AI Agents Just Paid Their First Real Invoices: Inside Visa's First Live Trusted Agent Protocol Pilot
TL;DR
Visa, BILL and Billtrust have launched a live pilot processing real invoice payments through supplier portals using AI agents.
The model links an authenticated agent and tokenised payment credentials to buyer-approved instructions, while the supplier portal retains its acceptance controls.
B2B invoice payments are a compelling first use case: the buyer authorises the payment, the supplier wants faster cash flow, and both sides have an incentive to participate.
The pilot is a concrete deployment of the trust problem discussed in the five-protocol race we covered at Money20/20's opening, not evidence that the market has settled on one standard.
Payment providers should watch supplier acceptance policies, agent authentication, liability and invoice reconciliation as agentic commerce develops.
Visa, BILL and Billtrust have moved AI agent payments from concept to live B2B activity. Announced on 20 October 2026, their pilot lets BILL’s authorised Payment Agent initiate and complete approved buyer invoice payments through Billtrust supplier portals. It is processing real transactions with participating buyers and suppliers: a meaningful step beyond a protocol announcement or demonstration.
What is happening in Visa’s live AI agent payment pilot?
The pilot lets an authorised agent make an already-approved invoice payment without requiring it to log in as a human user. BILL’s Payment Agent is registered through Visa’s Trusted Agent Protocol and Agentic Directory. It uses tokenised payment credentials and follows controls aligned with the buyer’s approved payment instructions.
Billtrust uses the protocol to recognise and authenticate the agent, then permits it to access the supplier portal so BILL can complete the invoice payment on the customer’s behalf. Visa says the process addresses protections traditionally provided by CAPTCHA and bot-blocking tools, without requiring an additional agent login. Suppliers continue to receive payments through their portals under their own acceptance policies.
The control boundary matters: the agent executes a payment the buyer has approved; it is not given open-ended authority to decide what to pay. Visa’s announcement describes the pilot as part of its Intelligent Commerce portfolio, while BILL automates card payments for supplier invoices.
As Visa senior vice president and head of product, commercial and money movement solutions Mark Nelsen put it: “Businesses should not have to choose between automation and control when making payments.” He said the pilot is testing how identity, payment controls and transaction visibility can make agents work across platforms.
Why does this live deployment matter more than another protocol announcement?
It matters because B2B invoice payments are a practical, high-volume test of whether agent identity and payment controls can work across separate platforms. The distinction is between a protocol proposed on paper and a protocol being used in a live payment flow.
The 19 October discussion of know-your-agent (KYA) focused on an identity problem and a market with competing approaches but no established standard. This pilot is the first live deployment of one of those approaches in the scenario described: an agent authenticated across buyer and supplier systems to complete a real invoice payment.
Invoices are a useful proving ground because they already have business approval processes and a defined payment purpose. That gives agentic commerce a narrower, more governable job than a consumer agent with broad discretion to browse and spend. It does not settle every question about agent payments, but it puts the trust model into an operational setting.
How does authentication without a portal login work?
The flow combines agent identity, tokenised credentials and buyer-defined payment instructions, rather than treating an AI agent as a human user with a shared password.
Registered agent identity: BILL’s agent is identified through the Visa Trusted Agent Protocol and Agentic Directory. For payment providers, this creates a way to distinguish an authorised agent from unidentified automated traffic.
Tokenised payment credentials: The agent uses tokenised credentials rather than relying on a customer’s portal login. For issuers, acquirers and PayFacs, this puts credential controls and transaction permissions at the centre of the design.
Buyer-approved instructions: The agent is constrained by the buyer’s approved payment instructions. Providers will need to understand how those instructions are represented and enforced within their own payment and risk systems.
Supplier-side acceptance: Billtrust recognises the agent and permits access under its portal flow. Suppliers retain acceptance policies, while operators may need to define how agent-authenticated transactions fit within them.
The reported flow ties the action to an invoice payment. That could support clearer invoice-level reconciliation, although the pilot’s announced details do not establish specific reconciliation outcomes or performance gains. For POS companies, ISOs and payment providers, the wider lesson is that agent authentication needs to connect with existing authorisation, risk and reporting processes, not bypass them.
Why is B2B a more natural starting point than consumer banking?
B2B payments align the interests of the parties more clearly: a buyer directs its accounts-payable provider, a supplier wants to collect faster, and both sides can agree to a controlled payment process.
