Agentic commerce and the battleground for new payments infrastructure – Bank Underground

Prem Munday

Agentic commerce, where artificial intelligence (AI) systems act on behalf of users to find products, negotiate purchases, and execute payments, is developing rapidly. This creates shared responsibility: developers must build legally sound systems, while regulators and infrastructure operators must consider how existing frameworks apply and where new approaches may be needed. The Bank of England operates, oversees and is co-ordinating the design of payment systems as part of its statutory responsibilities. Emerging agent‑based payments can have implications for how the private sector safely innovates and how regulators and payment infrastructure providers adapt. This post explores how agentic commerce could reshape future payment design.

How might agentic commerce be used in practice?

It is important to note from the outset that agentic AI always requires human deployment, and that deployers retain legal responsibility for an agent’s actions; responsibility does not sit within a ‘black box’.

Visa provides one example from industry for how agentic commerce could be used, setting out a four‑step adoption process:

• Recommending products: using Large Language Models (LLMs) to recommend better products. An agent could compare products and recommend the most suitable option.
• Initiating payments on your behalf: agents can make payments with user verification, such as one-off bill payments.
• Transacting on your behalf: agents execute payments according to predefined rules, such as renewing a service when usage hits a threshold.
• Orchestrating payments: an agent owns the whole lifecycle of payments and communicates with other agents to orchestrate complex payment flows.

This last point leads to a potential scenario where agentic payments become like ‘locals paying at bazaars’, with agents forming informal relationship-based agreements with other agents. This highlights a future state where agents might adapt the behaviours we see in payments, with potential downstream impacts:

• Payments move from being human-initiated to agent-initiated.
• There is an increase in speed and volume of transactions as agents may transact, negotiate, return and refund payments at speeds faster than humans.
• There are decreased transaction sizes, as agents may transact in small values to complete complex, orchestrated workflows.
• We need new authentication to resolve how humans and their agents interact, moving from Know-Your-Customer (KYC) to Know-Your-Agent (KYA) for payments, as highlighted by Dave Birch.

While some automated activity exists today in areas such as algorithmic trading, consumer and retail payments introduce distinct requirements around authentication, liability and consumer protection.

So how do payments and agents interact, and what are the responses to this?

A previous post examined how existing financial infrastructure can govern agents. I am developing this by highlighting how the infrastructure for managing agents can impact how payment systems are built.

Today’s agent payments landscape is fragmented, with multiple identity frameworks, payment protocols and communication layers that are not interoperable across providers. For example, some agent identity standards are only supported by specific card schemes, while agent payments protocols and how they integrate with checkouts vary across stablecoin and card‑based rails. Addressing this fragmentation is a shared task: the private sector needs to build and adopt interoperable standards; with public sector participants having a role in setting clear expectations and, where appropriate, common requirements.

We are already seeing new private sector solutions to solve the issues around fragmentation, standards and interoperability with different payments methods. These solutions tend to cover four aspects: how agents communicate with each other, how they pay, how they assure identity and how they settle payments.

• How agents communicate: New shared standards are emerging that allow AI agents to exchange information and instructions with each other. Examples include the Model Context Protocol (MCP) and Agent2Agent (A2A) frameworks. Think of these as like a common language that different agents can use regardless of who built them.
• How agents pay: New protocols are being developed to define how agents interact with online checkouts and payments processes. Examples include the Agentic Commerce Protocol (ACP), Universal Commerce Protocol (UCP), and Agentic Payments Protocol (AP2). These are the equivalent to giving agents a standard way to navigate the payments processes, like authentication of your card, that humans currently do manually.
• How agents prove their identity: For payment systems to trust an agent, they need a reliable way to verify who or what is acting. Card schemes are developing their own solutions (such as Visa Intelligent Commerce and Mastercard’s Agent Pay). Some solutions have also been developed by users for specific blockchains, such as ERC-8004 for Ethereum. The challenge is that these approaches are not consistent with each other.
• Which payment rails agents use: A payment rail is the underlying infrastructure that moves and settles money from one party to another, such as card networks, Faster Payments or blockchain-based systems. Agents will need to connect to these rails to complete transactions. Both established card providers and newer blockchain-based options (like the X402 protocol) are developing ways to accommodate agent-initiated payments.

