Your board has read the Visa and Mastercard announcements and wants to know where agentic AI in payments fits your roadmap. Networks and banks run live pilots, transaction volume is still small, and few teams have put a payment agent into production.
The short version: agentic payments already run in back-office queues and in a few customer-facing products, such as Bakkt's Zaira. Card-network checkout is still a pilot.
TL;DR
Networks run pilots, and agents touch 3% of merchant transactions.
Back-office agents already show measured time savings.
Wrong agent transfers can land on the PSP, so checks run before execution.
Bakkt's Zaira moves money to 63+ countries, built in five months.
How is agentic AI used in payments processing?
Agentic AI runs in three places in payments today: agent-initiated purchases on card networks, conversational money movement, and back-office operations. Only the first is still in pilots; the other two run in production.
Area | What the agent does | Status in 2026 | Example |
|---|---|---|---|
Agent-initiated purchases | Finds, selects, and pays within user-set limits | Live pilots | Visa Intelligent Commerce, Mastercard Agent Pay |
Conversational money movement | Runs onboarding, KYC, funding, and transfers in chat | Production | Bakkt's Zaira, 63+ countries |
Back-office operations | Validates, repairs, and investigates payments | Production, human-supervised | BNY digital employees |
For agentic AI payments processing, the dividing line is who confirms the payment. Agent-assisted flows, with a human approval step or a deterministic check, are live, while payments the agent completes alone stay in controlled pilots. Meta's Muse, launched in the US in September 2026, is agent-assisted: users approve each purchase total in chat.
So AI agents in payment processing reach production first where something other than the model stands between its output and the money. That boundary is a design decision, and teams building fintech software for regulated clients make it before they choose a model.
How many agent payments run today?
Fewer than the headlines suggest. Visa reported "hundreds" of completed agent-initiated transactions in December 2025, its only published count. In July 2026, it opened live agent purchases to a first cohort of more than 30 European issuers, including Barclays, ING, and Revolut.
Mastercard has completed live agentic transactions with Santander in Europe, DBS and UOB in Singapore, and HSBC and DBS in Hong Kong. In the UAE, the first live Visa Intelligent Commerce payment in December 2025 ran on an Emirates NBD card. Santander ran its March 2026 payment in a controlled environment and stated it "does not constitute a commercial rollout at this stage." None of these releases gives a transaction count.
Merchants surveyed by Checkout.com in 2026 say only 3% of their transactions involve AI agents. On Visa's July 2026 earnings call, CEO Ryan McInerney said: "We're in the very early stages."
The forecasts are large and far apart. McKinsey (2025) projects agentic commerce could orchestrate $3 trillion to $5 trillion in global revenue by 2030. Juniper Research (2026) puts 2030 spend at $1.5 trillion.
What blocks adoption?
The technology is ready, but trust and regulation lag behind it. Competing standards split identity and checkout, among them Visa's Trusted Agent Protocol, Mastercard Agent Pay, OpenAI and Stripe's Agentic Commerce Protocol, and Google's AP2. The Bank of England's DLT Lab (2026) calls them not interoperable across providers.
Surveys from 2026 show four barriers:
Trust. 27% of consumers trust no organization to run a shopping agent, and 24% will never delegate purchases (Checkout.com).
Security. 50% of US bank executives name securing agentic technologies among top emerging threats (KPMG).
Liability. 93% of merchants say the agent provider should bear the loss from a wrong purchase (PYMNTS Intelligence).
Regulation. 34% of UK online retailers name the lack of regulatory guidance as a barrier (The Payments Association).
The law also explains the liability worry. Under PSD2, Directive (EU) 2015/2366, Article 73, the payer's PSP refunds unauthorized transactions. Taylor Wessing (2026) argues a purchase outside the user's instructions can count as one.
The compliance and security risks of agentic AI start there, and Weavr's answer is to trace every agent payment to an accountable company or person. Its CEO, Alex Mifsud, puts it bluntly: "An AI can't go to jail or sue anyone."
Which use case should you ship first?
Start in the back office, where errors are reversible and a person already reviews the output. BNY's numbers show why.
BNY runs nearly 140 "digital employees," each with a human supervisor who assigns tasks and reviews output. In payments, validations that took five or six minutes now clear in under 30 seconds. Open payment investigations dropped by nearly 80%, and client inquiries that once waited days now close the same day (Microsoft WorkLab, 2026). Microsoft published these figures as a customer story, so read them as a supplier's case study.
Visa processed 106 million disputes in 2025, 35% more than in 2019, which makes disputes the next queue to automate. In April 2026, it released AI tools that summarize merchant documents for issuers and autofill dispute responses for acquirers. Fraud review and KYC follow the same pattern, covered in this breakdown of AI agents in fraud detection and KYC.
Agentic payments use cases that move customer money call for harder controls. Mifsud's simplest version is a spending cap: at a 100 euro limit, a wrong agent payment costs at most 100 euros. At Sibos 2026, Manoj Bohra, chief data and AI officer at State Street, said irreversible actions need "heavy human-in-the-loop" (FinTech Futures).
Customer-facing agents can still ship: Bakkt's Zaira routes every transfer through a deterministic check outside the model.
Pick your first use case with three tests: the error is reversible, a person reviews the output, and the agent calls APIs that already work.
What does shipping one take?
A small team and an architecture that never lets the model move money unchecked. Zaira reached production in five months with four engineers. A transfer executes only when its identifiers match what the user confirmed, and that check runs at the system level.
The guardrails that stop AI payment agents from hallucinating transactions follow the same split. The model captures intent, while deterministic code validates and executes. Fee and limit answers come from verified data. Model Context Protocol (MCP) sits between the agent and the payment tools, so swapping a provider changes a tool definition.
The agent also depends on the payment systems underneath it. At xpate, 3M+ monthly transactions run across SEPA, SWIFT, FPS, FX, crypto, and card rails. An agent triggering a transfer there is only as reliable as the integration layer it calls.
Key insights
Read together, the data points to three conclusions for your first agent project.
Network data describes pilots and issuer cohorts. Until a network publishes volume, merchant surveys are the only estimate, and they show small adoption.
Liability decides the order of use cases. Reversible, supervised work goes first, and customer money waits until a deterministic guard checks every transfer.
The integration layer caps the agent. Check the payment APIs before the agent project starts, because no model fixes a failing integration.
The teams that ship agentic AI in payments keep the model away from execution and connect it to integrations that already work. Both hold for Bakkt's Zaira, in production after a five-month build.
Authors
Andrzej Wysoczański
Frontend developer with 10 years of experience. With The Software House for almost 7 years, going from a regular dev to the Head of Frontend. He loves keeping tabs on the latest frontend technologies, especially React-related. Regular of the Taby & Spacje podcast (tsh.io/taby-vs-spacje) for Polish speaking programmers.