AI is becoming central to fraud detection and AML monitoring. For German banks, the real question isn’t whether AI can identify suspicious activity, it’s whether compliance teams can understand, challenge and evidence the decisions it makes.
Why the black-box problem matters
Part of why that question matters are how fraud actually behaves. Fraudsters rarely stick to the patterns a rule engine was built to catch. They adapt, spread activity across accounts and channels, and work deliberately to look legitimate. Rules-based systems still earn their place for known patterns (e.g., unusual amounts, rapid sequences, high-risk destinations), but they also have a real limit: too broad rules, they flood the queue with alerts, and too narrow, they miss the pattern entirely. This is where machine-learning models add something rules can’t. They compare a transaction against a customer’s own history, rather than a fixed threshold.
That said, a score of “high risk” on its own doesn’t tell an investigator much. They need to know which transaction characteristics drove it, how the activity differs from the customer’s usual pattern, whether a rule or a model made the call, and what to do next. Without that context, AI adds volume to the queue without adding clarity to the decision. In the end, genuine risk gets harder to spot in the noise, and only flagging transactions still don’t answer the question a regulator asks: Why did each one do?
German banks already working AI out
That’s the gap a few German banks and infrastructure providers are already working to close. They are building systems that show investigators the reasoning behind a score, not just the score itself.
Deutsche Bank has described its “Black Forest” model, which flags transactions that deviate from typical patterns such as amount, currency, destination country and transaction type, and then routes the anomaly to the anti-financial-crime team. The mode also learns from feedback to improve the quality of future classifications.
The Sparkassen ecosystem has taken a similar approach with KIWI, an AI-supported component developed by Finanz Informatik. The system evaluates suspicious transactions and shows investigators the specific indicators that made it look risky.
In both examples, the same underlying principle applies: AI speeds up expert judgment here; it doesn’t replace it. That balance matters more than it used to, as fraud tactics change faster than any team can retrain itself from scratch, and staying ahead of that means investigators need the reasoning behind a decision.
Explainability for Investigators, Compliance, and Regulators
That same extends well beyond investigators. Often framed as a regulatory box to check, explainability serves three distinct audiences.
1. Investigators
Investigators need plan-language reasoning they can act on at the moment. Something close to “this beneficiary was added 12 minutes ago, the amount is 8x the customer’s average, and the device hasn’t been seen on this account before“. This kind of explanation helps an investigator decide whether to close the alert, request additional information, contact the customer, or escalate the case.
2. Compliance
Compliance owners need to track how the model behaves over time. They monitor metrics such as alert volumes, false-positive rates, drift in behavior, or fraud patterns. This turns explainability into an ongoing control rather than a document created only for an audit.
3. Regulators and auditors
Regulators and auditors need vident that the model the bank uses was built for. That is becoming less optional by the month as Germany’s AI Market Surveillance Act took effect in July 2026. BaFin new oversight of how banks and insurers use AI. With it, a sharper expectation of traceability as banks need to show that the fraud score came from appropriate data and moved through a controlled auditable process.
How Vyntra supports explainability
Vyntra’s AML transaction monitoring platform is built around the same idea running through this whole piece: a score is only useful if the reasoning behind it is visible. Rules and AI models work side by side, and every decision comes with the context an investigator needs. Fraud, AML, and sanctions alerts land in that same shared context too, so nobody’s piecing together a customer’s activity by hand across four different systems.
Transform Finance Frankfurt - September 3, 2026
German fraud and compliance teams will be discussing explainability and the compliance challenges at Transform Finance Frankfurt this September.