These innovations highlight a future route to solving issues with how agents and payments infrastructure work together; but there still may be issues that arise that require further integration with payments infrastructure, new standards or reimaging payments infrastructure we have.

So, what are the design challenges for building payment rails that work with agentic payments? Some of these fall on private sector designers; others raise questions for regulators and infrastructure operators.

The potential use-cases and private sector innovations bring to the fore a few questions I grapple with when thinking about designing future payments infrastructure. These are:

• How to ensure consistent identity and authentication across human and agent actors.
• Whether payment systems should support higher‑frequency, lower‑value transactions.
• How deterministic requirements in payment law can be upheld when interacting with non‑deterministic AI systems.
• How regulation can encourage interoperability and enable integration between competing standards.

On the first issue, one question when integrating agent identity into payments is what role there should be for a central entity to mandate agent identity, like the conduct principles around how the private sector is required to implement KYC.

On the second issue, agents might need faster, lower value and high-volume payments. We need to consider if the existing rails can support these transactions or if new ones are needed. This also raises a larger infrastructure point: do payment rails need to be designed and built from the start to incorporate the ways mentioned above on how agents communicate, pay, and prove identity, as opposed to these being bolted on afterwards?

Thirdly, payment systems are designed to be deterministic: given the same input, they produce the same result. This predictability underpins reconciliation, fraud controls and legal certainty. Agentic systems rely on probabilistic AI outputs. An agent may phrase requests differently, pursue alternative paths to achieve a goal, or retry transactions in unexpected ways. This mismatch creates risks. Agents could generate excessive payment requests, submit non‑standard data, or trigger unintended transactions. Payment rails therefore need guardrails, clear policies and the ability to detect erroneous agent‑initiated activity. Designers of agentic payment systems – public and private – will need to manage this variability while preserving the predictability required for settlement. Regulators may also need to consider appropriate safeguards and standards.

Finally, agents will require a universal way to interact with online checkouts and allow interoperability. Today, each checkout journey varies by merchant, payment service and rail. For agents to participate meaningfully in commerce, we will need a layer of abstraction that allows them to complete checkout flows regardless of whether the underlying rail is Visa, Mastercard, Faster Payments, or emerging options like stablecoins. This means designing rails that provide interoperability with agent identity, payments protocols and communications standards so payment systems can interact with the many private sector frameworks that might get adopted. This also raises the question of to what extent a central authority should be the standard setter for agentic payments and commerce, to better enable this innovation.

These design choices have direct implications for how policymakers and payment system designers approach future infrastructure. While acknowledging payments sits in the context of a broader ecosystem where agents would not just interact with payments through the underlying rails but also via intermediaries (eg wallets, checkouts etc), there is an understanding that different payment technologies have different strengths in an agentic context. Blockchain-based forms of money, including stablecoins and tokenised deposits, can support programable, rule-based payments and small transaction values and flexible automated workflows. Existing card-based rails benefit from broad acceptance and established consumer protections. It is important that payment system builders, be it in the public or private sector, choose technologies and design them in ways that meet appropriate safety and resilience standards.

The broader challenge, for the public and private sector, is to determine how existing payments infrastructure can be adapted for agentic use, and where genuinely new approaches may be needed. In some cases, existing infrastructure may be sufficient; in others, new technologies such as blockchains could enable step changes in how agents, payments and commerce interact. Acknowledging these decisions helps us understand how to develop payment systems that remain trusted and fit for purpose in an economy where agentic payments may grow.

Prem Munday works in the Bank’s Distributed Ledger Technology Lab.

If you want to get in touch, please email us at bankunderground@bankofengland.co.uk or leave a comment below.

Comments will only appear once approved by a moderator, and are only published where a full name is supplied. Bank Underground is a blog for Bank of England staff to share views that challenge – or support – prevailing policy orthodoxies. The views expressed here are those of the authors, and are not necessarily those of the Bank of England, or its policy committees.