That contrasts with consumer banking, where financial institutions have reasons to restrict an outside agent’s authority over a customer’s money. In October, reporting described JPMorgan Chase, Citigroup, Bank of America and Charles Schwab as blocking or limiting third-party agents from moving customer funds. Approved connectors may provide read-only information, but that is different from permission to execute a payment.
In the invoice pilot, the buyer has authorised the payment and the supplier has chosen to accept payments through its portal. Faster cash flow is a practical incentive for suppliers to consider agent-authenticated payments. It is not a guarantee that every supplier will accept them; each portal still needs an appropriate policy and controls.
This split also reflects merchant priorities. PYMNTS Intelligence reports that close to nine in ten merchants with an owned site or app plan a hybrid approach, balancing direct customer traffic with agent-driven sales. The finding signals interest in agent-led commerce without necessarily handing over control of the customer relationship.
What should payment operators change or monitor?
Operators should start by deciding how their rules treat an authenticated agent, and who is accountable when its action results in a disputed payment.
For accounts-receivable teams and portal operators, acceptance policies may need to state whether registered agents are permitted, which credentials or protocols are recognised, and what controls apply. Fraud teams may need to shift from treating automation primarily as bot traffic towards verifying agent identity and authority.
Dispute and liability processes also need attention. Agent registration introduces another participant in the chain: the provider that establishes or vouches for the agent’s identity. Contracts and operating rules will need to clarify responsibilities if the agent is misidentified, exceeds its authority, or acts on incorrect instructions.
Finally, invoice-linked payments could make it easier to associate a transaction with the underlying obligation. That is a potential data benefit for reconciliation, not a proven result of this pilot. Payment providers should assess whether their data flows can preserve the relevant invoice and approval context from initiation through settlement and reporting.
How does Mastercard’s Agent Pay for Machines fit the picture?
Mastercard’s 20 October early-access launch addresses a different use case: autonomous agents buying digital services, rather than paying supplier invoices through a portal.
Its demonstration used the x402 open protocol and Mastercard’s Agent Pay for Machines Discovery Service. A marketing agency, Hype, gave an agent a task and a defined budget. The agent found providers, submitted a plan for approval, then covered nine payments across two merchants without separate accounts, API keys or invoices.
Together, the two moves point towards a similar design principle: a registered or identifiable agent, bounded payment authority and human-approved controls. But they remain separate stacks. Visa Trusted Agent Protocol and Mastercard Agent Pay are not one shared system, so fragmentation remains a strategic consideration for payment provider strategy.
What should businesses watch next?
The next test is whether the pilot expands beyond its current participants and whether other supplier portals adopt agent authentication. Businesses should also watch whether the model becomes generally available, how payment responsibilities are allocated, and whether consumer banks eventually support a comparable registered-agent approach for money movement.
For fintechs, issuers, PayFacs and payment companies, the immediate priority is not to assume that agents will be granted broad payment authority. It is to map how agent identity, customer permission, payment credentials and transaction records would fit into existing controls. The pilot makes that question operational.
Frequently asked questions about AI agent invoice payments
What is the Visa Trusted Agent Protocol pilot?
It is a live Visa, BILL and Billtrust pilot in which BILL’s authorised Payment Agent uses Visa’s Trusted Agent Protocol and Agentic Directory to access supplier portals and complete buyer-approved invoice payments. The pilot is processing real transactions with participating buyers and Billtrust suppliers.
Does the AI agent decide which invoices to pay?
No. In the announced flow, the agent uses tokenised payment credentials and follows controls aligned with the buyer’s approved payment instructions. It executes an approved payment rather than making an independent decision to pay.
Do AI agents need a separate supplier-portal login?
The pilot is designed to authenticate the authorised agent through the protocol without an additional agent login. Billtrust uses the protocol to recognise and permit access to the agent.
Why are invoice payments a useful agentic commerce use case?
They involve a defined business obligation, buyer-approved instructions and a supplier that may benefit from faster cash flow. Those conditions make invoice payments a practical setting for testing automation while maintaining control.
Are Visa and Mastercard using the same agent payment system?
No. Visa’s Trusted Agent Protocol pilot and Mastercard’s Agent Pay for Machines are distinct initiatives. Both emphasise controlled agent payments, but they remain separate approaches.
Key takeaway: the trust model is now being tested in live B2B payments.
Visa, BILL and Billtrust have demonstrated a live path for an authorised AI agent to complete approved invoice payments across buyer and supplier platforms. The significance is not that agentic commerce is solved; it is that identity, payment controls and supplier acceptance are being tested against real transactions. For payment providers, the task now is to prepare for a world where automated payments need to be both verifiable and bounded.