AI Agents Now Pay in Minutes, But Who's Liable When They Buy the Wrong Thing?

AI Agents Making Payments at Unprecedented Speed

AI agents have crossed a critical threshold in autonomous financial transactions. According to Chainalysis data, the median gap between funding an x402 wallet and its first payment collapsed from 73 days in late 2025 to just 18 minutes by the third quarter of 20261

1

. Across more than 138 million transactions, software has been purchasing data, compute resources, and model access using stablecoins, with no person approving each individual payment1

1

. This rapid adoption of agentic AI in commerce marks a fundamental shift from AI systems that merely provide recommendations to those that take direct action in the real world.

The Revocation Problem Exposes Critical Gaps

While AI agent commerce has solved the technical challenge of enabling payments, a more fundamental problem has emerged around revocation of authority. Mark Zalan, CEO at GoMining, argues that most frameworks handle credential issuance well but fail at revocation1

1

. When a person cancels an agent's mandate, every counterparty must know immediately, or "a dead credential" continues spending until someone notices. Abdul Rafay Gadit, co-founder and CCO at ZIGFinance, frames mass adoption around this need, stating that people require "a budget they can enforce, permission they can revoke and a clear route to resolve mistakes" before trusting AI agents with daily spending1

1

.

Spending Limits and Cryptographic Spending Caps Emerge as Solutions

One proposed solution involves building authority directly into the payment mechanism rather than treating it as a registry record. Zalan describes cryptographic spending caps that prevent an agent from exceeding limits regardless of what its credential states or whether a revocation message has arrived1

1

. Gadit compares this model to "a corporate card with an enforceable spending policy," featuring transaction caps, daily budgets, approved recipients, and expiry dates enforced by the wallet or payment provider1

1

. Card-industry players like FIS have floated virtual cards issued specifically for agentic AI with restrictions by amount, merchant, and time window.

Legal and Financial Challenges Remain Unresolved

The liability for AI actions presents a fundamental legal question that existing frameworks struggle to address. U.S. consumer payment protections under Regulation E and Regulation Z both test actual, implied, or apparent authority, but neither was written for consumers who hand card credentials to software3

3

. A transfer made by someone the consumer granted access to can fall outside the definition of "unauthorized" altogether, meaning the liability caps consumers rely on may not apply3

3

. The Consumer Bankers Association has asked regulators whether sharing credentials with an AI tool triggers this exclusion, but the question remains open.

Micropayments Dominate While Larger Transactions Wait

Most AI agent payments currently remain tiny, with most x402 payments worth ten cents or less1

1

. A BeInCrypto Intelligence report gave institutional financial workflows a readiness score of just 12 out of 100, compared with roughly 61 for micropayments1

1

. The same report documented two incidents where a combined $281,200 went to attacker-controlled wallets before funds were recovered1

1

. Logan Xie, leader of KuCoin AI Lab, argues that a business system separate from the AI's reasoning should check each payment before execution, with high-risk actions requiring human oversight1

1

.

Regulatory Gaps Create Three Layers of Uncertainty

The regulatory gaps in consumer protections for AI transactions create uncertainty at multiple levels. At the consumer layer, Regulation E and Regulation Z carve out transactions made by authorized persons, leaving software operated by third-party companies in an awkward position3

3

. At the network layer, Visa and Mastercard's zero-liability policies for unauthorized transactions don't clearly address purchases an agent completed correctly against a mandate the consumer now disputes3

3

. At the merchant layer, chargeback rules allocate losses assuming a human either did or didn't make the purchase, leaving no clear framework for agents exceeding their mandate.

Fraud Risks Transform from Impersonation to Integration Failures

According to Accenture, 78% of payments leaders expect fraud to increase significantly as agentic payments scale3

3

. The threat model changes fundamentally: fraud stops being impersonation of a cardholder and becomes impersonation of an agent. As Hogan Lovells describes, once an agent is compromised, an attacker can move laterally across every merchant that agent transacts with, making one breach resemble an integration failure rather than one stolen card3

3

.

Accountability Frameworks Lag Behind Technical Capability

Zalan emphasizes that the limiting factor is not technical capability but legal and liability frameworks1

1

. A five-figure payment to an unknown party across jurisdictions "needs settlement you can rely on and recourse you can actually use," which remains unsolved. However, he expects the question of how people delegate spending authority to software to be settled in years rather than decades. Once a workable legal framework exists, he believes "the technical solution will be available on the market in less than six months"1

1

. He also expects a common standard for agent authorization to emerge within one to two years, shaped by whoever processes the most agent payments.

Enforceable Budgets and Human Oversight Define the Path Forward

Experts anticipate larger payments will start where rules and accountability frameworks are already clear. Xie points to approved business invoices as an early use case1

1

. Gadit sees treasury and rebalancing within set limits, where an agent carries out policy but "accountability stays with the people setting it"1

1

. Transaction counts have dominated debate, but Gadit suggests a different measure: "Repeat customers and useful services will tell us more about adoption than raw transaction counts"1

1

. The next measure of agentic commerce will be whether people can trust AI agents with larger sums, stop them when necessary, and recover money when something goes wrong.