On October 20, 2026, Visa announced a live pilot (opens in a new tab) in which AI agents pay invoices that have already been approved, doing it through supplier payment portals. For finance teams, the main point is simple: the agent doesn't decide what to pay. It carries out a payment that was already approved. The pilot depends on three building blocks: a way for a supplier portal to confirm the agent is trusted, a registry where agents are recorded, and tokenized card credentials.
Key takeaways
- Visa describes this as the first live pilot of AI agents completing approved invoice payments through supplier portals.
- Agents prove who they are to supplier portals using Visa's Trusted Agent Protocol, which can take the place of protections that CAPTCHA and other bot-blocking tools used to provide.
- The agent acting for the buyer is registered through Visa's Agentic Directory and pays with tokenized card credentials, not a raw card number.
- For accounts payable, the work moves from typing in payment details to setting the approvals, limits and records that tell an agent what it may pay.
- This is a pilot. Treat it as a sign of where B2B payments are heading, not as a finished product with settled rules.
What exactly did Visa announce?
Visa's newsroom lists the announcement among its October 2026 releases (opens in a new tab), dated October 20. According to the press release (opens in a new tab), the pilot connects platforms on the buyer side and the supplier side. An AI agent working for the buyer goes to a supplier's portal and completes an invoice payment that has already been approved.
Many suppliers already ask business customers to pay through a web portal. Someone in accounts payable logs in, finds the invoice, enters card or bank details and submits the payment. The pilot hands that last step to software. A person still approves the payment, and the agent completes it.
Why do supplier portals need to recognize a trusted agent?
Most websites that accept payments are built to stop automated traffic. CAPTCHA puzzles, bot detection and similar tools assume that software trying to log in and submit a payment is probably up to no good. For a long time, that assumption has helped block credential-stuffing attacks and card-testing fraud.
AI agents complicate this. An agent paying an approved invoice for a real business is automated traffic, but it isn't malicious. If the portal blocks it, the payment fails. If the portal lowers its defenses for all bots, fraudsters benefit.
Visa's answer in the pilot is the Trusted Agent Protocol. The press release (opens in a new tab) says the protocol authenticates agents to supplier portals and can satisfy protections historically provided by CAPTCHA and other bot-blocking mechanisms. Put simply, the portal doesn't ask "is this a human?" It asks "is this an agent we can verify?"
What are the building blocks, in plain terms?
Visa names three components. The table gives a short summary, and each one is explained below.
| Building block | What it does | Everyday comparison |
|---|---|---|
| Trusted Agent Protocol | Lets a supplier portal confirm an agent is legitimate | Checking an employee badge at the door |
| Agentic Directory | Registers the buyer's agent | A list of authorized badge holders |
| Tokenized card credentials | Pays without exposing the actual card number | A single-purpose key, not the master key |
Agent authentication
Authentication answers the question "who is this?" With human users, that means passwords, one-time codes or biometrics. With an agent, the portal needs a dependable way to tell a known, accountable agent apart from an anonymous script. The Trusted Agent Protocol plays that role in the pilot.
An agent registry
Before an agent can be checked, something has to vouch for it. In the pilot, the buyer's agent is registered through Visa's Agentic Directory (opens in a new tab). The idea of a registry is that agents aren't anonymous. They are recorded, so the parties in a transaction have something to check against.
Tokenized card credentials
Tokenization replaces a card number with a substitute value, called a token, that can be used for payments. Tokens are generally limited in where and how they can be used, so a stolen token is worth much less than a stolen card number. In the pilot, the buyer's agent uses tokenized card credentials (opens in a new tab) to pay supplier invoices. That matters because the agent never needs to hold or type in the company's actual card number.
What could change for accounts payable teams?
Today, paying through a portal is often manual even when everything before it is automated. An invoice might be captured, coded and approved in an AP or ERP system, and then someone still has to log in to the supplier's site to pay it. Agents are aimed at that gap.
If this approach catches on, a few things could shift:
- Less time spent keying in payments. Staff who now log in to dozens of supplier portals could spend more of that time on exceptions and supplier questions.
- Approval becomes the control point. If an agent pays whatever is approved, the approval step carries more weight. Weak approval rules would be automated along with everything else.
- Different records to review. Payment records would show an agent completing the transaction. Finance teams will want to be able to trace each payment back to the approval and the person who gave it.
- Supplier-side changes. Suppliers would need portals that can recognize and accept trusted agents, not just block every automated visitor.
For many businesses, the payment method a supplier accepts will still depend on the relationship. Our guide on ACH vs. card payments for B2B invoices covers how those choices are usually made.
What should finance teams think about before letting an agent pay?
Nothing in the announcement suggests businesses need to act right away. Still, it's worth thinking through the questions an agent-based process would raise. What follows is general information, not legal, tax or financial advice. Your own controls should reflect your auditors' and advisors' guidance.
- Who approves, and how? Spell out which invoices need which approvers before any payment can go out, whether a person or an agent sends it.
- What is the agent allowed to do? Think about limits by supplier, amount and payment method, and about what should happen when an invoice doesn't match a purchase order.
- Which credentials does it use? Tokenized credentials kept separate from the cards employees use make it easier to see and control agent activity.
- Can you reconcile it? Agent payments still need to match invoices and post correctly in your books. Our article on payment and ERP integration explains what reconciliation involves and which steps integrations automate.
- How will you know if something goes wrong? Plan for alerts, review of unusual payments and a clear way to stop an agent quickly.
Frequently asked questions
Does the AI agent decide which invoices to pay?
No. In the pilot, agents complete approved invoice payments (opens in a new tab). Deciding what to pay stays with the buyer's approval process.
Why not just let bots through supplier portals?
Bot-blocking tools exist to stop fraud. Removing them would expose portals to attack. Authenticating trusted agents lets a portal accept the agents it can verify and keep blocking the rest.
Is this available to every business now?
The announcement describes a live pilot, not a general launch. Details, availability and participation requirements may change as the pilot goes on.
The bottom line
Visa's pilot is an early, concrete example of AI agents completing real B2B payments. The design rests on familiar ideas, verified identity, a registry of who is allowed to act and tokenized credentials, applied to software rather than people. For finance and AP teams, the useful step now is to make sure approvals, limits and reconciliation are solid enough that you would trust software to carry them out.
