- Key Insight: OceanFirst has started letting employees use Microsoft Copilot and is reaping time savings from it.
- Expert Quote: "Every app now has AI; we must leverage it responsibly" — Brian Schaeffer, CIO.
- Supporting Data: 150 employees have been onboarded to Microsoft Copilot.
At OceanFirst Bank, anti-money-laundering work is being done faster and with fewer people, thanks to AI, while the work needed to examine bond portfolios has gone from a six-to-eight-hour process to something that can be done in 15 minutes.
"This has been what seems like a lifelong pursuit of AI," Brian Schaeffer, the bank's chief information officer, told American Banker, adding that AI is "inherent in everything we do."
"I don't know that we chose that," Schaeffer said. "But every application we own now has an AI element to it, so it's incumbent upon us to leverage that."
OceanFirst, which has $13.3 billion in assets and is based in Toms River, New Jersey, started with an extensive data cleanup project, onboarded about 150 employees onto Microsoft Copilot and is now deploying a data infrastructure layer from Databricks to make the other AI models it's using more effective.
With these projects, OceanFirst is part of a broader industry trend among U.S. banks. In an
"High-quality data is critical for successful AI deployment," said David De Leon, fraud and financial crime lead in Accenture's Finance Risk Compliance practice. "Over the past 18 months, data governance and quality have been a major industry focus, including increased regulatory enforcement. While data cleanup is essential, it can be a never-ending challenge. Most banks are striking a balance between improving data quality and moving forward with AI initiatives — rather than waiting for perfection — because AI solutions depend heavily on the integrity of underlying data."
Step 1: Data cleanup
OceanFirst has been cleaning and organizing its data to feed it into AI models.
"It all starts with data," Schaeffer said. "I can't say that enough. Data is the lifeblood of everything. If your data is bad, nothing's going to work."
His team has spent the past year and a half on this data cleaning work, which is "not a fun process," Schaeffer acknowledged: "We literally go business line by business line and go through all their critical applications by fields, and start logging them and registering them." Then his team maps all the data elements and puts rules around them. Every department is audited and graded every year on the quality of their data.
In the next phase of data improvement, Schaeffer's team has begun feeding data into a data infrastructure and analytics layer from Databricks.
"I can pull in anything from a spreadsheet to a SQL Server and do comparisons automatically there," Schaeffer said. Databricks also uses the model context protocol, which means users can query data using large language models like ChatGPT. OceanFirst employees use this to ask questions about deposit and loan data.
"You can say, show me all the loans for this geography or this type of loan," Schaeffer said. "It kicks out a spreadsheet."
OceanFirst also uses a tool called MagicMirror that protects data from AI tools. "It lets you run any other AI model that you want, like ChatGPT or Anthropic, but it filters out all the sensitive data," Schaeffer said.
Money laundering investigators armed with Copilot
Meanwhile, the bank invested in Microsoft Copilot, the tech company's large language model, and started letting employees use it.
Money laundering investigators at OceanFirst use Microsoft Copilot for enhanced due diligence – deeper, risk-based investigations into high-risk customers and transactions that attempt to uncover potential financial crimes by scrutinizing ownership, source of funds and transaction patterns to prevent money laundering, terrorist financing and sanctions evasion.
The bank typically has 100 to 200 enhanced due diligence cases to work through per day. Often, these relate to complex companies, where determining the true nature of the entities behind a transaction can take six hours or more. Copilot helps get that down to five minutes.
The investigators copy data and files on suspicious transactions from the bank's Fiserv core system into Microsoft Copilot, as well as data from external sources like Verafin, which aggregates information about consumers to "help paint out the picture of risk that we need to look up," Schaeffer said. Copilot then helps them determine whether a transaction is truly high risk.
The primary benefit of using Copilot in money laundering investigations is that it saves time, Schaeffer said.
His team is building a process that will automatically feed transaction data into Databricks, and from there it can be queried by large language models. As Databricks starts to be used in the process, he expects it to get even faster.
At some point, AI may rank transactions based on how risky they seem, and recommend which ones investigators should take a closer look at, Schaeffer said.
"As we're looking through the workflows now, we're trying to understand how to do that better, smarter, faster, with the tools we have, and customize it to us," versus prepackaged AML software that doesn't always do things exactly the way bank wants to, Schaeffer said.
Many banks do use prepackaged AML software from companies like ACI, ComplyAdvantage, Thetaray and Quantexa. At Quantexa, leaders say these specialized models can do a better job at the all-important chore of entity resolution, the process of identifying and linking records that refer to the same real-world entity, like a person, company or product, across different data sources.
"The way we pattern that out through the use of graph technology and graph analytics on top, we feel reasonably certain that we find the right risk for the right people," said Andrea Walser, Quantexa's head of AML solutions for North America. "Because what we're looking for is not just a single pattern of behavior that looks like a red flag. We're looking at behavior patterns plus the attributes of risk that are associated in that particular customer's network. If you have the same fact pattern of a high-dollar wire going from point A to point B … you have a much higher probability of having a useful case brought to your investigators."
Large language models "often rely on less-governed datasets and carry a risk of hallucination, which can make outputs unreliable," De Leon said. "General-purpose LLMs without system-driven guardrails can lead to variation and inconsistency, which may be counterproductive."
Analyzing bond portfolios
OceanFirst employees are also using Microsoft Copilot to monitor the bank's bond portfolios for changes in rate and such, pulling data in from other sources like Moody's. "I can hit those comparisons instantaneously, every five minutes now," Schaeffer said.
They're testing using Databricks in this process. "We've invested heavily, energy-wise at least, back to the data layer, into Databricks," Schaeffer said. "What Databricks allows us to do is to build the hooks in for deeper analytics." This project is still in beta.
Schaeffer and his team are thinking about many potential AI projects, such as one that would automatically send helpful data to corporate customers.
"We already have customers now where we send them files through secure file transfer protocol to help them do their accounting well," he said. "What if that happened automatically every month, just to simplify their world?"
Schaeffer, who formerly was the bank's chief information security officer, is also looking to use AI in cybersecurity. AI could analyze data, looking for the threats that law enforcement partners tell the bank about, compare that to what customers are seeing on their side, "and send a message to a customer through internet banking that says, 'Hey, your pattern of transactions matches XYZ bad guy, you might want to change your password or check your balances,'" Schaeffer said.
With all these projects, part of the goal is to let people use AI on their own, without a lot of IT involvement.
"At the end of the day, what I really want is democratization of the data and the processes," Schaeffer said.
