- Key insights: OceanFirst has started allowing its employees to use Microsoft Copilot, which saves them time.
- Expert quote: “AI is in every app today, and we need to leverage it responsibly.” – Brian Schaeffer, CIO.
- Support data: 150 employees were onboarded to Microsoft Copilot.
At Ocean First Bank, AI has enabled anti-money laundering tasks to be performed faster and with fewer people, with the process of examining a bond portfolio taking 15 minutes instead of 6 to 8 hours.
“This seems to be a lifelong pursuit of AI,” Brian Shafer, the bank's chief information officer, told American Banker, adding that AI “is embedded in everything we do.”
Processing details
“I don't know if we would have chosen that,” Schaefer said. “But now every application we have has an AI element built into it, so it’s our duty to leverage that.”
OceanFirst, a $13.3 billion-asset Toms River, N.J.-based company, started with a large-scale data cleanup project, brought about 150 employees to Microsoft Copilot, and is now deploying Databricks' data infrastructure layer to make other AI models it uses more effective.
These projects make OceanFirst part of a broader industry trend in U.S. banking. in
“High-quality data is essential for successful AI implementation,” said David de Leon, fraud and financial crime lead in Accenture's Financial Risk Compliance practice. “Over the past 18 months, data governance and quality have been a key focus for the industry, including increased regulatory enforcement. Data cleanup is essential, but it can be a never-ending challenge. Rather than waiting for perfection, most banks are balancing improving data quality with advancing their AI initiatives, because AI solutions are highly dependent on the integrity of the underlying data.”
Step 1: Clean up your data
OceanFirst has been cleaning and organizing data to feed into AI models.
“Everything starts with data,” Schaefer said. “This is not enough. Data is the lifeblood of everything. If the data is bad, nothing will work.”
His team has spent the past year and a half cleaning this data, which Schaefer acknowledged is “not a fun process.” “We literally look at business line by line, we look at all our critical applications line by line, and we start logging and registering them.” His team then maps all the data elements and sets rules around them. All departments are audited annually to assess data quality.
As the next step in data improvement, Schaefer's team began feeding data from Databricks to the data infrastructure and analytics layer.
“You can take anything from a spreadsheet into SQL Server and automatically do the comparisons there,” says Schaeffer. Databricks also uses the model context protocol. This means users can query data using large language models like ChatGPT. Ocean First employees use it to ask questions about deposit and loan data.
“I can say, show me all the loans in this area or this type of loan,” Schaeffer said. “The spreadsheet will be launched.”
OceanFirst also uses a tool called MagicMirror that protects data from AI tools. “You can run any other AI model you want, like ChatGPT or Anthropic, but all sensitive data will be filtered out,” Schaeffer said.
Money laundering agent armed with co-pilot
Meanwhile, the bank invested in Microsoft Copilot, a large-scale language model for high-tech companies, and began making its employees use it.
Ocean First money laundering investigators use Microsoft Copilot for enhanced due diligence. This is a more in-depth risk-based investigation of high-risk customers and transactions that seeks to uncover potential financial crimes by scrutinizing ownership, funding sources, and transaction patterns to prevent money laundering, terrorist financing, and sanctions evasion.
The bank typically has to process 100 to 200 enhanced due diligence cases a day. These often involve complex companies, and it can take six hours or more to determine the true nature of the entity behind the transaction. Copilot helps reduce that to 5 minutes.
Schaefer said investigators are copying data and files on suspicious transactions from the bank's core system, Fiserv, to Microsoft Copilot, as well as data from external sources such as Verafin, which aggregates information about consumers, “to help us develop a complete picture of the risks we need to investigate.” Copilot helps you determine if a trade is truly high risk.
Schaefer said the main benefit of using Copilot for money laundering investigations is that it saves time.
His team is building a process that automatically feeds transactional data into Databricks, where it can be queried by language models at scale. He expects this process to become even faster once it starts using Databricks.
At some point, Schaefer said, AI could rank transactions based on their riskiness and recommend which transactions investigators should take a closer look at.
“Right now, as we're looking at workflows, we're trying to understand how we can do it better, smarter, faster with the tools we have and how we can customize it versus pre-packaged AML software that doesn't necessarily do things the way banks want them to,” Schaefer said.
Many banks use packaged AML software from companies such as ACI, ComplyAdvantage, Thetaray, and Quantexa. Quantexa leaders say these specialized models can do a better job at the crucial task of entity resolution, the process of identifying and linking records that refer to the same real-world entity (such as a person, company, or product) across different data sources.
“We have a high degree of confidence that by using graph technology and patterning it with graph analytics on top of that, we can find the right risks to the right people,” said Andrea Walser, head of AML solutions for North America at Quantexa. “Because what we're looking for is not a single pattern of behavior that looks like a red flag. We're looking at the pattern of behavior and the attributes of risk associated with that particular customer's network. If you have the same fact pattern of high-value communications going from point A to point B, you're much more likely to bring a profitable case to investigators.”
Large-scale language models “often rely on poorly managed datasets, run the risk of hallucinations, and can produce unreliable output,” de Leon said. “A generic LLM without system-driven guardrails can lead to fluctuations and inconsistencies, which can be counterproductive.”
Bond portfolio analysis
OceanFirst employees also pull data from other sources such as Moody's and use Microsoft Copilot to monitor things like interest rate movements in the bank's bond portfolio. “Now we can run these comparisons instantly every five minutes,” Schaeffer says.
We are testing this process using Databricks. “We've made significant investments in Databricks, going back to the data layer, at least on the energy side,” Schaefer said. “Databricks allows you to build hooks for deeper analysis.” The project is still in beta.
Schaefer and his team are considering a number of potential AI projects, including one that automatically sends useful data to enterprise customers.
“We already have customers who send files through secure file transfer protocols to ensure proper accounting,” he said. “What if it happened automatically every month to simplify the world?”
Schaefer, who previously served as the bank's chief information security officer, is also considering the use of AI in cybersecurity. The AI analyzes the data, looking for threats that law enforcement partners have communicated to the bank, and comparing them to what the customer is seeing, “so that through internet banking we can send a message to the customer saying, 'Hey, your transaction patterns match the bad guy at XYZ. Maybe you should change your password or check your balance,'” Schaefer said.
One of the goals of all these projects is to enable people to use AI on their own without much involvement from IT departments.
“At the end of the day, what I really want is democratization of data and processes,” Schaeffer said.
