Visa, BILL and Billtrust have launched what they describe as the first live pilot in which an AI agent completes approved business invoice payments end to end through supplier portals. Announced on 20 October 2026, the pilot has BILL's Payment Agent prove its identity to Billtrust-run portals using Visa's Trusted Agent Protocol, then pay with tokenized card credentials, with no shared portal login and no CAPTCHA. The agent only executes payments a buyer has already approved; it does not decide what gets paid.
What Did Visa, BILL and Billtrust Announce?
Visa's press release, datelined Las Vegas on 20 October 2026, says the three companies "have launched the first live pilot to enable trusted AI agents to initiate and complete approved buyer invoice payments" through supplier portals. The release states that the pilot "is currently live and processing end-to-end transactions."
The problem it targets is narrow and familiar to any accounts payable team. An invoice can be approved inside a buyer's AP software, yet the last step still needs a person to log into the supplier's payment portal and key in card details by hand. Visa's release says many portals have required human authentication, which has capped how far approved payments can be automated.
BILL is the buyer-side platform. The release describes it as an AI-enabled financial operations platform covering payables, receivables, expenses, forecasting and procurement, and states that "BILL's platform processes ~1% of US GDP annually." Billtrust runs the supplier side, the portals where invoices get paid. Visa supplies the agent identity layer, the tokenized credentials and the payment controls.
| Party | Side of the transaction | What it contributes to the pilot |
|---|---|---|
| BILL | Buyer (accounts payable) | The BILL Payment Agent, registered with Visa, which submits buyer-approved payments |
| Billtrust | Supplier (accounts receivable) | Portals that recognise and authenticate the agent instead of demanding a human login |
| Visa | Network | Trusted Agent Protocol, Agentic Directory registration, tokenized card credentials, payment controls |
Source: Visa press release, 20 October 2026.
How Does an AI Agent Get Past a Supplier Portal?
The flow runs in two halves, according to Visa's release. On the buyer side, the BILL Payment Agent is registered through Visa's Trusted Agent Protocol and Agentic Directory, and uses tokenized payment credentials with controls aligned to the buyer's approved payment instructions. On the supplier side, Billtrust uses the same protocol to recognise and authenticate that agent. Visa says this authentication "can satisfy" the protections that CAPTCHA and other bot-blocking tools have provided, without a separate agent login.
Trusted Agent Protocol (TAP) is Visa's specification for letting a website verify that an incoming automated request comes from a registered, legitimate agent. Visa's developer page describes it as "a verifiable, cryptographic standard" in which the agent attaches a message signature to its request and the merchant validates it against the agent's public key, citing RFC 9421. Cloudflare's October 2025 write-up explains the lineage: TAP and Mastercard's Agent Pay both "leverage Web Bot Auth as the agent authentication layer," which uses HTTP Message Signatures, and Visa's version adds a tag that tells the site whether the agent is browsing or paying.
The Agentic Directory is the registry side of the same idea. Visa introduced it on 10 June 2026, describing it as a list of "agents and merchants that Visa has verified as legitimate participants in agentic commerce."
Put together, the portal is no longer asking "is this a human?" It is asking "is this a registered agent, acting for an identified buyer, within limits that buyer set?" That is a different security question, and for a portal operator it is arguably a better one: a CAPTCHA proves very little about authority to pay. One caveat: TAP verifies that the agent is who it claims to be. How Billtrust then ties that agent to a specific buyer's account on the supplier side is not described in the release.
Why Start With B2B Invoices Instead of Consumer Checkout?
The release answers this directly in its FAQ: B2B payments "already operate within established approvals, controls and the buyer's rules." That one line explains why this use case suits a live pilot better than consumer checkout does.
In consumer agentic commerce, the hard problem is intent. An agent buying a jacket has to infer what the shopper meant, and when it gets that wrong, the question of who carries the loss is still unsettled. Our analysis of the agentic checkout trust gap covers how shoppers feel about letting agents pay.
Accounts payable has already solved intent. An invoice exists, it has been matched and approved by a named person inside an approval workflow, and the amount and payee are fixed. The agent in this pilot is not shopping. It is carrying out an instruction that a human already signed off. Visa's FAQ is explicit that the agent does not make payment decisions and executes only eligible payments the buyer has approved.
