Who Pays When an AI Agent Makes the Wrong Purchase?

The Payment Rails Are Ready for AI Agents. The Law Isn't. TL;DR: In a single week in March 2026, Stripe's Machine Payments Protocol went live, Mastercard agreed to acquire stablecoin firm BVNK for $1.8 billion, and Visa launched a command-line tool for AI agent payments. Every major payment rail is now being rebuilt for AI agents. But when your agent buys the wrong thing — or buys the right thing at the wrong price, or spends beyond what you intended — the law has no clear, tested answer for who absorbs the loss. The infrastructure is racing ahead of the legal framework. Nobody is talking loudly enough about what happens when the agent is wrong. Let me describe something that is no longer hypothetical. You authorize an AI agent to handle your weekly grocery order — $150 budget, "buy the usual." Overnight, the agent interprets that instruction liberally, catches a flash sale on a kitchen appliance you browsed six weeks ago, and buys that too. Total: $340. You didn't authorize that purchase. But you authorized the agent. So — did you authorize it? That's the liability gap sitting underneath one of the fastest infrastructure builds in payments history. What just got built Over the past 14 months, the entire payments stack got rebuilt for AI agents, quickly and largely under the radar. Mastercard unveiled Agent Pay on April 29, 2025, introducing Mastercard Agentic Tokens — an extension of its existing tokenization stack (the same rails that power contactless payments and card-on-file) that lets a verified AI agent transact on a consumer's behalf [1]. Visa followed the very next day, April 30, 2025, with Intelligent Commerce, its own suite of APIs replacing card details with tokenized credentials scoped to specific agents and consumer-set spending limits [2]. Then, in a single week in March 2026: Mastercard announced a definitive agreement to acquire stablecoin infrastructure firm BVNK for up to $1.8 billion on March 17 [3]; Stripe and its blockchain partner Tempo launched the Machine Payments Protocol (MPP) on March 18, an open standard letting AI agents pay for services autonomously, with Visa extending it to support card payments [4]; and Visa Crypto Labs unveiled a command-line interface for AI agent payments on March 19 [5]. Three moves, three days. Stripe has since positioned itself as connective tissue across the ecosystem, supporting both Mastercard Agent Pay and Visa Intelligent Commerce. The build is real, it's live, and it's moving faster than most people in financial services have processed. The authorization problem Here's the part that should make compliance officers uncomfortable. Regulation E — the US federal rule governing electronic fund transfers — was written for a world where humans authorize transactions. It defines an unauthorized transfer as one initiated by someone other than the consumer, without actual authority, from which the consumer gets no benefit; once a transfer is reported as unauthorized, the consumer's liability is capped and the institution must investigate [6]. That framework is clear when a fraudster steals your card: unauthorized, capped liability, bank investigates and reimburses. Straightforward. A mis-scoped purchase from an agent you did authorize doesn't fit that definition cleanly. The consumer arguably did benefit in some sense — the agent acted within its granted access — even if the specific purchase wasn't what was intended. That's newer ground for the recourse machinery, and Regulation E's binary of "authorized" versus "unauthorized" doesn't obviously resolve it. The same structural gap shows up in the UK, where a comparable consent requirement exists: a payment transaction is authorized only where the payer has consented to that specific transaction, and unauthorized transactions must be refunded [7]. Legal commentary on that framework makes the liability-allocation problem explicit: existing rules don't allocate liability between the payer's provider, a payment initiation service provider, and the AI technology provider when an autonomous agent — not the payer — decides the specific transaction [7]. The question regulators on both sides of the Atlantic haven't fully answered: when you grant an agent delegated spending authority and it exceeds the scope of what you intended — not what you technically specified, but what any reasonable person would have understood — who eats the difference? Consent law was designed for humans approving specific transactions. It was not designed for continuous delegation to an autonomous system making many micro-decisions on your behalf. That gap is structural. The liability chain nobody has drawn cleanly When something goes wrong, the chain includes you, the agent platform, the payment network, the processor, the merchant, and potentially a sub-agent the original agent delegated to without your knowledge. That last