TL;DR
- Agentic payments are payments an AI agent initiates and completes on a person's or company's behalf, inside a mandate that sets whom it can pay, how much and under what conditions.
- Most of the work so far sits with card networks, protocols and merchants, and very little of it says what the bank holding the account has to do.
- McKinsey expects payments value to move to a control layer of identity, consent and policy guardrails, with as much as $75 billion in legacy fee revenue at risk.
- Consumer agent checkout has stalled, so the first real volume is likely to come in B2B payments, which puts commercial banks in the path first.
- A mandate is a banking product: who may delegate, the limits, the KYC behind it and the handling of exceptions and disputes belong in the bank's own channels and banker workflows.
- Banks can build that control layer above the core they already run, so clients and relationship managers have one place to set, see and stop what agents do.
On October 20, 2026, Visa, BILL and Billtrust launched what they called the first live pilot of trusted AI agents paying supplier invoices end to end. A BILL Payment Agent, registered through Visa's Trusted Agent Protocol, uses tokenized credentials and controls tied to the buyer's approved payment instruction, and Billtrust's portal recognizes the agent without a CAPTCHA or a separate login. It solves a real B2B problem: supplier portals that needed a person to retype what should have been an automated step.
Read the list of what the pilot enables, though, and one party is missing. There's an authenticated agent, a tokenized credential, payment controls, a payment path tied to the invoice and visibility into completion. There's no bank. The account that funds the payment, the relationship manager who knows the client and the team that will field the dispute if the invoice was wrong don't appear in the flow.
We think agentic payments need a control layer on the bank's side. It decides who may hand payment authority to an agent, within what limits and under which KYC, and what happens when something goes wrong. Networks and fintechs are building the rails and the protocols, and they're doing it well. If banks leave the mandate to them too, they keep the ledger and lose the relationship.
What agentic payments are and how they work
Agentic payments are transactions initiated, authorized and executed by AI agents without human approval at each step. The agent decides when, how much and to whom, from real-time context. ACI Worldwide describes the agent acting on a mandate that specifies approved merchants, spending limits and other conditions. Tokenization lets the user hand it a credential scoped to that mandate instead of the real card details.
The plumbing is arriving quickly. Protocols such as x402 and the Machine Payments Protocol let an agent pay for an API call in the same request, with no pre-created account. Circle launched managed payments for agent traffic because sellers still need to receive stablecoins, reconcile millions of tiny payments and pay the proceeds out to a bank account.
The value in payments is moving to the control layer
McKinsey's 2026 Global Payments Report sizes payments at about $2.6 trillion in revenue, roughly 41 percent of all banking revenue. As software embeds transaction flows, McKinsey says, value moves from the card tap and the checkout screen to the control layer underneath: machine identity credentials, consent frameworks, dynamic routing and policy guardrails. It puts as much as $75 billion in legacy fee revenue at risk from agentic commerce, against a productivity opportunity of about $110 billion from agentic AI.
Consent frameworks and policy guardrails are what a mandate is made of. Whoever holds them holds the client's trust in the payment. Corporate clients are already shopping around: 83% of midsize and large organizations are engaging or planning to engage directly with fintech vendors for cash management and payment services, according to Datos Insights.
Picture what that looks like with one more step automated. A client's treasurer delegates supplier payments to an agent set up in a fintech's dashboard. The operating account stays at the bank, the salary run and the credit line stay at the bank, and the bank learns about the agent when a payment nobody can explain lands on a relationship manager's desk.
Consumer checkout stalled, which is why business payments come first
The fair objection is that agentic payments are mostly hype, and Forrester's take on B2C agentic payments supports it. US consumer adoption of OpenAI's Instant Checkout stayed low from debut to discontinuation, and OpenAI shut it down in March 2026. Most merchants with ChatGPT apps send shoppers back to their own sites to pay.
Forrester also expects agent-to-agent payments to take off in B2B first, where payments repeat and agents already manage complex upstream workflows. Our data also support this. That's accounts payable, supplier invoices and treasury, the territory of the BILL pilot. Consumer hesitation buys retail banks some time. It buys commercial banks very little.
A mandate is a banking product, not a token setting
Crossmint, which builds agent wallets, lets developers set spending limits, whitelist merchants and require human approval above certain thresholds. From the perspective of a commercial banker, it's a familiar list. It's entitlements, payment limits and dual approval, the controls corporate clients already agree with their bank, usually through a form, an RM and a back-office team.
