Artificial-intelligence systems are moving from helping consumers decide what to buy toward acting on those decisions. Meta's Muse can browse, fill forms, negotiate, and seek approval before purchases. Banks are now warning that the institutional rules governing this form of agentic commerce remain less mature than the technology enabling it.
On September 22, NatWest, Bank of America, Capital One, Commonwealth Bank of Australia, ING, and ASB Bank published joint principles for trusted agentic commerce. Their concerns include identity, authorization, fraud prevention, liability, privacy, interoperability, and customer protection. Reuters reported the same day that the banks view AI shopping agents as creating risks of scams, fraud, and data-privacy breaches as technology companies increasingly promote systems that can select products and make purchases for users.
Recommendation is becoming execution
Digital platforms have long shaped consumer choice through rankings, advertising, personalization, and recommendation systems. Agentic commerce changes the institutional position of the system. The agent can search, compare, decide, communicate with sellers, and initiate payment. Several stages that were previously separated across a consumer, merchant, search engine, payment network, and bank can become part of one automated sequence.
That compression creates convenience and a governance problem. A recommendation can be rejected with little consequence. An executed transaction creates obligations. Money moves. Credentials are used. Goods are ordered. A contract may be formed. The relevant question becomes whether the agent had authority to act and which institution bears responsibility when the action is wrong.
Authorization becomes an infrastructure
Meta's design for Muse illustrates one response. The company says the agent runs inside a dedicated secure virtual machine, uses separate safeguards before internet actions, keeps credentials in protected storage, and asks users for approval before sensitive actions such as sending an email or making a purchase. Payment arrangements can also use tokenization or one-time credentials rather than exposing a consumer's underlying card details.
These features matter because consumer control in an agentic system cannot depend entirely on a general instruction such as "find me the best option." Authority needs granularity. A user may authorize research without purchase, purchase below a price ceiling, recurring payments to specified merchants, or a transaction only after explicit confirmation. The design of authorization therefore becomes part of the market's institutional architecture.
Liability follows the chain of action
Traditional consumer-payment systems contain established roles for merchants, issuers, networks, payment processors, and account holders. Agentic commerce inserts another decision-making actor into that chain. If an AI agent selects a fraudulent merchant, misinterprets a return policy, chooses an insecure payment method, or completes a transaction outside the consumer's intended constraints, existing protections may apply unevenly depending on how the transaction was authorized and represented.
The banks' principles recognize this problem by emphasizing transparency, safety, privacy, customer choice, and interoperability. These are useful starting points. Their practical meaning will depend on technical standards and legal rules that specify how agents identify themselves, how consent is recorded, what evidence survives after a dispute, and how responsibility is allocated among agent providers, banks, merchants, and consumers.
Consumer choice is also a problem of incentives
Agentic commerce introduces an additional issue: whose objective does the agent optimize? A system may receive revenue from merchants, prefer payment methods associated with commercial partnerships, rank products according to platform incentives, or steer consumers toward services within its own corporate ecosystem. Disclosure alone may be insufficient if users cannot meaningfully inspect how those incentives affect action.
This creates a familiar institutional problem in a new technical form. Financial advisers, brokers, marketplaces, and payment platforms have long faced conflicts between user interests and intermediary incentives. AI agents can intensify the problem because search, recommendation, negotiation, and execution may occur too quickly and opaquely for consumers to reconstruct the path from request to transaction.
From interface design to financial governance
The decisive governance question is whether agentic commerce will remain an interface feature or become a new layer of financial infrastructure. If agents begin to mediate large shares of search and payment activity, rules governing their identity, authority, records, conflicts, and liability will shape market access and consumer protection across the economy.
The transition is already visible. Banks are attempting to establish principles while technology companies build agents capable of acting across websites and payment systems. The institutional sequence matters. Rules created after market concentration and technical standards have hardened may have less influence over system design. Agentic commerce therefore presents a near-term governance problem: establish accountable authority before automated action becomes ordinary enough to disappear into the background of everyday consumption.