At Northern Trust, finance experts and Microsoft engineers partner to move faster on AI

The AI agent's performance over time, or hill-climb in AI terms, has improved to hit a 96.5% FI first-pass quality score, a standard AI evaluation metric that combines precision and recall. The remaining 3.5% - stuff the AI agent is unsure how to categorize - is referred to a human to review.

This part is key; the Northern Trust team wanted to keep a human in the loop to avoid potential wrong information.

The company then proceeds with their regular reconciliation process - making sure their books and records match that of the previous administrator - before going live.

Using the AI agent has slashed the time needed to map transaction data from new clients from two and a half weeks to three days.*

"By using Microsoft's FDEs, I got speed to market for something that might have otherwise been deprioritized," says Cherecwich. Now, "speed to revenue is much faster. The client's happier."

All client data stays within Northern Trust and isn't used to train any outside AI models. The bank is now looking at how AI agents can be used with other products and services.

Northern Trust has laid the ground for using more AI, offering training to all 23,000 employees on Microsoft Copilot. Employees have created over 15,000 Copilot agents to do things such as dig out specific data points from company prospectuses. To spur interest, the bank recently ran a contest with a cash award for best Copilot agent.

They're also addressing employees' questions. "Somebody asked me, 'What does this mean for jobs?'" Cherecwich says. "My answer was: One of my first jobs was looking for municipal bonds on microfiche. Nobody is looking at microfiche anymore. So OK, jobs will change. You have to learn new skills."

And while some things are changing fast, others will stay the same.

"We can be cutting edge if it's low risk," he says, adding: "We are not going to have AI do $100-million payments on its own anytime soon."

Footnote*: Following deployment of the client onboarding agent, Northern Trust measured preparation times for real client onboarding scenarios. The analysis included data transformation, analyst review, exception handling, and final validation. The resulting improvement was calculated by comparing the prior workflow baseline of approximately 2.5 weeks with the average time required to complete measured production cases using the agent-assisted process.

Top photo: Pete Cherecwich, chief operating officer of Northern Trust, stands in front of the bank's headquarters in Chicago.

Second photo: Pete Cherecwich, chief operating officer of Northern Trust, at work inside the bank's Chicago headquarters.

(All photos by Alex Garcia for Microsoft.)

Chen May Yee writes about AI and innovation for Microsoft. Previously, she was a staff correspondent for Agence France-Presse and the Wall Street Journal and an editor at the Singapore Straits Times and the Minnesota Star Tribune, covering topics from business to health to technology. You can contact May Yee on LinkedIn.