When the agent sits outside the bank, those controls get rebuilt outside the bank too, in a wallet or a network token, disconnected from the client's entitlements, its KYC file and its credit lines. A bank-side mandate answers questions a token can't. Which signatory delegated authority? Does the delegation survive a change of directors? Does the agent's activity fit the client's risk profile? What happens to its limit when the RM extends a facility?
Today those answers are scattered. Sixty percent of frontline work lives in the whitespace between a bank's systems: the handoffs, exceptions and manual checks no system owns. KYC review alone costs many institutions up to $175 million a year. Add the agent as a new actor on that foundation and you add another seam.
Exceptions and disputes decide whether agentic payments work
A mandate that works on the happy path is easy. The test comes when the agent pays the wrong supplier, pays twice, or pays someone impersonating a supplier. "Faster Payments equals faster fraud. At the end of the day, bad guys are going to be going after whatever form of payment that we have," Peter Tapling recently told us on the Banking Reinvented podcast when we spoke to him about the speed-trust tradeoff in commercial banking. An agent that pays in seconds, around the clock, is a very attractive new target.
ACI says securing agentic payments depends on verifying each agent's identity, scoping its credentials to the mandate and keeping a tamper-resistant record of everything it was authorized to do. That record earns its keep when a client calls the RM and asks why the money left. If the evidence is split between a network, a wallet provider and a fintech, the banker rebuilds it by hand. Disputes already stall when the case is spread across systems, and fragmentation is already costing banks the fight against fraud. Agent-initiated payments make everything much harder.
What a bank-side control layer for agentic payments looks like
We built the AI-native Banking OS to sit above the systems a bank already runs, the core, the payment processors and the CRM, rather than replace them. It's how we work with 120+ banks, and it suits agentic payments because the mandate needs the bank's context far more than it needs a new ledger.
For a commercial bank, that layer puts four things in one place. The client sets up and reviews agent mandates in the same digital channel where it manages users, approvals and payment limits. The RM sees those mandates, the agent's activity and the KYC status in their workspace, next to the rest of the relationship. Payment initiation and KYC checks run as shared banking capabilities that every workflow and agent calls the same way, so a limit set once applies everywhere. And when something breaks, the exception reaches a person with the evidence already attached, through the same resolution loops that the solutions built on the Banking OS run for disputes and KYC updates.
Governance is what has to hold up in front of a regulator. AI can guide, summarize, classify and recommend, but something has to govern execution. That's where the Sentinel, the authority layer of the Banking OS, comes in. No action runs without a Decision Token recording the policy applied, the actor's identity, the model version, the outcome and the context. That's the tamper-resistant record ACI describes, and the bank holds it.
Start with one payment flow
Pick the commercial payment flow where a client is most likely to hand work to an agent first. Supplier invoices are the obvious candidate. Walk it from the bank's side: who at the client can delegate, where that's recorded, which limit applies and who can change it, and how many systems the RM opens to answer a dispute. The answers tell you whether your bank will hold the mandate or hear about it secondhand.
FAQ
What is an agentic payment platform?
An agentic payment platform gives AI agents the means to pay: credentials, wallets or virtual cards, and rules for what they can spend. Crossmint, for example, offers agent wallets with spending limits, merchant whitelists and human approval above set thresholds. For banks the question is wider, because the mandate, the client's entitlements, KYC and dispute handling also need a home inside the bank's own channels.
Can AI help me pay my bills?
Increasingly, yes, starting with businesses. In a live pilot from Visa, BILL and Billtrust, AI agents pay approved supplier invoices end to end with tokenized credentials and controls tied to the buyer's payment instructions. Consumer use is slower: Forrester found US adoption of agent checkout stayed low.
What is a fintech payment service?
A fintech payment service is a non-bank provider that moves or manages money, from processing and cash management to stablecoin settlement. Circle's CPN Managed Payments, for example, handles wallets, reconciliation and payouts to bank accounts for businesses that accept payments from agents. Many corporate clients use such services alongside their bank, and 83% of midsize and large organizations engage or plan to engage fintech vendors for cash management and payments.
What are the types of digital payments?
Common types include card payments, bank transfers, instant payments, digital wallets and stablecoin transfers. Agentic payments can run on any of these rails; the difference is that an AI agent initiates the payment within a mandate instead of a person pressing pay. Protocols such as x402 and the Machine Payments Protocol let agents pay over the web directly.
How should a bank control payments made by AI agents?
Treat the mandate as a banking product. Record who delegated authority, the limits and approved payees, and the KYC behind the client, keep a tamper-resistant record of every action, and route exceptions to a banker with the context attached. Building that layer above the existing core lets it draw on the client's entitlements and history without a core replacement.