The conventional framing treats B2B as the slower, more conservative market that will adopt agent payments after consumers prove them. This pilot points the other way. The industry analysis case is that agent payments go live first wherever human approval already exists as structured data, and accounts payable is the clearest example of that.
What the Announcement Does Not Disclose
A "live pilot" can mean a handful of test payments or a meaningful share of a platform's volume. The release does not say which. These are the gaps, checked against the Visa release and Visa's follow-up commercial payments release of 21 October.
| Question a finance or payments team would ask | Disclosed? |
|---|---|
| How many buyers and suppliers are in the pilot | No |
| Number or value of payments processed so far | No |
| When the flow becomes generally available | No; the release says the companies "will continue exploring opportunities in agentic payments" |
| Pricing or fees for buyers or suppliers | No |
| Payment rails beyond tokenized card | No; the flow described is card only |
| How the portal maps a verified agent to the right buyer account | Not described; the release says only that authenticated agents act for identified buyers |
| What happens when a payment fails or a card is declined | Not described |
| Whether other AP platforms or portals can use the same flow | Not stated |
None of this makes the pilot less real. It does mean the article-worthy fact is the architecture and the live status, not scale. Anyone citing this pilot as proof that agent payments are mainstream in B2B is reading more into it than the companies have said.
The Card Question: Who Pays for the Automation?
The pilot runs on card. Visa's FAQ frames one benefit as letting buyer platforms extend card payments to "suppliers that historically supported automated payments primarily through ACH or check," and says supplier platforms may improve Days Sales Outstanding (DSO), the average time it takes to collect payment after an invoice.
That benefit has a cost attached that the release does not discuss. Billtrust's own guidance for suppliers notes that card payments "typically require the merchant (the supplier) to pay an interchange fee on every single transaction," and suggests suppliers use payment policies to "steer customers toward more cost-effective payment methods like ACH."
So the pilot sits on a real tension. Faster collection and less manual work for both sides, paid for by a card acceptance fee that falls on the supplier. Whether suppliers accept that trade will depend on their margins and on how much the speed is worth to them, and the pilot's terms on this are not public.
Visa's interest in the card side is not hidden. Its 21 October release on commercial payment capabilities lists the BILL and Billtrust work as "a live agentic pilot" alongside new supplier-validation and virtual-card authentication tools, which Visa says are in pilot with general availability expected in 2027. The same release says its supplier analytics tool has identified "more than $600 billion in potential cardable spend." Agent payments, in Visa's framing, are one more way to move invoice spend that currently travels by ACH or check onto card rails.
What AP Teams and Portal Operators Can Check Now
The pilot is closed, but the pattern it sets will reach other platforms. A practical readiness checklist follows directly from how the flow works:
- Find out what your portal does with automated traffic. If a supplier portal blocks all non-human requests by default, a registered agent will fail at the same wall as a scraper. The question to ask your portal provider is whether it can verify signed agent requests (TAP or Web Bot Auth) rather than only challenge them.
- Map where human approval lives today. The pilot works because approval happens before the agent acts. If your approval step is an email thread rather than a recorded workflow state, an agent has nothing reliable to execute against.
- Decide your card acceptance policy deliberately. For suppliers, agent-initiated card payments will arrive with interchange attached. Know which invoice sizes you will accept on card before buyers' agents start offering it.
- Check the audit trail. Visa lists visibility into authorization and completion as a pilot capability. Any agent payment flow you adopt should show which agent paid, on whose instruction, and against which invoice.
Where This Fits in the Agent Identity Stack
The companies call this the first live pilot of its kind, and it tests an idea the payment networks have been building for a year: that an agent should be identified once, cryptographically, and then recognised wherever it transacts. Visa launched TAP with Cloudflare in October 2025, added the Agentic Directory in June 2026, and has now shown both working in a live payment flow outside consumer retail.
The same thread runs through our coverage of the Know Your Agent framework and Visa's Agent Score and Agentic Directory announcements. What changes with the BILL and Billtrust pilot is that the identity layer has a working business case attached: removing a manual step that finance teams already resent, inside a process where the hard question of what to pay has been answered before the agent arrives.
