In 2026, the long-standing chore of a person logging into a supplier portal to finish an already-approved invoice is beginning to give way to software agents. Visa and Mastercard, which have spent decades governing how money moves, are now building the means to recognize and authorize those agents. The question they are answering is less about moving payments than about who is permitted to act on a company's behalf, and that question will shape how much of corporate spending is handed to machines.
Card Networks Use B2B Payments to Write the Rules for Agentic Commerce
The networks are trying to make authorization their indispensable contribution.
Mark
Why does it matter that Clink now supports both Mastercard and Visa?
Mimi
It shows developers don't want to be locked to one network, while each network wants its credentials built into every agent experience. Clink sits right in that tension.
Mark
And the agent isn't deciding what to pay?
Mimi
Right. The business approves the invoice first. The agent only executes, which keeps existing approval processes intact.
Luke
That's the framing, but it's analysis, not a finished market. The piece says networks are 'moving to' build this infrastructure. Nothing here shows who actually wins.
Mark
What about the adoption numbers?
Mimi
Nearly 7% of U.S. enterprise CFOs had agentic AI in live finance workflows, and 5% more were piloting.
Luke
Worth noting that figure is from September 2025, from one PYMNTS report. It's a snapshot, not today's number.
Mark
Where do banks fit?
Mimi
They supply credit, accounts and liquidity, which could become inputs when agents choose payment methods, if bank products are embedded in the workflow.
Luke
'Could' is doing the work there. The piece doesn't say that has happened.
Mark
And the biggest open question?
Mimi
Liability. An authenticated agent can still pay wrongly, and the source names that question without answering who bears it.
Luke
Agreed. Who is responsible when an automated payment goes wrong remains unresolved.
Il Polso
- Two card giants are racing to certify AI agents before they touch high-value supplier payments.
- Clink's addition of Mastercard Agent Pay puts it on both major networks, widening the contest for machine-initiated invoices.
- Agents are being confined to executing decisions businesses already made, a deliberate limit meant to calm CFOs and auditors.
- Open questions about fraud, liability and agent authority remain unresolved, and the outcome is still forming.
In 2026, the long-standing chore of a person logging into a supplier portal to finish an already-approved invoice is beginning to give way to software agents. Visa and Mastercard, which have spent decades governing how money moves, are now building the means to recognize and authorize those agents. The question they are answering is less about moving payments than about who is permitted to act on a company's behalf, and that question will shape how much of corporate spending is handed to machines.
For many finance teams, a payment that has already been approved still requires a person to log into a supplier's portal, enter credentials and finish the job. Even sophisticated CFOs find it hard to automate authentication and reconciliation for sensitive, high-value payments. Agentic AI promises a way out, and the market is moving toward it fast.
In early November, payments technology company Clink added support for Mastercard Agent Pay, joining its existing Visa Intelligent Commerce integration. The move followed an October Visa initiative with BILL and Billtrust that let authenticated agents complete previously approved invoice payments through supplier portals. Both efforts target some of the most stubbornly manual corners of B2B payments.
Removing that final human step raises harder questions: who authenticates the agent, who sets its spending authority, and who answers when an automated payment goes wrong. Card networks are building the infrastructure to answer them, using the credentials, authentication systems and commercial rules they already run.
The agent is not deciding whether to pay. It carries out a decision the business has already made. Industry voices stress that authentication is foundational and that governed agency, not unconstrained AI, is the realistic path. Banks, meanwhile, may find a role as agents begin choosing among payment methods based on cost, credit and terms. Still, a properly authenticated agent can make a wrong payment, so the real test will be exception handling, recovery and disputes.
In 2026, a B2B invoice that has already been approved often still needs a person to finish the job. Someone logs into a supplier's portal, enters payment credentials and completes the transaction. Even two digitally sophisticated CFOs dealing with each other can find it hard to streamline and scale the authentication and reconciliation of high-value, sensitive payments. Agentic AI may offer a way out, and the payments and fintech market is moving toward it quickly.
On Monday, Nov. 9, payments technology company Clink added support for Mastercard Agent Pay. It already supported Visa Intelligent Commerce, so it now works with both networks. The step followed an October Visa initiative involving BILL and Billtrust, which allowed authenticated AI agents to complete previously approved invoice payments through supplier portals. Both efforts target some of the most stubbornly manual parts of B2B payments.
Removing that last human step raises harder questions: who authenticates the agent, who defines its spending authority, and who is responsible when an automated payment goes wrong. Card networks such as Visa and Mastercard are moving to build the infrastructure that answers them. For decades they have competed on acceptance, transaction volume and the economics of moving money. Agentic commerce adds a new layer: establishing whether a machine has permission to initiate a payment at all.
Their advantage is that they already run the credentials, authentication systems and commercial rules that connect banks, buyers and suppliers. The agent, importantly, is not deciding whether an invoice should be paid. It is carrying out a decision the business has already made. Chris Trainor, head of platform strategy, innovation and AI at Paymentus, told PYMNTS that authentication remains foundational, and that the mistake is expecting the AI model to become the system of record or a transaction engine. In his words, the future is governed agency rather than unconstrained AI. Automating execution lets companies keep their existing approval processes, whereas delegating the decisions themselves would raise far thornier questions of corporate authority, fraud and liability.
Mastercard's Agent Pay, like Visa's offering, combines tokenized credentials, authentication and transaction controls for payments started by AI agents. Software developers want to build purchasing experiences without being tied to a single network, while the networks want their credentials and authorization tools embedded wherever those experiences run. Each participant can control a different piece: an enterprise software provider may decide when an invoice is ready for settlement, a bank may supply the funding, a network may authenticate the agent, and a supplier platform may decide whether to accept the payment. The networks are trying to make authorization their indispensable contribution.
Demand is real. A PYMNTS Intelligence report found in September 2025 that nearly 7% of U.S. enterprise CFOs had deployed agentic AI in live finance workflows, with another 5% running pilots.
Banks need not lose out. They supply the credit, operating accounts, liquidity services and relationships underneath corporate payments, and as agents begin choosing payment methods, those could become inputs, provided bank products and controls are built into automated workflows. Andrew Ng, head of payments and embedded finance at Tungsten Automation, told PYMNTS that payment execution alone is increasingly commoditized and that the lasting value lies in making the right payment decisions. An agent might eventually pick between a commercial card and an account-to-account transfer based on supplier acceptance costs, available credit, cash flow needs or payment terms.
Recognizing an agent solves only part of the problem, though. A properly authenticated agent can still make an incorrect payment. For CFOs, the case for agentic payments must therefore count exception handling, payment recovery, dispute resolution and the cost of maintaining controls across several software and financial providers. Whoever solves those problems may gain an edge that reaches well beyond processing transactions.
Citazioni salienti
The future is not unconstrained artificial intelligence. It is governed agency.— Chris Trainor, Paymentus
Payment execution alone is increasingly commoditized; the durable value is in making the right payment decisions.— Andrew Ng, Tungsten Automation (paraphrased)
Vuoi la storia completa?
Leggi l'originale su PYMNTS.com ↗