Visa reported nearly $11 billion in value-added-services revenue in fiscal 2025. Its new pilot with BILL and Billtrust takes that strategy to the last manual step in many B2B card payments: getting an approved invoice paid inside a supplier portal.
The October 20 announcement describes a narrow but important job. A buyer has already approved an invoice. BILL’s payment agent must get into a supplier’s Billtrust portal and pay it by card. Historically, that last step has involved a person logging in and typing payment details into fields designed for a human. The invoice may be approved and the card available, but the payment is still stuck behind a login screen.
According to the three companies’ announcement, the BILL agent is registered using Visa’s Trusted Agent Protocol and Agentic Directory, carries tokenized payment credentials and follows controls tied to the buyer’s approval. Billtrust recognizes it at the portal and applies the supplier’s acceptance policy. Visa’s developer documentation explains the underlying trust mechanism: the agent signs a request, and the receiving site verifies the signature against its public key. The pilot is live with participating buyers and suppliers, but the companies have not disclosed participant counts, payment volume, completion rates or a broad commercial rollout.
That distinction matters. The agent is not deciding whether the buyer owes the money. It is executing a payment the buyer already authorized.
The service is access, not just a better card
Visa has spent years selling services around its network: fraud controls, acceptance tools, issuer technology, data and advisory. Its 2025 annual report says those businesses generated nearly $11 billion in revenue. This pilot extends the same logic to a new problem. When software, rather than a person, arrives at a supplier portal, somebody has to establish who sent it, what it may do, which payment credential it may use, and how both sides can see what happened.
Those are services that sit beside the transaction. The announced pilot demonstrates one controlled path between BILL and Billtrust. My inference is that Visa wants to make the same identity and control layer useful across many buyer platforms and supplier portals, giving it a position before authorization reaches the card network. The harder thing to copy is not the model that fills out the form. It is the trusted connection between the buyer’s instruction and the supplier’s acceptance rules.
I have argued before that the AI interface does not have to own the underlying financial relationship. This is the payments version of that argument. The interface may be an agent, but the durable value can remain with whoever supplies identity, permissions, credentials and a path to settlement.
The startups are already in the room
Sunlight markets an agentic card-payment engine that identifies card-eligible invoices and executes payments through supplier portals without requiring suppliers to join a new network. That is close to the work this pilot addresses. Natural is building a broader stack of payment primitives for agents, including identity, observability, wallets and processing. It is not the same supplier-portal product, but it is contesting part of the trust and infrastructure layer Visa wants to serve.
This is direct competitive overlap, not proof that the companies are commercial adversaries. A startup can supply execution inside an incumbent’s distribution, or an incumbent can absorb the capability and make the startup a feature. Those are very different businesses to underwrite.
Startups have a legitimate opening in the messy work of navigating thousands of different portals. Visa has existing payment credentials, risk infrastructure and institutional relationships. Billtrust, however, controls the supplier portal in this pilot. Other portal owners can choose which agents to admit and on what terms. Mastercard is building agent recognition and trust services too. No one has won a standard merely by announcing a pilot.
The difficult payment starts before the agent reaches the button
A utility is a better stress test than a single invoice in a cooperating portal. The buyer may operate many sites, each with its own utility account, service address and bill. The agent has to identify the right obligation, determine which portal holds it, match the amount and due date to the buyer’s approved instruction, and confirm that the payment actually posted against that account. A successful card authorization is not proof that the utility applied the money to the right meter or service address. An ambiguous confirmation creates two bad choices: retry and risk paying twice, or stop and risk leaving a bill unpaid.
Multi-location suppliers produce a similar problem with different identifiers. Cintas asks customers for account numbers to register for billing. FedEx lets customers find a bill by invoice number, while its billing tools can filter invoices by account. Cisco’s Commerce Express guide walks the user through selecting an invoice before choosing a card. These are examples of different portal workflows, not evidence that any of the three companies has joined Visa’s pilot. For a buyer with multiple locations, the hard part is maintaining the mapping among vendor, location, account, invoice and approved payment. A bot that gets through the login screen but cannot establish that mapping has automated the wrong step.
The failure state matters as much as the happy path. The agent needs a rule for a partial payment, a changed balance, a card surcharge that violates buyer policy, a portal that times out after submission, and an invoice already paid outside the portal. Each exception needs an owner and a record of what was attempted. Otherwise the AP team inherits a faster version of the same swivel-chair reconciliation it was supposed to escape.
Will vertical specialists emerge?
Yes, if they own completion rather than just navigation. A utility specialist could maintain account and service-address mappings, detect billing changes, monitor payment status and handle the exceptions unique to that category. A specialist serving multi-location national vendors could do the same for location IDs, account hierarchies and invoice-level remittance. That is a product with a measurable unit: correctly completed payments per dollar of operating cost, including human rescue and duplicate-payment losses. It is more defensible than a generic agent clicking through a page.
But specialization does not automatically mean a standalone payments network. Visa or Mastercard can provide identity and credentials; an AP platform can distribute the workflow; a vertical operator can supply the data, portal coverage and exception handling. The specialist could be a supplier to the incumbent rather than its replacement. The question for a founder is whether the proprietary asset is reusable across customers in that vertical, or whether every new portal and buyer requires another custom implementation.
For an investor or a CFO, the tests now become more demanding. First, does the agent complete approved payments across unrelated portals and multi-location accounts, or only inside one cooperative environment? Second, do suppliers accept the card economics, including interchange and any surcharge, once the novelty wears off? Third, who can enroll or revoke an agent, and who pays when the agent submits the wrong amount or cannot reconcile an exception? I would ask for invoice-match accuracy, confirmed completion, duplicate-payment, and human-rescue rates by vendor category. The announcement supplies none of those numbers or operating rules.
I wrote earlier about the fragmented virtual-card acceptance problem. Agentic portal payments attack one piece of it: the last manual step between an approved invoice and a supplier who will accept the card. If portal owners adopt competing identity standards, if suppliers reject the card economics, or if completion rates collapse outside Billtrust, this remains a useful integration rather than a Visa-wide service. If the flow works across utilities and complex national vendors, Visa’s value-added-services strategy moves closer to the actual work of paying a supplier. Specialists who solve those categories may capture a valuable layer of their own.
The payment is not finished when the agent clicks submit.