Summary
As AI agents begin to act on behalf of users, payment systems built around apps, checkout pages and confirmation buttons may no longer fit every transaction. The central challenge will be establishing verifiable authorization, liability and settlement rules when no person is manually approving each payment.
|- Digital settlement assets may support machine-native settlement, but they do not by themselves solve identity, permissioning, fraud, liability or dispute resolution.
- Banks, card networks, payment companies and wallet infrastructure providers will need stronger policy controls, audit trails and escalation mechanisms for agent-triggered transactions.
- The institutional custody question is likely to expand from protecting keys and accounts to governing what automated agents may view, initiate and approve.
Payments were designed around a visible human action
A recent TechFlow article citing David Marcus, a former PayPal president and former head of payments and crypto at Meta, argues that the rise of AI agents is exposing a foundational assumption in digital payments: the system expects a person to sit in front of a screen, open an app, complete a form, receive a verification code and press a confirmation button.
That assumption has shaped much of the modern payments stack. Bank applications, checkout pages, card forms, one-time passwords and confirmation screens are not simply user-interface elements. They are also points at which identity, consent and risk controls are applied. Payment companies have spent years removing steps and reducing the time required to complete a transaction, while fraud detection, chargebacks and dispute processes have been designed around a human-initiated payment.
Agentic commerce changes the sequence. A user may describe an objective in natural language—arranging travel, finding a service or completing a purchase—and an agent may compare options, communicate with a merchant system and execute the task. The user may not open each application involved, and may not enter a new verification code before every individual action. The transaction therefore moves from a person operating a visible interface to an agent acting on that person’s behalf.
Authorization becomes more important than the button
When there is no single, obvious moment of manual confirmation, the key question is no longer simply whether a user clicked “confirm.” It becomes whether the agent had sufficient authority to perform the action in the first place.
That authority could involve several dimensions. What categories of goods or services may the agent purchase? What spending limit applies? Which merchants, jurisdictions or time periods are permitted? Can the agent delegate part of the task to another agent? Does an instruction to arrange a trip authorize a change fee, an upgrade or a purchase from an unfamiliar provider?
These questions point toward a more granular form of authorization than ordinary account login. Authentication can establish that a user, device or account is connected to a transaction, but it does not necessarily prove that every action taken by an agent reflects the user’s intent. A future payment system may need to combine identity, task description, amount, merchant category, timing and risk level into an auditable authorization credential.
Liability is equally important. If an agent misunderstands an instruction, if a merchant-side agent fulfills an order incorrectly, or if a compromised integration initiates an unauthorized payment, responsibility may be contested among the user, agent provider, merchant, payment institution and underlying infrastructure provider. Existing rules for fraud, account takeover and chargebacks may not map neatly onto these situations.
The challenge, in other words, is not merely enabling an agent to move money. It is creating a permission system that can be verified, limited, revoked and reviewed after the fact. The participants also need a reliable record of how an automated decision was made when a transaction is disputed.
Agents could become a new digital gateway
Marcus’s argument draws on earlier experiments in conversational assistants and internet-native digital assets. The article describes his view that previous efforts were early because the relevant models were not capable enough and the regulatory environment for digital assets was less mature. He now sees advances in models and the development of digital-asset legal frameworks as conditions that could allow agentic interaction and digital settlement to develop at the same time.
Under this model, users would increasingly express goals to a personal agent rather than navigate a list of applications or websites. Businesses could deploy their own agents to handle procurement, customer service, inventory or settlement. Interactions between a user-side agent and a merchant-side agent could replace portions of the conventional web journey.
That would shift the boundaries between operating systems, platforms and payment networks. Historically, an operating system has helped users launch applications, while platforms have controlled important parts of discovery, identity and distribution. If users rely more heavily on agents, the new gateway may be the system that understands intent, connects to services and executes tasks. The companies controlling that gateway could influence attention, transaction flows and commercial rules.
The outcome is not predetermined. An agent must understand a request reliably, operate across multiple services and explain why it selected or rejected an option. It must also behave appropriately when money is involved. An agent that cannot execute tasks consistently will not be a dependable payment interface. An agent optimized only for completion rates, meanwhile, could create new forms of manipulation, unwanted purchases or fraud.
Digital assets may help settlement, but they are not the full answer
The article also considers the role of digital settlement assets. Marcus argues that these assets can serve as a way to express and regulate digital value transfer, and that they may be useful for internet-native settlement. But the asset used for settlement does not resolve the broader problems created by autonomous agents.
Whether funds are transferred through a bank account, a card balance or a digital settlement asset, a payment system still needs to determine who is paying, whom the agent represents, what authority the agent has, whether the transaction meets applicable requirements and what happens when the transaction is wrong. A digital settlement asset can change the settlement asset and the way value moves. It does not automatically provide identity verification, permission management, fraud prevention or consumer protection.
The source article points to emerging regulatory frameworks for digital settlement assets as examples of regulatory development. Such rules may provide greater clarity around issuers, reserves and compliance obligations. Once agents participate directly in transactions, however, regulatory attention may also need to cover agent identity, service-provider responsibility, transaction monitoring and the preservation of records across platforms.
The exact implementation of those requirements remains a matter for regulators and industry participants. Legal recognition of a settlement instrument should not be confused with a complete framework for automated commerce.
Implications for payment infrastructure and institutional wallets
For banks, payment companies, card networks and digital-asset infrastructure providers, agentic payments could expand the meaning of account security. The relevant control is no longer only whether an account is protected. It is also whether the agent has the right permissions for a particular action.
Systems may need policy engines that support granular limits, temporary authorizations, approval tiers, automatic suspension and detailed audit records. They may also need to distinguish between an agent that can read account information, one that can prepare a transaction and one that can submit or approve it.
Institutional wallets and custody systems face a related challenge. Existing controls often focus on protecting private keys, separating assets and limiting human operator access. If an agent can trigger or coordinate transactions, institutions will also need to define what the agent may inspect, initiate or approve, and which actions require human review or multi-party authorization. Automation does not eliminate governance; it requires governance rules to be expressed in advance and applied consistently across connected systems.
This is unlikely to be solved by one model provider or one payment company alone. Payment networks need interoperable approaches to identity, authorization and messaging. Merchants need a way to evaluate whether an agent is permitted to place an order. Financial institutions need continuous monitoring and escalation controls. Regulators need to clarify consumer protection and liability when an automated system acts without a person approving every step.
A redesign may be more realistic than a complete replacement
The claim that apps and buttons will disappear is best understood as a directional forecast rather than a literal prediction that all visual interfaces will vanish. Users will still need to review products, approve high-risk actions, resolve disputes and intervene when an agent behaves unexpectedly. The more plausible change is that a button will no longer be the only evidence of consent, and a screen will no longer be the only entry point into a payment process.
The next stage of development may therefore focus less on connecting an existing payment API to an AI system and more on building transaction infrastructure designed for machine execution. Agents must be able to act on a user’s behalf within explicit boundaries. Merchants must be able to verify an agent’s authority while retaining the ability to reject suspicious requests. Financial institutions must be able to monitor compliance without requiring a person to manually approve every low-risk action.
If those mechanisms emerge, agents could become a new coordination layer for payments and commercial services. If authorization, liability and dispute rules remain unclear, automation could transform visible human-confirmation risks into less transparent system-level risks. The central issue is not simply how money moves after the interface changes. It is who is authorized to decide, how that decision can be verified, and who is accountable for the result.
Source: link