“Just getting a sense of this also helps us to prepare,” says Pattanaik, the head engineer in charge of BNY’s artificial intelligence division. “And this is exactly what we do to our salespeople and relationship people.” Fortunately, the AI’s assessments of a wide array of business opportunities—and of me—remain locked within a database on-premises alongside the bank’s $2 trillion in assets under management.

BNY’s new AI tool is named Eliza in honor of the wife of Alexander Hamilton, who founded the bank’s corporate predecessor. The physical part of Eliza’s brain consists of Nvidia microchips and cloud infrastructure powered by Microsoft’s Azure and Google Cloud. But Eliza’s mind, the way she approaches problems, is a BNY original, using OpenAI’s GPT-4, Google’s Gemini and Meta’s Llama.

While most open-source AIs are essentially one-size-fits-all virtual assistants, Eliza lets employees create bespoke assistants (BNY prefers the term “agent”) and fill them with proprietary data in order to take on particular tasks. The agents operate much like the consultants who have long helped bankers by offering niche subject matter expertise on matters like payments, collateral and compliance. But unlike human consultants, Eliza’s army of agents are low cost and easy to cast aside when the task is done.

“I don’t think people realize,” says BNY managing director, Victor O’Laughlen, “that if it takes you five hours to research a topic to answer a question, why wouldn’t you spend two minutes to upload all the relevant information, ask the question and just throw the agent away?”

Eliza adoption spreading rapidly

Eliza is currently being used by a quarter of the bank’s employees—about 14,000 people— including O’Laughlen. As the head of BNY’s global clearance platform, which conducts more than $15 trillion worth of securities transactions a day, O’Laughlen is currently training what he calls an “operating model Subject Matter Expert.” Its mission is to implement the Agile project management system, a popular tool for software developers, in a way that gives front-line employees more autonomy.

By relying on Eliza to implement the process, rather than expensive Agile consultants, BNY saves on training costs. Agile training is, of course, just one example of the multitude of one-off tasks that the bank pays human specialists to conduct and, if BNY can replace most of those with Eliza-driven agents, the savings could potentially add up to billions.

Regardless of how that plays out, O’Laughlen says the goal for his Eliza virtual assistant is to identify second-order connections on his team’s org chart—people who don’t report to each other, but who rely on each other—and track the value of the collective workflow. This might include how work done in operations informs an operations lead, which then informs decisions made by the product manager—but on a massive scale.

“Getting it wrong means that maybe someone’s expertise was missed early in the process, you identified the issue much later in the development, and then it costs you two or three times the amount of money to fix it later than it does earlier.” Thirty percent of his team is already using Eliza, a number he expects will double by the end of the year, especially as employees come to appreciate the efficiency it offers.

“When people talk to you, sometimes there are barriers,” says Pattanaik. “The agents have no barriers. They do not have any political agenda. They are just going to get information for you.”

A hub and spoke model for AI

The genesis of Eliza goes back to the summer of 2023 when Robin Vince, CEO of BNY, called Pattanaik into his corner office on the bank’s executive floor. Vince wanted to address the problems the bank was experiencing as a result of having disparate AI projects being undertaken in offices around the world—resulting in development times as long as 18 months.

Vince told Pattanaik to fold the projects into a single centrally controlled AI platform. In addition to being more efficient, Pattanaik hoped he could make it more resilient against employees who might accidentally leak the bank’s secret sauce or client data. He called a meeting with the managing directors of each bank business line, including Securities Services, Market and Wealth Services and Investment and Wealth Management. The ensuing sessions resulted in hundreds of possible applications.

What they ended up with is a hub and spoke system, per Vince’s orders, where Eliza herself is at the center, and engineers at each business line, representing the spokes, customize aspects of her software for their team. The thousands of employees who have passed a two-hour session on how to safely handle data and interpret and double-check the responses can build their own agents.

BNY isn’t sharing exactly how many agents have been created, but some of them include a benchmarking tool that analyzes funds, flags deviations and helps accountants address them; and a tool to help investment managers customize portfolios. Twenty of these agents have percolated to the surface, the cream of the AI crop, and been approved by BNY management for broader use.

In the future, the savings will be amplified, according to Pattanaik, when the virtual assistants start to “stitch together” observations from agents with different expertise. “Some of the capabilities that we have,” says Pattanaik, “specifically where you have multiple agents that talk to each other and create an economy of their own: mind blowing. That is what I think is the future.”

Banking jobs on the block

Of course, all this efficiency comes with a trade-off: jobs. Already, Goldman Sachs and JPMorgan have launched their own proprietary Eliza competitors, and a Citigroup report in June estimates that only 12% of banking jobs will actually be augmented by AI. Instead, the report predicts that more than half—54%—of banking jobs will be automated out of existence entirely.

To get a sense of the jobs at stake, note that BNY has 53,400 employees. JPMorgan has 240,000 employees and Goldman Sachs has 45,300. Across the U.S. there are 1.3 million commercial banking employees, according to Federal Reserve Bank of St. Louis estimates. If Citi’s report plays out, that would mean 702,000 U.S. banking jobs would eventually be eliminated.

In anticipation of AI’s impact on jobs the European Union earlier this month introduced a framework for dealing with what it considers “high risk” use of AI, including for “monitoring or evaluation of persons in work-related contractual relationships.” Last year President Joe Biden signed an executive order instructing the Secretary of Labor, Julie Su, to publish best practices for how employers should deal with AI displacing jobs.

Back at BNY, the bank is preparing training sessions for employees using AI and Pattanaik is working on Eliza 2.0, which he hopes will be able to converse with branch managers, help accountants read trust agreements and lengthy new regulations to see what exactly the bank needs to do to comply. As my conversation with Pattanaik wound down I started to ask him if there was “anything else that you want to add” —then corrected myself—“that you were hoping maybe the AI told you I was going to ask? “No,” he replied. “The AI actually did well in terms of preparing me.” Add PR to the list of skills at risk.