Agentic Payments Expose Liability Gap as Card Network Rules Lag Behind AI Automation

When Bots Buy: The Liability Gap at the Heart of Agentic Payments

AI agents are booking flights, renewing subscriptions and completing purchases — autonomously, without asking for human sign-off on each transaction. The payments infrastructure those agents are using was not designed with them in mind, and the resulting legal vacuum is now a practical problem for every party in the transaction chain.

The core tension is structural. Card networks have not published rules specific to agent-initiated transactions, which means the frameworks that govern disputes, chargebacks and liability were written for a world in which a human being — or at least a clearly delegated human instruction — stood behind every card-present or card-not-present payment. They did not anticipate a software process acting repeatedly, at scale, on a stored credential.

Who Bears the Loss?

In the absence of agent-specific dispute rules from any card network, the default position — consistent with how existing chargeback frameworks operate — is that the merchant of record carries the loss when a disputed transaction cannot be defended. That outcome may be commercially tolerable in isolated cases, but it becomes structurally significant as agentic commerce scales.

The deeper difficulty lies one step upstream. Legal commentary reviewed for this article, including analysis published by Goodwin Procter via Mondaq and by Astraea Counsel, identifies a foundational ambiguity: a consumer's decision to authorise an AI agent does not automatically constitute authorisation of every individual transaction that agent subsequently executes. Whether a specific purchase counts as "authorised" under existing payment frameworks is, in the current absence of clear rules, genuinely open.

The consent a user gives to an AI agent at set-up may not, in regulatory terms, extend to each autonomous transaction the agent makes — leaving the boundaries of "authorisation" legally unresolved.

That ambiguity matters in both directions. A consumer disputing a transaction their agent made — perhaps because the agent acted beyond its intended scope, or simply made a choice the consumer would not have made — can plausibly argue the transaction was not individually authorised. Whether that argument succeeds under existing consumer payment rules is not settled.

The Mandate Problem

The scenario that troubles issuers most is the one in which an AI agent exceeds the boundaries of what the consumer intended. Commentary from Goodwin Procter and Astraea Counsel makes the point directly: there is no clear legal rule determining who bears the loss when an AI tool acts beyond its mandate. That gap is particularly acute for card issuers considering making credentials available to third-party AI tools — if the agent oversteps and a dispute results, the loss allocation is unresolved.

Agency-law principles, discussed in National Law Review commentary, add a further dimension. How concepts of authorisation and delegated action apply when the "agent" is software rather than a person, and when that software acts across many transactions rather than one, is not addressed by existing rules.

Issuers and acquirers currently navigating this environment are doing so by analogy — drawing on frameworks that may have partial relevance rather than frameworks designed for the purpose. That is a workaround rather than a solution, and its adequacy as a risk-management approach depends on whether regulators and card networks treat agentic transactions as functionally equivalent to existing categories.

Why the Gap Matters Now

The practical stakes are asymmetric. Merchants sitting as the default loss-bearer have limited ability to verify whether a transaction was within the agent's scope at the point of sale. Issuers extending credentials to AI tools face residual exposure if those tools act beyond their mandate. Consumers who dispute agent-initiated transactions may find that the very convenience of agentic payments — no per-transaction approval — weakens the clarity of their own authorisation chain.

No card network has yet published rules that address these questions directly. Until that changes, every party processing agent-initiated transactions is operating in a framework built for a different era of payments — and bearing whatever risks that mismatch creates.