Northern Trust's client onboarding agent cut institutional data preparation from about 2.5 weeks to 3 days, and the useful part for anyone copying it is how the bank measured that. The time saving was measured on real production onboarding cases, end to end, with analysts reviewing and approving the agent's work. The 96.5% quality score comes from seven anonymized test workbooks. If you run operations at a bank, custodian or asset servicer, copy the measurement design and the review step, then build a much larger quality sample before you scale.
Northern Trust and Microsoft announced the deployment on November 9, 2026. The agent is live for clients joining Northern Trust Asset Servicing's Front Office Solutions platform, which provides performance reporting and analytics to foundations, endowments, pension schemes and large family offices.
What the Northern Trust Onboarding Agent Actually Does
The agent converts a new client's historical transaction data into Northern Trust's standard format so the bank can rebuild the client's performance history. Before it existed, staff mapped and transformed that data by hand before loading it into the reporting environment.
That job is harder than it sounds. Alternative investments can produce "an enormous amount of data" in mixed formats with years of transaction history, Northern Trust COO Pete Cherecwich told CIO. In the Business Wire release, the bank says 57% of asset owners it surveyed named consolidating data from disparate sources as their biggest challenge. The release gives no sample size or survey date for that figure.
The stack is all Microsoft. Per the Microsoft announcement, the agent runs on Azure, uses Microsoft Foundry, and hands its work to analysts in Excel and SharePoint, where they review, refine and approve it. An Azure "governed intelligence layer" keeps approved onboarding knowledge so later onboardings can reuse it.
The division of labor is plain. The agent does the pattern-matching transformation, and analysts take the cases it is unsure about. Cherecwich described the agent as doing "a lot of the heavy lifting" while staff handle uncertain cases (CIO). Analysts use it during implementation, before go-live, to catch and reconcile errors in the client's starting data (Microsoft).
How Northern Trust Measured the 83% Time Cut
The 83% figure is a production number, and the release footnote says what it covers. Northern Trust measured real client onboarding cases after deployment against a baseline of roughly 2.5 weeks, and the measured time includes transformation, analyst review, exception handling and final validation. Three days is the average across those production cases. The footnote does not say how many cases there were.
Counting review and exceptions inside the clock is the decision worth copying. Most agent pilots report the agent's run time and leave out the human time spent checking it, which is how a "minutes instead of days" claim turns into a team that is no faster. Here the analyst's review hours sit inside the three days.
Do the arithmetic before you quote the number internally. An 83% cut from 2.5 weeks to 3 days works if both figures are calendar days (3 of 17.5). If the 2.5-week baseline means 12.5 working days, the cut is closer to 76%. Either way it is a large saving, but your finance team will ask which one, and the release does not say.
Cherecwich framed the return as speed to revenue, client experience and efficiency together rather than any single metric (CIO). For a reporting platform, that framing makes sense: a client whose history is rebuilt two weeks sooner starts paying for the service two weeks sooner. Neither company disclosed what the build cost, and Banking Dive's report gives no budget, no count of Microsoft engineers involved and no rollout timeline.
What the 96.5% F1 Score Does and Does Not Tell You
The 96.5% F1 score is a first-pass quality number from controlled testing, and it rests on seven workbooks. The release footnote says the test set was seven representative cases built from real, anonymized historical client transaction workbooks, scored against analyst-prepared reference outputs that were withheld while the agent ran. The footnote says the score "reflects first-pass results from controlled testing," measured "before analyst feedback or corrections."
An F1 score is the harmonic mean of precision and recall. In this case, precision is the share of the agent's mappings that were right, and recall is the share of the correct mappings it found (scikit-learn's definition). The design has two things to like. The reference answers were held back, so the agent could not be tuned to the answer key during the run. And Microsoft tuned its evaluation measures on anonymized data only, an "eyes off" approach that kept client data under Northern Trust's control.
The score also has three limits a buyer should weigh.
- Seven is a small number. If an event never occurs in n trials, the 95% upper bound on its rate is 1 - 0.05^(1/n), which the rule of three approximates as 3/n for larger samples. If none of the seven workbooks had a serious failure, the data still cannot rule out a per-workbook failure rate as high as about 35%. Northern Trust may have more evidence internally. The public number does not show it.
- F1 counts every error the same. A missed mapping and a wrong mapping cost the same in the score. In a rebuilt performance history they do not cost the same. A missed field surfaces as an exception an analyst has to clear, while a confidently wrong one can flow into a client's returns unless the reviewer catches it.