one is a recognized problem in AI security circles: OWASP's LLM Top 10 names "Excessive Agency" as a core vulnerability class, covering cases where an agent has been granted more functionality, permission, or autonomy than its task requires — including delegation chains where a sub-agent has no direct relationship with the original user and only knows what the parent agent told it [8]. The clearest legal answer so far comes from the UK's Competition and Markets Authority (CMA). Its March 9, 2026 guidance on AI agents and consumer law states plainly that a business is responsible for what its AI agent does in the same way it would be for an employee's actions — including where a third party designed or provided the agent [9]. Most companies building on top of agent platforms haven't fully internalized what that means for them. What's actually being done about it Not nothing, to be fair. Mastercard's tokenization framework lets consumers set spend caps, merchant restrictions, and expiration windows at the moment they grant an agent access, and authorization is revocable in real time [1][10]. That's a real technical safeguard against clear overreach. But it's not a legal framework for the grey zone: the agent that acted within its technical parameters but outside your reasonable intent. The FCA's Payments Regulatory Priorities Report, published March 25, 2026, signals that UK regulators will actively consider whether the payments regulatory framework needs to change to accommodate autonomous agents initiating transactions — a step beyond the regulator's usual approach of applying existing rules to new technology without amendment [11]. In the US, the posture so far has been to apply existing consumer protection law to AI systems rather than write new rules — better than nothing, but it still leaves the core consent problem unaddressed. The honest state of play The infrastructure is impressive. The liability framework is not. If your AI agent makes a purchase you didn't sanction, your best recourse right now is a chargeback process designed for human-initiated fraud — which this isn't — or a dispute with a platform whose terms of service limit their exposure. McKinsey's research puts the scale of what's coming in perspective: under a moderate adoption scenario, agentic commerce could orchestrate $900 billion to $1 trillion in US B2C retail revenue by 2030, with global figures reaching $3 trillion to $5 trillion [12]. Consumer adoption is already moving faster than that infrastructure timeline suggests: a TD Bank survey found that 55% of Americans now use AI to help manage their finances, up from just 10% a year earlier [13]. The adoption curve is not waiting for the legal framework to catch up. That gap is exactly where the first major consumer harm incident will originate. Every company shipping agentic commerce features without asking "who pays when the agent is wrong" will wish they'd had that conversation sooner. The rails are ready. The answer to that question is not. Sources Mastercard. "Mastercard unveils Agent Pay, pioneering agentic payments technology to power commerce in the age of AI." Press release, April 29, 2025. Visa. "Find and Buy with AI: Visa Unveils New Era of Commerce." Press release, April 30, 2025. Mastercard. "Mastercard to acquire BVNK to connect on-chain payments and fiat rails." Press release, March 17, 2026. Stripe. "Introducing the Machine Payments Protocol." Stripe blog, March 18, 2026; see also Fortune, "Stripe-backed crypto startup Tempo releases AI payments protocol, launches blockchain," March 18, 2026. Reporting via coinalertnews.com aggregation of financial news wires, "Visa Crypto Labs Unveils CLI for AI Agent Payments," March 19, 2026. Consumer Financial Protection Bureau. 12 CFR § 1005.6, "Liability of consumer for unauthorized transfers" (Regulation E), and § 1005.2(m) definition of unauthorized EFT. Bratby Law, "Agentic AI Payments: Consent, Compliance and UK Law," analysis of the FCA's 2026 Payments Regulatory Priorities Report and the Payment Services Regulations 2017 (regulations 67 and 76). OWASP GenAI Security Project. "LLM06:2025 Excessive Agency," OWASP Top 10 for LLM Applications. UK Competition and Markets Authority. "Complying with consumer law when using AI agents," guidance published March 9, 2026. Mastercard. "Agentic token framework: Driving trusted AI transactions," Mastercard Global. FCA. "Payments Regulatory Priorities Report 2026," published March 25, 2026, as summarized by Browne Jacobson, "FCA payments regulatory priorities 2026: What to do now." McKinsey & Company. "The agentic commerce opportunity: How AI agents are ushering in a new era for consumers and merchants." October 2025. TD Bank. "Second Annual U.S. AI Insights Report." Press release, March 31, 2026. Disclosure: I used an AI assistant for research synthesis and copyediting. Analysis, framing, and all final wording are my own.