Visa, BILL and Billtrust pilot AI agent that pays invoices end to end
On Tuesday, October 20, 2026, two AI-agent payment setups went live in the market: a B2B pilot in which Visa, BILL and Billtrust let a registered agent complete approved invoice payments end to end, and an integration between Sierra and Stripe in which customers pay bills over the phone. These are the first concrete examples of agents moving real money — and both share the same structural trait: the approval sits with a human before the agent does anything at all.
How the invoice pilot works
The pilot was developed by Visa together with the payments platform BILL and the invoice automation company Billtrust, and was announced via Business Wire on October 20, according to Briefs and MSN.
The mechanics, as described in secondary reporting that reproduces the companies' own statements, run like this:
- An authorized BILL Payment Agent uses tokenized payment credentials — a digital substitute for the customer's payment details — tied to the buyer's instructions on which invoices to pay, and how.
- The agent is authenticated through Visa's Trusted Agent Protocol.
- Billtrust then uses Visa's technology to carry the payments all the way to completion inside the suppliers' portals.
BILL itself describes the setup as "the first live pilot that lets trusted AI agents complete approved invoice payments through supplier portals, without shared logins" — a company statement from the press release of October 21, not an independently verified fact.
On the supplier side, Billtrust's customers are to be able to recognize and accept payments initiated by trusted business agents, which the companies present as a way to reduce "trapped cash" — money sitting idle because the payment cannot be matched or processed. The payments are handled according to the suppliers' existing acceptance rules.
Mark Nelsen, the Visa executive responsible for products, commercial services and money movement, framed the pilot this way in a quote reported by Briefs: "This pilot applies agentic commerce to a real B2B-payment problem."
Why B2B first? Approval before agent
Why did the first live agent payments launch in invoice and bill workflows rather than in the consumer's shopping cart? That is the core of Forkast's analysis by Tessa Vaughn (October 22), and it is an analytical interpretation — not an independently verified finding — but it explains both deployments well.
In the invoice model, Forkast writes, "the approval exists before the agent shows up." The buyer has already defined the amount, named the invoice, and registered the agent's identity with the supplier. The agent does not make any new decision — it executes one already made. This is what allows the agent to act "without humans stepping in to authenticate," as Briefs puts it, without bypassing any control mechanism.
Consumer commerce lacks this head start. According to ChannelEngines' Marketplace Shopping Behavior Report 2026, as cited by Forkast, 58 percent of consumers have used AI tools to research products, but only 17 percent are comfortable purchasing directly through AI. The same report finds that 76 percent require a human review and approval before money changes hands. The figures are cited secondhand, and the underlying report is not available in this evidence base, but the pattern is clear: consumers research with agents, but will not let them pay.
Sierra and Stripe: the human in the payment moment
The same day, a phone-based integration between Sierra and Stripe also went live, according to Forkast. It follows the same logic through a different channel: at the payment moment, the agent switches to a secure mode in which the customer keys the card number on the phone's keypad. The keystrokes are routed directly to Stripe. The agent never sees the full card number — only the payment status and the last four digits.
Here too, human control sits within the payment act itself: the customer enters their own card details, and the agent is little more than a voice interface around a secure transaction.
Sierra states that 71.3 percent of calls in which a patient starts a payment with the agent Kora (from Cedar) end in a successful payment. This is a vendor-reported figure without published methodology, as Forkast itself points out, and should be read with that caveat.
What is company claims, and what is verified
Nearly everything mechanical in this story rests on the companies' own announcements, relayed through secondary reporting. There are no primary documents from Visa, Billtrust, Sierra or Stripe in the available evidence base — Visa's press release and the Sierra–Stripe announcement are available only via secondhand retellings.
BILL states in the press release that the company has surpassed $2 trillion in cumulative payment volume, and that the platform in the first half of 2026 stopped 9.64 million fraud attempts and saved customers $925 million in potential losses. These are vendor figures from a promotional press release, not independently verified numbers.
The background from Visa itself: earlier this year, the company said it would reduce its workforce by around 7 percent, and CEO Ryan McInerney said, according to Briefs, that AI is helping to accelerate how work evolves at the company. That places the pilot within a larger transformation at the payments company — but says nothing about the pilot's own scale.
Open questions
What is not documented is just as important as what is known. None of the sources state how many participants the pilot has, how long it will run, what transaction volumes it covers, or which industries the buyers and suppliers come from. It is impossible to say whether this is a broad rollout or a highly limited trial arrangement.
The second open question is structural: Can the "approval before agent" pattern be transferred to consumer commerce? In the B2B setup, there is already an invoice, an amount and an approval process for the agent to attach to. In a consumer shopping cart, all of this must be created in the moment — exactly where 76 percent of consumers, according to the ChannelEngine figures, require a human in the loop. If Forkast's interpretation holds, the barrier is not the technology but the ordering: the agent must come after the approval, not before it. Whether the consumer market can be built the same way — with pre-approved frameworks instead of free-roaming agents — is something none of the available sources answer.