AI in payments refers to the use of artificial intelligence, mostly machine learning, by banks, card networks, payment processors and merchants to make decisions and automate work across the life of a payment. The most established use is scoring transactions for fraud in the fraction of a second before authorisation, but models now also choose how payments are routed, decide when to retry declined ones, monitor accounts for money laundering, sort disputes and answer customer questions.
Payments suit these techniques because they produce large volumes of labelled data: every transaction has an outcome, such as approved, declined, disputed or confirmed as fraud, that can be used to train and test a model. The field draws on artificial intelligence in general, with machine learning as its main toolset, and more recently on language models and software agents that can act for a customer. Its use is shaped by data protection law, supervisory expectations on model governance and, in the European Union, the AI Act.
Fraud detection and risk scoring
A machine learning fraud model gives each transaction a score based on features such as the amount, the merchant, the device, the location, the card's past behaviour and patterns seen across many merchants. A risk engine combines that score with fixed rules to approve the payment, decline it or ask for more checks. Compared with rule-only systems, models can pick up complex combinations of signals and adapt as fraud patterns shift. Issuers, acquirers and card networks each run their own models at different points in the same transaction.
In 3-D Secure, issuers use risk scores to let low-risk online payments through without a challenge, the basis of risk-based authentication. Network or graph analysis looks at links between accounts, cards, devices and addresses to find organised groups such as money mule rings, which single-transaction scoring tends to miss.
Authorisation, routing and retries
For merchants, the main commercial target is the authorization rate. Smart routing picks an acquirer or network for each payment based on predicted approval and cost. Retry engines predict whether and when a payment retry after a soft decline is likely to succeed, which matters most in subscription billing. Other models flag merchant-side rules that are causing false declines, so that genuine customers are not turned away. Issuers use similar techniques to decide approvals, spending limits and credit line changes.
Compliance and financial crime
Transaction monitoring for anti-money laundering has traditionally relied on threshold rules that produce large numbers of false alerts. Machine learning is used to rank alerts by risk, detect unusual behaviour that fixed rules would not describe, and reduce false positives. Language tools help with fuzzy name matching in sanctions screening and with drafting case notes and reports. Supervisors generally accept such tools when a firm can explain how they work, document their testing and show that they do not miss risks a simpler system would catch.
Disputes, operations and customer service
Within dispute management, models classify chargebacks by likely cause, gather evidence and estimate whether a dispute is worth contesting. In back offices, matching algorithms reconcile payments against bank statements and invoices, and document extraction reads remittance details from emails and attachments. Chat assistants built on large language models answer routine questions about payment status, card blocks and refunds, passing complex or sensitive cases to staff.
Agentic commerce
Agentic commerce describes AI agents that search for, choose and pay for goods on a person's behalf. Card networks and payment firms have begun publishing frameworks in which an agent receives its own payment token with spending limits, and the customer's instructions are recorded, so that issuers can tell an authorised agent from a fraudster. Open questions include who is liable when an agent buys the wrong item, how consent is proved and how existing dispute rules should apply.
Risks, governance and regulation
Models can be hard to explain, can drift as customer behaviour changes and can repeat biases in their training data, for example by declining more payments from certain regions or customer groups. Banks apply model risk management, covering independent validation, monitoring and documentation; in the United States, supervisory guidance known as SR 11-7 set expectations from 2011 until April 2026, when revised interagency guidance, SR 26-2, replaced it. Under the GDPR, individuals have rights in relation to decisions based solely on automated processing that significantly affect them.
The EU AI Act, Regulation (EU) 2024/1689, treats AI systems used to assess the creditworthiness of individuals or set their credit scores as high-risk, while expressly excluding systems used to detect financial fraud from that category. Its obligations apply in stages. Criminals use the same technology, producing deepfake voices, faces and documents and automating attacks such as card testing, so defensive models have to be retrained continually.