On 20 October Visa, BILL and Billtrust announced what they call the first live pilot of trusted AI agents completing approved invoice payments end to end through supplier portals. The problem is mundane and expensive: an invoice has been approved, and someone in finance still has to log in to the supplier's portal and type the card details in. In the pilot, BILL's payment agent, registered through Visa's Trusted Agent Protocol and Agentic Directory, reaches the supplier's portal, where Billtrust recognises and authenticates it and lets it pay. It uses tokenised card credentials instead of shared portal logins, its controls are tied to the buyer's approved payment instruction, and each payment follows a path tied to a specific invoice, with both sides able to see its authorisation and completion.
The agent did not decide to pay. Someone approved the invoice, and the agent did everything after that. The approval is the asset.
The sentence that matters is in the release: the agent executes only eligible payments the buyer has already approved and does not make payment decisions itself. Earlier this month we asked who carries the loss when an agent buys the wrong thing. The first live business deployment answers by making the question small. The agent has little room to buy the wrong thing, because it does not choose.
The same day, The Block reported that Mastercard had opened early access to Agent Pay for Machines, with its first buyer and seller APIs, a developer toolkit and a sandbox, after end-to-end pilots with Cloudflare, Ant International, BVNK, AIsa, Base and Coinbase. In the demonstration it described, an agent working for a marketing agency bought research and eight campaign images in nine payments to two providers, all inside a budget the agency had approved, using Coinbase's x402 protocol. Mastercard's own description of the service, from its June launch, says organisations set authorisation rules and spending limits that are enforced programmatically, and that every agent is credentialed. The Block notes it is a pilot and not commercially available.
Put the two side by side and the design is the same. A person makes one decision, an approved invoice or an approved budget. Software does everything after it, at whatever speed and volume the work needs, and every step is identified, bounded and logged. That is not a co-pilot. The agent runs the function end to end. It is also not an agent with a wallet and its own judgement, which is the version most pitch decks draw.
Notice too who is in the middle. On the Visa pilot, the party that holds the approved payment instruction is an accounts payable platform that, by its own account, nearly 500,000 businesses use to manage their finances. On the supplier side, Billtrust says it reaches more than 20 million buyers. The networks supply identity and the rails. The approval itself lives in software the finance team already uses.
Business payments are going first for a reason: they already run on approvals. An approved invoice already has an approver, a supplier and an amount before anyone pays it. The agent inherits a mandate that already exists rather than needing a new one invented for it.
Sources
- 01Visa: Visa, BILL and Billtrust launch first live pilot for trusted AI agent payments between buyer and supplier platforms (2026-10-20)
- 02Fintech News Singapore: Visa, BILL and Billtrust launch AI agent pilot for B2B invoice payments
- 03The Block: Mastercard opens early access to AI agent payments following pilots with Coinbase, Base and Cloudflare (2026-10-20)
- 04Mastercard: Mastercard launches Agent Pay for Machines (2026-06-10)
Tell us what you are building. Every note gets a real read.