Fraud Learns Fast. Banks Must Think Even Faster: How to Protect Transactions, Customers, and Reputation

Authorized Push Payment (APP) fraud—where the customer is manipulated into sending money to a fraudster—is no longer an exception. It’s become a daily reality. What used to be sufficient for prevention—rigid rules and manual review—simply doesn’t cut it anymore.

says Jiří Kaplický, Director, Meonzi, a Trask company

Milliseconds That Cost Millions: APP Fraud Is Redefining the Balance Between Risk and Trust

The fraud economy runs on speed. In instant payments, funds disappear within seconds—chargebacks or card blocks are useless if the customer has “confirmed” the transaction. Impersonation, social engineering, and mule accounts shorten the time between manipulation and transfer. A delayed response means not just direct financial loss, but a loss of trust—something far harder to restore than any accounting entry.

While SEPA instant payments open new attack vectors, most fraud still occurs in domestic payment environments. Regulation is intensifying the pressure. PSD3/PSR sharpen liability for APP fraud, expand mandatory controls (e.g., IBAN/name check), strengthen authentication, and standardize APIs. The practical impact? Less room for excuses, more room for lawsuits and penalties—unless the bank can prove it did everything possible.

Loss doesn’t only occur at the moment of fraud. It multiplies over time as customers lose trust—and if the bank responds insensitively or imposes overly strict measures, friction in digital channels increases. That directly affects customer relationships and business outcomes.

adds Jiří Kaplický

AI as a Defense Tool, Not a Magic Button

Organized fraudsters use cutting-edge technology—AI is already part of their toolkit. Banks must respond with equal sophistication, but with systemic discipline. It’s not just about fraud detection—it’s about full-spectrum risk management in real time, from geolocation and behavioral analysis to UX-driven decisioning.

Effective defense doesn’t come from deploying a single system—it comes from connecting technologies, processes, and people, including the customer. Rule-based systems react to known patterns. But when attackers change the playbook, rules fail. AI models, on the other hand, detect anomalies, relationships, and behavioral sequences that neither humans nor rules can catch in time. They decide instantly—before the money moves—and adapt with every new interaction.

What AI Actually Does in Fraud Prevention

• Behavioral profiling and sequential models
  Instead of “is amount > X?”, you track dynamics: typing speed, device changes, unusual navigation paths, step order. Sequential models (e.g., transformer architectures for time series) understand context—what’s normal for this customer, here and now.
  ‍
• Graph learning and network detection‍
  Fraud is rarely a solo act. AI builds entity graphs (accounts, devices, IPs, cards, recipients) and identifies suspicious clusters: mule accounts, recycled devices, payment “washers.” Graph models reveal silent risks without traditional triggers.

• Streaming feature store
  Decision quality depends on what signals the model has at time T. A streaming feature store aggregates signals across channels (web, mobile, call center), from external sources (blacklists, sanctions) and internal ones (history, limits, device fingerprint). That’s the difference between 500 ms of “guessing” and 20 ms of “knowing enough.”

• Real-time action orchestration
  An AI score isn’t the goal—it’s a trigger. High risk → immediate block or step-up (biometrics, call-back), medium risk → delay and verification, low risk → frictionless flow. Real-time decisions must be tightly integrated with UX, so protection doesn’t punish legitimate users.

Real-world example: Geolocation in a mobile app can be a critical signal—but only if the technology is properly integrated.

Not all fraud happens inside the system

Some fraud activities unfold partially—or even entirely—outside the information systems of both the bank and the customer. That’s why effective protection requires more than just technical integration. Banks need access to behavioral signals, such as digital channel patterns or call center interactions, where deceived customers often turn as a last line of defense.

AI is a tool, not a solution. True protection only emerges when AI is connected to data, processes, decisioning, and customer experience—in real time, across the entire organization.

“It’s not about how much time you have. What matters is the quality of data available at the moment of decision,” says Anton Vaskovskyi, Vice President, Services Business Development. Czech Republic and Slovakia, Mastercard.

