DailyPay, an earned wage access provider that processes $30 billion a year, has to keep an eye out for money laundering and other forms of financial crime, just as banks need to do. To do this work efficiently, the company recently began using agentic AI — autonomous systems that act as agents, leveraging large language models to reason, plan and take multi-step actions.
"We use agentic AI in the sense of, let's get ourselves another agent that we can work with that filtrates all the prompts, but we don't give it decision-making power," Gregory Schipilliti Cid, New York-based DailyPay's director of international, told American Banker. Human experts make the actual decisions.
DailyPay is an example of an early adopter of agentic AI for anti-money-laundering work. In recent years, AML software has used machine learning to analyze transactions and put red flags on the ones that might be criminal behavior. In the current generation, vendors of AML compliance software, such as Quantexa, ComplyAdvantage, Thetaray, Nasdaq Verafin, SymphonyAI, Genpact and Unit21, have recently begun offering agentic AI, which can act autonomously to research those flagged transactions further and make a preliminary decision about which are likely criminal or benign.
Experts say agentic AI can help streamline AML work, provided it's fed high-quality data and used with appropriate supervision.
How it works at DailyPay
It's unlikely that anyone would launder money through an earned wage access program like DailyPay's. The company's users are typically hourly, modest earners, and DailyPay only makes about 40% of their pay available. The typical transaction is $100 or less, most commonly to buy gas or groceries or to fill prescriptions. Nevertheless, the company needs to comply with regulations in both the U.S. and Canada, as it's expanding into the Canadian market.
The agentic AI software DailyPay is using, from ComplyAdvantage, does work at a level equivalent to that of a recent college graduate, Cid said. It finds essential data points and hands them to a human decision-maker.
"You use agentic AI to provide a summary," he said. "You're going to have it be an additional analyst and a digital trust and safety consultant or an auditor."
It's helped reduce the number of false positives and the amount of research DailyPay's human analysts have to do, according to Cid. It's cut the workload of human analysts by 50%, he estimates, so the company's existing team can do more work without having to add anyone new.
"There is a deliberate cost savings that we get as an organization, because we can handle increasing volume," Cid said.
ComplyAdvantage's AML software helps identify anomalous transactions, such as someone who normally works 35 hours a week suddenly working 60 hours a week, or a person who previously used DailyPay twice a week now using it seven times a week. Often, this turns out to be a technical issue, such as a user's company switching from one HR management software service to another, rather than crime.
Like other AML compliance software programs, ComplyAdvantage's platform analyzes money laundering and financial crime risks during onboarding screening and across sanctions data, politically exposed persons data, negative press and such. It risk-rates customers and monitors their activity after they're onboarded.
The company's AI agent is built on machine learning models that analyze data provided by the client and by ComplyAdvantage, such as sanctions lists. It was trained by some of ComplyAdvantage's customers.
Alert fatigue
With agentic AI releases, AML vendors say they are addressing a common challenge for banks in their anti-money laundering and sanctions screening work: alert fatigue. AML software can generate thousands of red flags daily, and sorting through those messages to determine which are true indicators of criminal behavior is where a lot of the work lies.
"Think about someone who is reviewing hundreds of alerts a day, and 90% or 95% of the alerts they review are just low quality alerts — it's very clear they're false positives, and they have to go through the exact same motions every time to clear them," Jeff Fox, associate director, U.S. Advisory Services at Wolters Kluwer, told American Banker.
"That's when you run the risk of complacency and making mistakes, not documenting something appropriately, because it's always a false positive, and you're just going through the motions," Fox said. "If the tools can pull out and extract those high-value alerts, analysts will be much more engaged, and accuracy would then improve, because you're going to be reviewing stuff that is actually a potential risk to the organization, versus something that is very clearly a false positive."
Unlike other regulatory requirements, anti-money laundering work requires judgment, Fox pointed out.
"AML is very contingent on who you're dealing with, what products they're using, what is normal and appropriate behavior for them, and then taking what's expected and what you learn from account opening and your customer due diligence, and then comparing that to what their transaction activity actually is," he said. "So even with a human involved, it's very judgmental. So it's not black-and-white by any means. But having AI help with that data collection, data gathering and collation and summary, and then giving that to the human, is really worthwhile."
Data dependent
Agentic AI's effectiveness at catching money laundering depends on the quality of the data it can access, according to Fox.
"In an ideal world, the tool an organization is looking to use has solid, direct access to everything that it's going to need access to. That's where the really big opportunity is," he said. Where this is true, agentic AI can collate and consolidate information, cross-reference, double-check, and present an initial summary that a human can review for accuracy and make adjustments as needed, he said.
The challenge, in Fox's view, is that there are typically a myriad of systems, locations, drives and documents that investigators need to pull data from, and the data tends to be in many different formats.
"If you're missing one of those pieces, whether it's international wire transfers or account-opening documents. because the system doesn't have functionality to do that, that's where I think everything is going to completely break down," he said. "You might have 10 or 15 sources that you need to develop a staging area for, and make sure you have everything you need. Then I think these tools have the ability to be much more useful for analysts and to kind of streamline the investigations."
Banks that have high volumes of suspicious activity reports and Office of Foreign Assets Control sanctions screening work are the most likely to be interested in agentic AI for AML, he said.
And it's better suited for alert review and adjudication than for monitoring and reporting suspicious activity, which involves many data sources, he said.
"If you receive a sanctions alert or a negative media alert, those types of tasks are much easier to have AI help you manage, because you have finite discrete information for the sanctions party, and then you have that information in your core system for your own customer," Fox said.
