- Key insights: First National Bank of Omaha was a relatively early adopter of AI agents for financial crimes investigations.
- What’s going wrong: The bank says it has cut investigation time in half. Experts say there are risks for banks to be aware of.
- Future outlook: FNBO and others are considering employing AI agents for fraud detection efforts, among other use cases.
At First National Bank of Omaha, AI investigators have begun taking over some of the work of human financial crimes investigators, cutting the amount of time humans spend on each case by 50%, bank executives said. Agents also help human investigators consistently respond to fraud, money laundering, and sanctions violation alerts.
Processing details
The adoption of AI is part of a broader trend in which U.S. banks are seeking returns from investments in agent AI. At the same time, financial crime is also on the rise. Nasdaq Verafins
“The consistent challenge of financial crime means we are always looking for better ways to solve problems,” Nick Baxter, FNBO’s chief risk officer, told American Banker.
Chuck Subrt, director of fraud and AML countermeasures at Datos Insights, said he sees the use of agent AI in financial crime activity as a rapidly emerging trend, although banks are still in the early stages of adoption.
“Across the financial services industry, agent AI is gradually moving beyond pilots and into operational financial crime operations,” Subrt told American Banker. “Many of the foundational elements of an investigation, such as collecting transaction history, entity data, and contextual signals and distilling them into structured summaries, are among the first areas in which agencies apply them. Investigators have now completed many of the steps necessary to complete the large amounts of iterative evidence collection required to support their analysis and decisions, but the final decision-making remains with analysts and investigators.”
What the AI agent does
FNBO was already using Nasdaq Verafin’s anti-money laundering and fraud analysis software. The vendor showed the FNBO team two AI agents it launched a year ago to enhance due diligence and sanctions inspections, and the company was open to trying them out.
“From a Nasdaq perspective, we recognize that the industry needs to help address increasing regulatory complexity and illicit financial activity,” Rob Norris, senior vice president and head of product at Nasdaq Verafin, told American Banker. “The need to fight these bad guys is growing, and our customers across the board say they are looking for technology to help them do it.” The company’s AI agents are developed using AWS Bedrock, Amazon’s service for building generative AI applications, which allows them to be model agnostic, according to Nasdaq Verafin.
Enhanced Due Diligence AI agents perform some of the work required when a bank receives an AML, fraud, or sanctions violation alert, determining whether the alert truly indicates that a financial crime has occurred that warrants a suspicious activity report, or whether the alert is a false positive.
Another AI agent helps banks investigate sanctions review alerts. If a sanctions alert flags that a customer appears to be doing business with a person or company on the U.S. government’s Office of Foreign Assets Control list, this AI agent can examine the customer’s transaction history and pull up context that may prove or disprove the need to file a suspicious activity report.
“From the beginning, we’ve focused on what’s called Level 1 alert triage,” Norris said. “So you can think about all the different types of alerts that the system generates for anti-money laundering and anti-fraud purposes, and you can also think about sanctions screening and things like that. There’s usually a level one screening of what are good alerts and bad alerts, which alerts need to be investigated further, which alerts are explainable.”
More than 650 banks use these agents, according to Nasdaq Verafin. According to the company, Agentic Sanctions Analyst has reduced sanctions alert review workload by up to 90% for some banks, and Agentic Enhanced Due Diligence Analyst has reduced the time it takes to complete enhanced due diligence reviews by up to 50%.
With FNBO, agents quickly investigate and document why some alerts are false positives or indicate low risk, while others are serious and need to be escalated to a human for review. According to David Dawson, FNBO’s BSA/AML Officer, one tangible result is improved consistency.
“In many cases, existing processes are made to look like they are already perfect, free of defects and errors,” Dawson told American Banker. “You can do your best to use templates, but if you hire 100 investigators and analysts and ask each of them to do 100 enhanced due diligence reviews, you’re going to end up with variations in how information is documented.” With AI agents, basic investigations are automated in a consistent way.
