01 Trigger Event

On November 12, 2026, Bloomberg Technology reported that Mastercard Asia Pacific President Richard Wormald said agentic AI could create a new machine-to-machine commerce channel. Whether this channel can scale will depend on consumer trust.

The original report disclosed no transaction volumes, fee rates, merchant adoption figures, or specific launch timelines. Nor was this a Mastercard product announcement. It was a thesis about a new channel: AI agents will participate in product discovery, comparison, decision-making, and payment. Payment networks must therefore prepare for an entry point that is neither exactly e-commerce nor a conventional API.

02 What This Really Means

The issue is not whether AI agents will place orders for consumers, but who will control the underlying customer relationship.

The traditional e-commerce journey runs as follows: a consumer discovers a product, submits an order, and the payment network handles authorization and clearing. Agentic commerce adds a layer of machine decision-makers. A consumer may say only, “Buy a monitor suitable for a small team,” and the agent will search for candidates, read reviews, select a merchant, construct the order, and initiate payment.

Payment companies are therefore facing not merely a new frontend, but a new distribution layer. The point of entry may migrate from apps, search results, and e-commerce platforms to models, agent runtimes, and MCP toolchains. Whoever defines which payment capabilities agents can invoke may influence transaction paths, authorization granularity, identity credentials, and dispute handling.

Mastercard’s potential advantages include its existing tokenization, identity verification, risk controls, and merchant network. But network position does not guarantee a default win. If OpenAI, Google, Anthropic, or independent agent platforms control consumer intent and package payment as a native capability, Mastercard could still be reduced to back-end clearing infrastructure.

What will ultimately be priced may not be “support for agentic AI” as a feature. The more important question is whether Mastercard can upgrade its existing payment rails into a trusted machine-to-machine commerce protocol that developers prefer to integrate first.

This is also a variant of the Stratechery-style aggregation problem: when user demand converges on a small number of agent gateways, an infrastructure company’s value depends on whether it stands behind those gateways or is squeezed into a cost position by them.

I have not operated this transaction flow inside Mastercard, so my assessment of its current progress may be mistaken. Strategically, however, this is not about adding another payment button. It is about redistributing control over transactions.

03 Historical Analogies / Structural Comparison

When the iPhone introduced a gateway to mobile payments in 2007, Visa and Mastercard did not disappear. Apple, however, reshaped value distribution through its hardware, account system, and App Store. The payment networks supplied secure rails, while Apple controlled the consumer gateway and some payment rules.

Around 2014, the spread of AWS APIs was often misread as the replacement of traditional IT. In reality, AWS did not eliminate servers and databases. It packaged infrastructure capabilities as a distribution layer that developers could invoke directly. The party controlling the calling interface often had more bargaining power than the company merely supplying the underlying resources.

Agentic commerce is closer to a combination of the 2007 iPhone and the 2014 AWS model:

  • The iPhone controls the consumer gateway.
  • AWS provides developer-accessible primitives.
  • Mastercard controls identity, authorization, and clearing rails.

This time, however, one previously less prominent variable becomes decisive: trust.

If an agent spends money on a consumer’s behalf, mistaken purchases, prompt injection, profile misuse, and questions of liability translate directly into payment losses. Consumer trust is not marketing copy; it is a product constraint. Authorization interfaces, spending limits, merchant provenance, audit logs, and dispute mechanisms must all be sufficiently clear.

Visa and Mastercard historically became networks because merchants accepted them, consumers recognized them, and a common rule set absorbed transaction risk. For agentic commerce to develop similar network effects, it must establish comparable rules rather than merely enable a demo to complete checkout.

The agent wave following the launch of ChatGPT in 2022 has repeatedly demonstrated that generating an answer is not the same as assuming responsibility. Payment is one of the most consequential settings for accountability: an erroneous code request can fail, while an erroneous debit can trigger complaints, refunds, and regulatory scrutiny.

04 What This Means for AI Builders

First, payment should not be treated as just another tool call inside an AI product. It should be designed as a constrained transaction workflow. Builders need to distinguish three separate functions: the agent proposes intent, the user gives explicit authorization, and the payment network executes settlement. Combining all three into a single “buy” button creates serious gaps in the allocation of responsibility.

Second, the MCP and tool ecosystem may become a distribution layer for payment capabilities. If merchants expose product catalogs, inventory, pricing, and refund policies through MCP servers, agents can handle discovery and transactions directly. The competition over developer tooling will therefore extend into commerce primitives: whoever can expose identity, payment, authorization, audit, and fulfillment through standard interfaces will be better positioned to control the agent ecosystem.

Third, agent products should record a traceable decision chain from day one. At a minimum, it should include the agent version, model, prompt, candidate products, final selection, points of user confirmation, payment credentials, and refund outcomes. Without this data, teams will struggle to determine whether failures originated in model reasoning, tool permissions, or merchant systems, and will struggle to explain responsibility to users.

Fourth, token economics must incorporate transaction economics. A product search may consume substantial tokens, while the order value and gross margin may not be sufficient to cover inference and payment-dispute costs. Agentic commerce will make “inference cost per order” more important than “price per request.” Routing lower-capability models through product filtering and reserving stronger models for high-value decisions may therefore be a more effective optimization than switching models alone.

Fifth, independent AI product founders should not assume that checkout will remain platform-owned by default. If a product’s value depends heavily on cross-merchant comparison and execution, a platform may provide distribution while also locking users, data, and transaction relationships into its own system. A more resilient design gives agents the ability to execute intent across channels while preserving portable user authorization and order history.

This week, I will prioritize two checks: whether existing agent workflows rigorously separate user confirmation from actual payment, and whether the team already has end-to-end payment auditing and failure-recovery mechanisms. The latter is often harder than integrating a payment API.

05 Counterarguments / Risks

I may be overestimating how quickly machine-to-machine commerce will emerge as an independent channel. Agentic AI may remain merely an assistive layer within traditional e-commerce. Consumers may be unwilling to grant models full purchasing authority, especially for luxury goods, medicines, and frequently replenished products. Merchants will also continue subsidizing platform traffic, manipulating rankings, and restricting agent access to inventory.

The second risk is that network effects may not accrue to payment companies. During the Android and iPhone eras, Apple and Google demonstrated the power of controlling gateways. The AI agent market, however, has no stable winner. Model providers, browsers, super-apps, and vertical platforms may all intercept demand. If payment becomes a replaceable provider within an agent’s calling stack, Mastercard’s network may not automatically carry over into the agent era.

The third risk is that trust may be mistaken for consumer reassurance. Consumers may trust a brand without trusting an agent operated by a third party that continually changes models or tool permissions. The scarce capability in the future will not be “making users feel safe.” It will be making every automated transaction explainable, reversible, and accountable.

The fourth risk is that liability costs may suppress adoption of automation. Fraud, prompt injection, and fulfillment errors can increase payment-network risk costs or lead to limits on the amounts and use cases allowed for agent transactions. If every transaction requires multiple rounds of identity verification and manual review, the unit economics of agentic commerce may not work.

I therefore view this report as a directional signal, not evidence that Mastercard has secured a new distribution layer. Over the next 12 months, the metrics worth tracking are not the number of agent-payment demos. They are standardized authorization interfaces, genuine repeat-purchase rates, dispute rates, contribution profit per order, and whether agent gateways retain portable user intent.

If those metrics fail to emerge, Mastercard may have bet on a long-running technical narrative rather than a new commercial channel.