Northern Trust deploys AI agent built with Microsoft to speed client onboarding

Northern Trust is putting an AI agent to work on years of investment records before new institutional clients go live. Announced with Microsoft on November 9, 2026, the deployment converts historical transactions into the bank’s standard format. The companies say it cuts data-preparation time by 83%, with analysts still reviewing and validating the work.

The agent serves Northern Trust Asset Servicing’s Front Office Solutions platform, which provides performance reporting and analytics for foundations, endowments, pension schemes and large family offices. Bringing those clients onto the platform can mean combining investment records from multiple sources and reconstructing performance histories, particularly for alternative assets.

Teaching the agent what the spreadsheets mean

The two companies began exploring the onboarding problem earlier in 2026. Northern Trust’s finance specialists worked alongside Microsoft Frontier Company’s forward-deployed engineers—technical staff embedded with customers to build systems for specific business tasks. Their joint assignment was to translate transaction histories and cash flows arriving in different formats into records Northern Trust could use.

A same-day Microsoft feature describes how the team built a library of rules around the bank’s financial vocabulary. Different firms may use different terms for the same transaction. The team used anonymized records from previous onboardings, including how those records were ultimately organized, and expanded the rules through feedback from Northern Trust experts.

Northern Trust chief operating officer Pete Cherecwich said access to Microsoft’s engineers accelerated a project that might otherwise have been deprioritized. He linked the faster onboarding work to reaching revenue sooner and improving clients’ experience.

Two measurements, different evidence

The timing result covers more than the agent’s initial conversion of data. Northern Trust says its production measurements included transformation, analyst review, exception handling and final validation. It is a measure of the preparation workflow, not a claim that the entire process of bringing a client onto the platform takes three days.

The companies also report a 96.5% first-pass F1 score. F1 combines precision and recall into one measure of output quality. That result came from controlled testing across seven representative cases, using real but anonymized historical client transaction workbooks supplied by Northern Trust.

For that test, agent-generated values were compared with reference outputs prepared by analysts and withheld during execution. The score was measured before analyst feedback or corrections. It describes a bounded first-pass test, rather than the quality of the final records after people have checked them; both performance figures are company-reported.

Analysts keep the approval step

The agent runs on Microsoft Azure, uses Microsoft Foundry and connects with Excel and SharePoint. Those integrations let analysts review, refine and approve its work. An Azure-based knowledge layer retains approved onboarding knowledge so the system can draw on the bank’s accumulated expertise.

Analysts can examine outputs and reasoning, resolve exceptions and feed corrections back into the process. Northern Trust also retains its regular reconciliation before a client goes live: checking that its books and records match those of the previous administrator. The agent operates during implementation, when errors in starting data can still be identified and reconciled.

The companies say Microsoft used only anonymized data to improve evaluation scores, while client data remained under Northern Trust’s control. Microsoft’s feature also says client data is not used to train outside AI models. Those are the stated data boundaries around this deployment, alongside the human checks built into the workflow.