Graph analysis, which maps every party and payment in a dataset into a single connected picture, helping illuminate possible suspicious links. For example, a first-time payment that looks fine on its own can prove risky once you see that the sender is potentially linked to fraud flagged at another company. This is particularly useful with virtual cards, which are designed for single use.

Transformer models—the same technology behind large language models—help surface subtle relationships in a dataset that no rule was written to catch. Using attention mechanisms to focus on the most relevant signals, they can help detect nuanced patterns that point to potential fraud.

1. New types of data: AI can help fraud prevention systems use data that couldn’t previously be analyzed at scale. For example, language models can interpret non-standard, free-text records linked to ACH payments—pulling more information out of each transaction. 

2. Large-scale data processing: AI and machine learning systems can analyze billions of data points, surfacing subtle patterns that signal fraud. “With more traditional approaches, the more features you add to the analysis, the more complex and difficult it can become,” Smallwood said.

3. Targeted detection: Traditional fraud detection models require heavily structured data. Newer machine learning models need less structure to identify important signals. “The fraud we see is truly a needle in a haystack, if you look at how many transactions we process. Newer models’ ability to focus helps with that problem,” Smallwood said. 

Stick to your processes: “People have a tendency to think they couldn’t get scammed. It’s this bias we all have, that we believe in our own faculties and abilities,” Smallwood said. Following strict standard operating procedures—for example, verifying any change in payment instructions using a phone number from a reliable system of record—can help your team interrupt that bias and spot red flags.  

Strengthen validation practices: The rise of affordable, convincing deepfakes makes it even more important to authenticate requests for payments, credentials or other sensitive information. “We’re seeing more emphasis on validation of knowledge,” Smallwood said. “If someone who sounds like your CFO asks you to make an urgent payment, do you have pieces of information you can use to validate their identity?”    

Communicate with your banker: The more you tell your bank about your company’s payments, the more precise its fraud detection can be. Wires and ACH payments have built-in mechanisms for sharing transaction information that can help your bank spot potential fraud or avoid false positives. And if your company expects unusual payment activity—for example, a spike in volume or transactions in a country where you don’t typically do business—tell your banker in advance. “It takes strong collaboration,” Kalamba said. “The more accurate data we have, the better we’re going to be—and the better the client experience will be.”