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. This bifurcation suggests that the industry is not waiting for better code, but for a clearer understanding of liability.
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 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. (Mastercard’s own pages returned 403 to our fetcher; the disclaimer rides search-confirmed snippets of those exact pages)
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.
When examining these gated launches, a clear verdict emerges: the constraints are driven by risk and liability, not product capability. In each of these instances, the documentation confirms that the underlying technology is functional. The Visa/BILL pilot is actively clearing transactions, and the Mastercard demo proved the capability to execute multiple payments within defined parameters. Capability is never cited as the blocker. The gate exists because the network or issuer has not yet determined how to allocate the financial risk of an agentic error.
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, though it notably avoids addressing liability or error allocation. 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 were able to launch without gates precisely because the liability was already contractually assigned to the merchant of record or the end user. In these cases, the provider does not need to absorb the loss, so they do not need to gate the onboarding.
Ultimately, the industry is not waiting for agents to become more capable of making payments. It is waiting for a standardized way to handle the inevitable errors that occur when machines act on behalf of humans. Until a liability-assignment framework is established that allows networks and issuers to protect themselves without manual oversight, the gate will remain the primary feature of agent-payment infrastructure.
