What it is
“Who counts as the authorising party” is shorthand for a gap in payments law that has opened up as AI agents begin making purchases and transfers on people’s behalf. Payment systems — credit cards, bank transfers, commercial wires — were built around a simple assumption: a human initiates a transaction, and that act of initiation is what “authorization” means. AI agents break that assumption. A consumer might tell an agent “book me a flight under $400 to Chicago next week,” and the agent then searches, selects, and pays without further human review. The question the payments industry has not settled is whether that payment was authorized by the consumer (who gave a general instruction), the agent’s developer (who built the decision logic), or no one in the legally meaningful sense at all.
This is not a hypothetical. Existing US consumer-protection law was not written with autonomous intermediaries in mind, and the two main statutes that govern payment authorization point in different, awkward directions. Under the Electronic Fund Transfer Act and its implementing Regulation E, a bank transfer is presumed authorized once a consumer hands over credentials to an agent, even if that agent then acts outside what the consumer actually intended. Under the Truth in Lending Act and Regulation Z, a credit card charge is authorized if the party who initiated it had “actual, implied, or apparent authority” — a standard broad enough to cover an agent that has technically overstepped its instructions. Neither statute was drafted to answer whether an AI system acting on broad, standing permission can authorize a specific transaction the way a human swiping a card does (Goodwin Law, June 2026).
The practical effect, according to legal analysis from Goodwin Law, is that losses tend to fall on whoever has the weakest legal footing to resist: consumers may find they are liable for mistakes their own agents make, merchants absorb a rising tide of chargebacks dressed up as authorization disputes, and the agent platform itself — the company that built and ran the AI — often bears the least exposure of all, because no statute currently says the AI provider is responsible if its agent completed a purchase badly.
Why it matters for AI governance and narratives
The authorising-party question sits at the center of a broader contest over how much autonomy AI agents should be given in consequential, real-world domains, and who answers for the outcome when that autonomy goes wrong. Industry framing — from payment networks eager to capture a share of agent-driven commerce — tends to present the gap as a solvable engineering problem: build a cryptographically verifiable “chain of authorization,” and trust follows. That framing is doing real work in the narrative, because it reframes a liability and accountability question as a protocol question, implicitly settled by whichever company’s standard wins adoption rather than by regulators or courts. The absence of government action on this specific question — while plenty of government attention goes to more visible AI risks — is itself a signal about where regulatory capacity is and is not currently being deployed.
Key facts and dates
The technical response has moved faster than the legal one. Mastercard launched Agent Pay on April 29, 2025, and Visa announced Visa Intelligent Commerce the next day; Visa’s more specific Trusted Agent Protocol followed on October 14, 2025, with early partners including Stripe, Shopify, Microsoft, and Coinbase. Google introduced its own open standard, the Agent Payments Protocol (AP2), in September 2025, built around cryptographically signed “Intent,” “Cart,” and “Payment” mandates intended to create a verifiable record of what a user actually authorized an agent to do. In May 2026, Google and Mastercard contributed AP2 and a related “Verifiable Intent” standard to the FIDO Alliance, aiming to make the authorization framework platform-agnostic rather than owned by any single company. These protocols create a technical record of authorization — a signed trail showing what instruction a user gave and when — but none of them resolve the underlying legal question of who is liable when that trail shows the agent executed the user’s instruction poorly, or a bad actor compromised the agent itself. Commercial payments, by contrast, already have a workable model: under UCC Article 4A, a bank and a business customer can agree in advance on a “commercially reasonable security procedure,” and a payment order that follows it binds the customer even if it later turns out to be unauthorized. No equivalent, clearly-settled framework yet exists for consumer transactions routed through an AI agent.
Where to learn more
- Authorizing Agentic Payments — Goodwin Law (June 2026) — legal analysis of EFTA/Reg E, TILA/Reg Z, and UCC Article 4A as applied to agent-initiated payments
- Announcing Agent Payments Protocol (AP2) — Google Cloud Blog — primary source on the mandate-based authorization standard
- Visa Introduces Trusted Agent Protocol — Visa Newsroom (Oct. 14, 2025) — primary source on Visa’s merchant-facing agent-verification framework
- Google donates Agent Payments Protocol to FIDO Alliance — Google Blog (May 2026) — primary source on the move toward a platform-agnostic standard