Citi deploys AI to detect real-time fraud, errors in payments

Register now

In the face of well-publicized cyberattacks and the threat of fraud, Citigroup is deploying artificial intelligence in an attempt to better monitor corporate customers' payments.

Citi has partnered with Feedzai, a fintech startup in which Citi Ventures began investing in 2016. The bank expects to take the new system live early in 2019.

“The entire payment industry is worried about payment security, banks are worried about it and corporate customers in general are worried about it,” said Manish Kohli, global head of payments and receivables, Citi’s Treasury and Trade Solutions. “As payments move faster, it’s easy for errors and fraud to happen. And the more our clients rely on technology, automation and straight-through processing on their side, the higher the risk is of a cyber event or a machine on their side not following instructions and sending an incorrect payment.”

AI software can instantly analyze a current transaction against all historical ones a client has conducted and recognize anomalies. It can learn the customer’s behavior over time and take into account how it changes. More traditional rules-based fraud detection systems typically compare a payment against a checklist of preset rules.

Payment fraud and error is just the first use case Citi has in mind for the Feedzai technology. It plans to apply it to anti-money-laundering programs and other uses.

Under the current plan, if a customer sends a payment to Citi that seems out of character compared to historical activity — perhaps it’s being sent at an unusual hour, in an unprecedented amount, or to a country the company has never sent a payment to before — Citi may send the payment request back to the client for review.

“This can be used to reduce the risk of cyberhacks or this could be used to prevent even general errors if a customer sends a payment that appears to be unintended,” Kohli said.

Some clients will start off using this in offline mode, where payments are observed and flagged in real time, but they aren’t stopped. Once the customer is comfortable with the false- positive rates, typically clients prefer the false positives to be below 20 basis points, they will move to online mode, in which suspicious transactions are halted and returned to the client for review.

For reprint and licensing requests for this article, click here.