in agentic commerce.

The hardest unsolved problem in agentic commerce is not moving money. It is deciding who carries the loss when a purchase goes wrong.

Today's dispute rules assume two kinds of problems: fraud ("I didn't buy this") and merchant failure ("it never arrived"). Agents add a third: "my agent bought it, but not what I meant." Was that the user's fault for a vague instruction, the agent developer's for a bad decision, or the merchant's for a misleading listing? Existing chargeback reason codes were not designed to answer that.

This is why the signed records in foundation 2 matter so much. As one analysis of AP2 notes, mandates move proof of purchase from inference to signed intent. They give each party evidence of what the user allowed and what the merchant promised. Fraud models will also need new inputs, such as which agent acted and whether a human confirmed, and issuers will need to see that an agent was involved at authorization.

Expect liability frameworks to be negotiated network by network and market by market. Until they settle, risk teams should treat agent-initiated volume as a separate segment with its own limits and monitoring.

6. Rails and settlement: many ways to move the money

The headline: Agent payments will run on many rails, and in Asia, routing across them is the hard part.

Agentic commerce will not run on one rail. Consumer purchases will mostly ride cards, wallets and bank transfers, while machine-to-machine payments are pulling stablecoins into the picture.

The machine side looks different. An agent paying a few cents for data, an API call or compute cannot absorb card fees or wait for a monthly invoice. Coinbase's x402 protocol builds payment into the web request itself, settling in USDC. Activity is growing fast from a small base: Chainalysis counted more than 100 million cumulative x402 transactions through the first quarter of 2026, yet Bernstein estimated dollar volume at no more than US$25 million over a 30-day period. The transaction counts are large; the dollar values are still small.

For Asia, the rail question is sharper than in the US. Payments here run on many systems at once: real-time account-to-account schemes, QR networks, super-app wallets and local cards, across dozens of currencies. An agent shopping across borders has to route, convert and reconcile across all of them. That is old, hard Web2 work, and it is why agent builders will keep needing partners who know how traditional payments actually settle.

What this means for you

None of these foundations is finished, and that is the opportunity. The standards are still being written, and the people who understand how payments really work have a seat at the table.

It is the same movie with new actors. Identity, authorization, credentials, acceptance, risk and settlement have always been the core of payments. Agents do not remove any of them; they add a new party to each. If you run payments for a platform or build for those who do, five questions are worth asking now:

  • Identity: How will we recognise a trusted agent, and who vouches for it?

  • Consent: What record will prove what the customer authorised, and when must a human confirm?

  • Credentials: Can our tokens, wallets and accounts be scoped and revoked per agent?

  • Acceptance: Which protocols will we support, and who stays merchant of record?

  • Liability: When an agent gets it wrong, who pays, and how will we know?

In our next pieces, we will go deeper into each of these, starting with how leading players in Asia are approaching these defining issues.

Comments


bottom of page