How this regulator is using AI to probe financial fraud

Register now

When governments settle major fraud cases against banks, the punishment often seems to come long after the initial wrongdoing occurred. Case in point: the Justice Department’s $4.9 billion settlement with Royal Bank of Scotland in August over the way the bank misled investors in residential mortgage-backed securities between 2005 and 2008.

But some regulators believe that artificial intelligence can help speed such cases up.

The U.K.’s Serious Fraud Office, which holds bankers and others accountable for large-scale misdeeds that range from the rigging of the London interbank offered rate to bribing government officials and beyond, has been using two forms of AI to comb through the millions of documents, emails and text messages that are evidence in such cases to determine charges more quickly.

“We’re dealing with some of the most document-heavy, complex criminal investigations anywhere,” said Ben Denison, chief technology officer of the group. “It can be five years before you can reach a charging decision, because there’s so much material you have to review.”

But using AI can change that, he said, by looking for patterns in communications between people, including Word documents, emails, text messages and spreadsheets.

“Imagine you had a keyword you were looking for, it will show you on a particular day there were lots of communications with that word and then show you who the communications were between,” Denison said. “Instead of you having to know which person you’re looking for, it will help you to identify who might be of interest and then in turn, who they were speaking to or emailing.”

The Serious Fraud Office, or SFO, is not alone in turning to AI to catch fraud faster. The Securities and Exchange Commission uses AI to monitor market conduct by gathering and analyzing reported data and big data. Other agencies are also eyeing it if not employing it already.

Regulators “will almost certainly expand their use of these tools, and so will everyone else,” said Jo Ann Barefoot, CEO of Barefoot Innovation Group. “At the same time, policymakers will have to figure out rules of the road to be sure AI meets basic standards for accuracy and fairness.”

Fighting fraud is one of the most advanced use cases for AI in financial services, for both governments and banks around the world, she said.

"One reason is that there is high motivation to invest in it — saving money and catching crime," Barefoot said. "Also, fraud investigations generally occur after the crime has been discovered, which makes it relatively easy to train the AI to recognize which patterns are in fact fraudulent and which aren't.”

Catching fraud

In the financial services arena, the U.S. has tended to prosecute firms over executives. But the SFO generally goes the other way, seeking to jail individuals over penalizing corporations.

One of the challenges of working on big cases, however, is the sheer volume of material that’s produced.

“That’s a problem faced by the SEC, DOJ and others in the U.S.” Denison said.

For instance, 10.5 million documents were leaked in the Panama papers case. When the SFO was investigating Rolls Royce for bribing officials in Indonesia to buy jet engines, it had to go through 30 million documents. In two current cases, for which Denison couldn’t divulge the accused, one has 65 million documents, while the other has 100 million.

Artificial intelligence and machine learning are helping investigators get to the more relevant material at a faster rate.

One such technology is an AI-powered document review system, OpenText’s Axcelerate, which is employed by the SFO.

It tries to learn what it can from the information it processes and then identify similar material, in the same way Amazon will recommend an item based other people’s buying patterns. The algorithm gets better at finding relevant material as the case progresses.

Another technology, Ravn’s iManage, scans documents to find those that are covered by legal professional privilege and therefore cannot be used in an investigation. The fraud investigators sometimes refer to this as a robot or a robo-lawyer.

In the past, the SFO has hired independent lawyers to review material and figure out what might be privileged. But that can be a lengthy process.

The SFO ran a test in which it gave the same sample set of documents to the Ravn software and to a group of lawyers to see whether they came to the same conclusions.

“What we found was the robot results were more consistent than the lawyers,” Denison said. “If you have a team of five lawyers, they can make slightly different determinations, they’re people and have independent thought. Whereas the robot was applying the same rules every time. We were able to demonstrate to the defense team that this approach was as good as the way we were doing it in the past but it was much faster.

Because it’s in the interest of those we’re investigating to conclude the case earlier, this was something they were willing to sign up to.”

Denison makes the case that such technology isn’t depriving lawyers of jobs.

“There is an element of that, that it’s taking away work we were previously using lawyers to do,” he said. “But the other side of that is this isn’t work they particularly want to be doing. It’s quite low level, quite repetitive. We’re not going to stop employing lawyers but we’re making better use of their time.”


On the Rolls Royce case, the Serious Fraud Office was able to cut 80% of the cost of sorting out privileged material with the use of the software.

“At its peak the robot was reviewing 600,000 documents a day, as opposed to 300 documents a day a lawyer could do,” Denison said. “So the savings is substantial and had we used it at an earlier point in that case, we could have made even greater savings. That’s what we’re hoping to do in the future.”

Going forward, the Serious Fraud Office will expand its use of this technology and find technologies to help it extracting evidence from mobile phones, laptops and other devices that are equipped with ever-stronger encryption.

“We continue to invest in that area and we continue to look at the market in terms of analytics and AI software to see how that can benefit us,” Denison said. “Given the volume of materials is increasing and the technologies get better and more things get developed, it’s very important for us to stay up to speed with what’s coming out and be able to use it.”

More banks and regulators are adopting a similar mindset, because they have to.

“Our current investigative systems were originally designed on paper, when data was scarce and expensive,” Barefoot said. “Now it’s cheap and ubiquitous.”

Editor at Large Penny Crosman welcomes feedback at

For reprint and licensing requests for this article, click here.
Artificial intelligence Regtech Document management Machine learning RegTech Conference