AI in payments in Germany

Stripe sorts AI in payments into the steps of a transaction: a personalized checkout, authentication with less friction, fraud detection across hundreds of signals, higher authorization rates through routing and retries, and dispute handling. In the survey of more than 2,000 business leaders that Stripe cites, 43 percent already used AI or machine learning in payments, and 32 percent planned to within two years.

Much of this is older than the current wave. Rich Turrin points out that machine learning has been used in fraud prevention since the early 1990s, and that three of the four main uses still rely on machine learning, not generative AI. In his view AI changes how people reach a payment, while the rails underneath stay the same.

The EU rules on strong customer authentication contain an exemption that only works with good risk models. Under Article 18 of Delegated Regulation (EU) 2018/389, a payment provider may skip the second factor for a remote payment if its real-time transaction risk analysis finds no warning sign, such as unusual spending, an unknown device, malware or an abnormal location, and if its fraud rate stays below a reference value. For remote card payments that value is 0.13 percent up to EUR 100, 0.06 percent up to EUR 250 and 0.01 percent up to EUR 500.

A provider whose fraud rate goes above the reference value for two quarters in a row loses the exemption. The quality of the fraud model thus decides how often a customer at the checkout is asked for a second factor. Strong customer authentication covers the rule itself, and AI fraud detection covers how the models work.

The joint EBA and ECB report counted EUR 4.2 billion of payment fraud in the EEA in 2024 and found that payments with strong authentication were generally less exposed to fraud. Instant transfers in euro now come with a name check before the money leaves, described on verification of payee.

The next change is a payment that an AI agent starts for a person, within a mandate the person gave in advance. It raises the question every payment rule answers first: who authorized this? Agentic payments covers the protocols proposed for it and the open questions of mandate and liability.

It scores each payment against the usual behavior of the customer, the device and the merchant, and blocks or challenges the ones that do not fit. Under the EU exemption for transaction risk analysis, a provider whose models keep fraud below the reference rates may also skip the second factor for low-risk payments.

Technology and payment firms have proposed protocols for it, among them Google's Agent Payments Protocol. The open points are proof of what the customer authorized and who pays when an agent buys the wrong thing, as agentic payments explains.