How to Protect Customers, Reputation, and Your Own “Money Mules”

AI-driven detection helps fulfill three key objectives simultaneously—aligned with growing regulatory responsibility, which now applies not only to outgoing but also incoming payments and anti-money laundering. PSD3 will require banks to compensate losses if they fail to demonstrate adequate measures. Protection thus applies not only to victims but also to those unknowingly used as tools of fraud—so-called money mules.

Transaction security
Real-time intervention stops the attack before funds leave the account. If refunds occur, they’re exceptional—based on proven failure, not systemic policy.

Customer experience
The goal isn’t blanket tightening, but micro-targeted intervention where real risk exists. AI helps minimize false positives and blunt rejections that harm NPS.

Regulatory certainty
Demonstrating adequate measures: IBAN/name check, strong authentication, consistent consent management, auditable decisioning. AI contributes to explainability—making decision logic accessible for compliance and dispute resolution.

Fraud Is Evolving. Defense Must Be Fast, Adaptive, and Resilient  

Latency budget: within tens of milliseconds for scoring
To truly protect instant payments, decisioning must happen in milliseconds—including loading relevant features. Anything that takes “seconds” is no longer prevention—it’s incident response.

Human-in-the-loop decisioning
AI assesses fraud probability, humans review ambiguous cases. The system learns from these decisions—and next time, it handles similar situations without operator input. Queues are prioritized by risk, not order.

Monitoring model drift and performance
Fraudsters change tactics. Track metrics like TPR, precision, recall, FPR by segment, latency, and feature drift. When behavior shifts (e.g., seasonal waves or new attack types), models must be continuously validated—without downtime or performance loss.

Data security and cyber resilience
AI needs data—and that data must be trustworthy. The cybersecurity layer protects signal integrity (e.g., against device spoofing or session hijacking), secures the feature store, and defends against data poisoning.

Even the best AI model can’t make good decisions without clean data. It’s like having a top pilot but broken instruments—you won’t fly safely.

emphasizes Jiří Kaplický

What to Watch For: Blind Spots in the Fight Against Online Fraud

Online fraud is growing—and it’s no longer just about cards. Attackers target instant payments, digital identity, and gaps in business processes. Even well-designed AI can fail if it’s not connected to UX, data, or operational reality.

False positives can damage business—especially if there’s no sensitive step-up strategy. Instead of hard rejection, dynamic verification (e.g., mobile token or biometrics) is more effective—it respects context and minimizes friction. Fraud management methods must be connected not only to technology, but also to people and customer experience.

A great model doesn’t guarantee a working solution. Without streaming features, fast APIs, and clear action scenarios, results won’t show up in practice. AI must be part of orchestration—not an isolated calculation. And explainability must be built in from the start—not added as an afterthought during complaints.

Finally: data without cyber hygiene is a risk. If signal integrity is off—e.g., device fingerprint or session ID—AI learns noise instead of signal. The cybersecurity layer must protect not just the data itself, but also its context, so decisions remain reliable and resilient.

It’s not just about connecting data, AI, and decisions—it’s about having the right data to begin with. Without clean, timely, and context-rich inputs—including behavioral and off-system signals—even the most advanced models will struggle. Education campaigns that help customers avoid manipulation are just as critical as technical safeguards.

It’s Not About the Model. It’s About the Results: How to Know AI Is Actually Working

Success isn’t measured by lab “accuracy.” Truly effective AI has visible impact: it reduces net transaction losses, limits false positives, speeds up decisioning, shortens dispute resolution time through explainability—and improves customer experience, measured for example by post-intervention NPS. If the numbers don’t move in these areas, the solution isn’t finished. It’s just installed.

AI is only successful in combating fraud when it works on the basis of sufficient high-quality data and continuously evaluates changes in the behavior of both clients and fraudsters.

Concludes Jiří Kaplický

One-Sentence Summary for the Board

Fighting fraud isn’t a race against time—it’s a matter of systemic design: the winner is the one who connects data, AI, and decisioning into a controlled ecosystem that operates in milliseconds and earns trust from both customers and regulators.