“I think if there was a 10-step process, we used to start at step one,” Dawson said. “Today we begin Step 3, and it is exciting to progress incrementally through this type of innovation in existing embedded processes.”
The AI agent provides artifacts and summaries that human analysts can review, Dawson said, so they can free up their time to do more valuable analysis.
For example, sanctions alerts may flag transactions by customers whose names are similar to people on the OFAC list.
“But if you research, you may find that the jurisdictions, dates of birth and contact information are different,” Dawson said. “You can say this name sounds similar to this name, but they’re clearly different people, and that information is used by the agent to streamline and document why it’s interesting to the bank, or they can clear it.”
AI agents are helping banks examine financial crime investigations “at a much more granular level than manual processes,” Baxter said. “At the end of the day, the biggest benefit is the human resources that are freed up to work on higher-value work. So we’re more productive, we have better work products, we can monitor work more effectively, and we free up resources that can be used for higher-value work. This means we have a greater impact in taking criminals off the streets.”
Dawson said the new AI agents will not replace humans. “Financial crime is not going away,” he said. Agents help human investigators “put bad guys behind bars and protect their communities’ banks.”
To maintain security and control access to the system, AI agents are given the same access controls and privileges that human agents receive for the same tasks within the Nasdaq Verafin system.
“It allows us to preview how it will work before banks are granted access to take further action, so we can think about the various safety and governance measures we will put in place at the bank,” Norris said. He added that AI agents can be monitored more closely than humans.
Dawson said FNBO’s 15 financial crimes investigators and analysts are excited to be freed from some of the basic investigation so they can focus on higher-value analysis.
Dawson estimates that these AI agents save financial crimes investigators about 50% of their time.
“if [enhanced due diligence] “On average, an incident took 30 minutes, but now it may take as little as 15 minutes,” he said. AI agents are also helping banks achieve their real purpose: identifying and detecting crimes and empowering law enforcement to take action. “This allows us to balance our resources with interdiction and meeting needs,” Dawson said.
what’s next
In the future, FNBO may use AI agents to investigate fraud cases. Nasdaq Verafin announced its Agent Fraud Analyst earlier this month.
“Start easy, build confidence and validate your ability to monitor and manage these agents,” Baxter said. “We started with relatively low-risk work that was easy to explain and monitor with consistent output. This gives operators, supervisors, managers and regulators confidence that we are moving forward with the right platform, and then allows us to be more adventurous.”
Subrt issued several warnings to banks looking to use AI agents to investigate financial crimes. For example, in the case of work related to sanctions review or enhanced due diligence, he said that reduced accuracy “has direct regulatory implications and requires close attention throughout design, development, deployment, and ongoing maintenance and scrutiny over the life of the agent.”
Explainability is key, he said. “An agent’s work and profile must be traced back to its underlying source,” Sabuto said. “Otherwise, you’re introducing more risk than you are mitigating it.”
Establishing and maintaining data integrity, quality, and integrity is equally important. “For example, if an agent omits relevant signals or overestimates unimportant signals, false or incomplete records may be created, potentially negatively impacting investigative efforts before analysts or investigators have even begun the process,” he said.
Educational institutions must also be wary of biases and illusions that can impair the human judgment that solutions are intended to support. Financial crime syndicates should be trained and encouraged to monitor and challenge the performance of their agents, rather than just blindly consuming, Sabuto said.
Governance is a further consideration, he said. “Agents need to operate within a financial institution’s own policies, risk appetite and escalation thresholds, backed by clear audit trails and historical review prior to deployment, so that performance is demonstrated rather than assumed,” said Sabuto. “Ultimately, these agents fall within existing model risk and third-party governance expectations, and institutions remain accountable for outcomes regardless of the tools used.”
Despite these caveats, Subrt expects the adoption of these types of AI agents by banks to accelerate over the next 12 to 24 months, particularly when it comes to enhanced due diligence and reviewing sanctions alerts, which could place the greatest burden on manual investigations.
He said the banks he is currently investing in are “setting the operating standards by which other banks will be evaluated.”
