Editor’s note (Oct 29, 2026): This article has been corrected. The original version carried causal claims not supported by the cited sources: that gated launches exist because liability is unallocated, that self-serve products launched precisely because liability was assigned, and the title’s assertion that the pilot wall is a liability wall. The sources show working capability and the presence or absence of published liability terms, but do not establish any causal direction. The claims – including the title, subtitle and final section – have been restated to the observable facts: what each product publishes, and where the gates sit.
October has been defined by a distinct pattern in agent-payments infrastructure. While the industry continues to struggle with the gap between announced intent and actual retailer readiness, as detailed in our retailer-readiness analysis, the launches that did arrive this month were almost exclusively placed behind pilot programs or sales gates. That piece measured the retailer side – who plans to open a catalog and who actually holds the data agents need. This one is the payments-rail side: who is letting an agent move money, and under what gate. Two patterns sit side by side: gated launches that are technically functional, and self-serve launches that opened without waiting for liability terms to be written.
Visa, BILL, and Billtrust launched the first live pilot for trusted AI agent payments between buyer and supplier platforms, utilizing tokenized credentials and controls aligned to pre-approved payment instructions. The system is functional, with the release FAQ explicitly stating that the agent does not make payment decisions but only executes eligible payments already approved by the buyer. Despite this operational success, the pilot remains closed, with no disclosed timeline for general availability or a published framework for liability assignment. Similarly, Mastercard opened early access for its Agent Pay for Machines at Cloudflare Connect. The program provides developers with APIs and a sandbox, and a demonstration at the event showed an agent successfully completing nine payments under a single approved budget. Yet, Mastercard maintains a strict disclaimer that the program is not available commercially, and enrollment is by request only, with no published criteria for general access.
PayPal’s Agent Ready page runs a hybrid approach for its agentic commerce services. While catalog exposure is self-serve for existing merchants, the broader onboarding process requires a direct conversation with a sales team. The barrier here is administrative, requiring businesses to navigate a formal inquiry process rather than accessing tools directly. American Express presents the most restrictive case, having announced its Business Playbook for Agentic Commerce in early October. The company’s proposed Agent Purchase Protection remains an announced intent rather than a functional product. Amex has explicitly stated that it intends to protect card members and merchants from agent-driven errors, but the technical specifications and the definition of a registered agent remain under development. The gate here is entirely structural; Amex is the party that would absorb the loss, and it has not yet defined the terms under which it is willing to do so.
What the gated launches have in common is observable: the technology works, and no liability allocation has been published. The Visa/BILL pilot is actively clearing transactions, and the Mastercard demo proved the capability to execute multiple payments within defined parameters. Neither release names a published framework for who absorbs the loss when an agent errs. Whether that absence is the cause of the gate or simply its most visible companion is not something these two releases settle – what they show is a working system behind a wall whose terms have not been written in public.
This reality is underscored by the contrast with three self-serve launches that opened on day one. Crossmint released its Agent Commerce Toolkit as a self-serve product; its terms do not address liability or error allocation at all. Robinhood has allowed agents to interact with its trading accounts since May, but its terms explicitly shift all risk to the user. Sierra and Stripe launched a phone-based payment system that allows agents to process payments without ever seeing card data, with Sierra reporting – vendor-reported, no published methodology – a 71.3% success rate for patient-initiated payments. These services launched without gates. Robinhood’s terms and Sierra/Stripe’s arrangement assign risk to the end user or merchant of record; Crossmint’s terms are silent on it entirely. In these three cases, open onboarding shipped alongside terms that either assign risk somewhere or say nothing about it – whether any of that produces open onboarding is not something the releases state.
Demonstrated capability in these launches is real but bounded by the conditions tested: the Visa/BILL pilot is clearing transactions within its pilot, and Mastercard’s demo executed nine payments under a single approved budget. Neither demonstration establishes that deployment faces no constraints beyond those tested. What none of these releases publishes is a standardized way to handle errors when machines act on behalf of humans. The pattern visible this month is straightforward: pilots and request-only access holding the field, with liability terms unpublished across the gated launches. Whether that pattern shifts when liability rules arrive is not a question these releases answer.
