Northern Trust: AI agent cut client data prep from 2.5 weeks to 3 days

Northern Trust's client onboarding team now takes three days, not two and a half weeks, to prepare a new institutional client's investment history for loading, the bank and Microsoft said in a joint press release on 9 November 2026. They put the cut at 83%. The work is done by an AI agent, software that carries out a multi-step task on its own, with analysts reviewing what it produces.

Microsoft co-wrote the release, so read it as promotion.

The job

The agent serves clients joining Front Office Solutions, Northern Trust's performance and analytics platform for foundations, endowments, pension schemes and large family offices. A new client arrives with years of transactions and cash flows in whatever spreadsheet layout its last provider used. Until now, CIO reports, Northern Trust staff mapped and converted that data by hand.

Everybody's spreadsheet is different. But there are core things you can figure out: 'Oh, this is a capital call. This is a valuation.'

Pete Cherecwich, chief operating officer, Northern Trust, in Microsoft's feature

The agent converts each history into the bank's standard format, and analysts review and approve its work in Excel and SharePoint. After that, Microsoft's feature says, the bank runs its usual reconciliation, checking that its books match the previous administrator's, before the client goes live.

The numbers, and what sits behind them

• Time. A footnote says Northern Trust measured real client onboardings after deployment and counted data conversion, analyst review, exception handling and final validation. My arithmetic on the roughly 2.5-week baseline: 2.5 weeks is 17.5 calendar days, and 3 divided by 17.5 is 17%, which gives the 83% cut. In working days (12.5) it would be 76%. The release does not say how many onboardings were measured.
• Quality. The release reports a 96.5% "F1 first-pass quality score", a test score that combines how much of the agent's output was right with how much of the correct answer it found. A footnote says it comes from seven test cases built from anonymised past client workbooks, scored against answers analysts had prepared and the agent could not see, before any human correction.
• The other 3.5%. Microsoft's feature describes it as items the agent is unsure how to categorise, which go to a person. The footnote describes a test score. Neither is an error rate on live clients.

The people side

Microsoft Frontier Company, a venture Microsoft launched in July with $2.5 billion according to CIO, placed its engineers inside the bank. They and Northern Trust's finance staff wrote a library of rules for the agent, which feedback from the bank's team helped expand. Microsoft says it tuned the agent on anonymised data only and that client data is not used to train outside models.

Cherecwich told CIO the gain is staff spending "less time moving data around and more working with clients". No source says how many analysts do this work or whether that number has changed. Microsoft's feature says the bank has offered Copilot training to all 23,000 employees and that staff have built more than 15,000 Copilot agents. Asked by an employee what AI means for jobs, Cherecwich said "jobs will change. You have to learn new skills." He also drew a line: "We are not going to have AI do $100-million payments on its own anytime soon."

What the sources leave out

• Cost. Nothing on what Microsoft's embedded engineers cost, what the agent costs to run per onboarding, or how long the build took.
• Scale. No count of clients onboarded with the agent or of the production cases behind the three-day average.
• Errors in production. No figure for mistakes that got past analyst review.
• Upkeep. No word on who maintains the rule library once Microsoft's engineers leave.

What to do

If you run client onboarding or finance operations, look at the step Northern Trust chose: converting data that arrives in many formats, where a reconciliation already exists downstream to catch mistakes. Before a pilot, time your current process end to end, review included. Hold back finished cases with known-good answers and score the tool against them. Ask any vendor quoting a figure like 83% how many live cases it covers, what the error rate was after human review, and what the embedded engineers cost. Keep the reconciliation step.