- "Representative" is the bank's word. The release does not say whether the seven cases include the messiest alternative-asset histories, which are where Cherecwich says the data volume and format problems sit.
The agent may well be as accurate as the score says. The public evidence for that is seven cases, though, and a bank planning its own version should test on many more. We have covered how many eval cases it takes before a difference is real and why hand-labeled cases beat synthetic ones. Both apply here.
Why the Analyst Review Step Is the Real Control
Analysts review the agent's outputs, evaluate its reasoning and resolve exceptions, per Microsoft's description, and that review is what lets a 96.5% first-pass score ship into a regulated workflow. The agent proposes, and a person signs off on the client's starting data before go-live.
Large banks are adding controls of this kind. Evident's AI index, published in October 2026, found that more than two-thirds of the 50 banks it tracks now run additional controls on AI systems, including checks on inputs and outputs, limits on which tools systems can use, and oversight once they are live. US supervisors have started asking banks to explain their agent safeguards directly.
A review step only works while reviewers keep reading. When an agent is right most of the time, people start approving its work without checking it, and our guide to human-in-the-loop approval gates covers how often that happens. The fix is cheap: seed known errors into a sample of the agent's output and track how many reviewers catch. Northern Trust has not said whether it does this, so build it into your own version from the start.
What Microsoft's Forward Deployed Engineers Brought
The agent was built with forward deployed engineering teams from Microsoft Frontier Company, the $2.5 billion unit Microsoft launched on July 2, 2026 with about 6,000 industry and engineering experts and a plan to send forward deployed engineers into customer offices. AWS had announced its own $1 billion forward deployed engineering organization two days earlier, and it describes customer engineers moving from observers to co-builders to operators who run the system themselves.
Cherecwich's main lesson, per CIO, was that the project needed Northern Trust's domain knowledge and Microsoft's AI engineering together, and that having Microsoft engineers embedded with bank staff sped up delivery. That is a credible account. It also leaves the questions a buyer has to answer before signing a similar engagement: who owns the agent, the prompts and the governed knowledge layer when the engineers leave, and what it costs to run once they do. We covered those terms in the context of IBM and Microsoft's Imagine Enterprise deal and Anthropic's own forward deployed program.
Onboarding is one of several agent projects at the bank. CEO Michael O'Grady has said client action plan agents save relationship managers two to three hours per client memo, and Banking Dive reports that AI-enabled digitization has nearly doubled the bank's straight-through processing. The pattern across them is the one Meta used when it cut data-column onboarding to under an hour: pick a narrow, repetitive data job with a clear right answer and a person who already checks it.
What to Do If You Run Client Onboarding
This Week:
- Time one recent onboarding end to end, including review and exception handling, in calendar days and in working days. That is your baseline, and without it no vendor's percentage means anything.
- Pull the last ten onboardings your analysts finished and keep their final mapped outputs. Those are the start of your withheld reference set.
This Month:
- Build an anonymized test set of at least 30 client workbooks, weighted toward the messiest alternative-asset histories, before you trust any F1 score. Report precision and recall separately so you can see wrong mappings apart from missed ones.
- Write down which errors are worse. A wrong cost basis matters more than a missed memo field; score them that way.
Before You Sign a Forward Deployed Engagement:
- Put ownership of the agent, its prompts, the evaluation set and any knowledge layer in the contract, along with the run cost after the vendor's engineers leave.
- Require a reviewer-accuracy check: seeded errors in a sample of agent output, with catch rates reported monthly.
The Bottom Line
Client data onboarding has been automated before. Rules engines and RPA handled the clean, repetitive formats and left analysts with everything else. Northern Trust's agent goes after the long tail those tools could not, and the bank measured its time saving on live cases with the human work counted. The quality claim needs a bigger sample before anyone treats it as settled. Start by timing your own last onboarding.
Continue Reading
- Only 3 of 36 Model Gaps Were Real. Size Your Eval Set.
- Human-in-the-Loop for AI Agents: Most Approval Gates Rubber-Stamp
- Meta Cuts Data-Column Prep to Under an Hour With 14 Subagents
- Enveda's Claude Reviewer Cut Review Time From 1,553 to 166 Minutes
- US Regulators Ask Banks to Explain Their AI Agent Safeguards
- At Sibos 2026, Bank AI Agents Fix Payments a Person Still Releases