A back-office problem finally gets a real fix
Client onboarding at a custody and wealth bank has always been a data-assembly problem disguised as a relationship-management task. Analysts pull account details, compliance documentation, and account structures from systems that were never designed to talk to each other, then manually reconcile the result into something a relationship manager can act on. Northern Trust says that process used to take two and a half weeks per client. Its new AI agent, built with Microsoft's forward deployed engineering teams on Azure and Microsoft Foundry, now does the same consolidation in three days.
That is an 83 percent cut in elapsed time on one of the least glamorous parts of private banking, and it matters precisely because it is unglamorous. Onboarding delays are the kind of friction that shows up in client satisfaction scores and lost mandates long before anyone notices it in an earnings call. By going after the data preparation bottleneck specifically, rather than the onboarding process as a whole, Northern Trust picked a narrow enough target that the agent could be built, tested, and trusted in a reasonable timeframe.
The numbers behind the headline figure
The 83 percent figure is the one getting attention, but it sits alongside other internal metrics that show this runs deeper than a single demo. Northern Trust has also deployed client action plan agents that CEO Michael O'Grady says save relationship managers two to three hours per client memo, time that previously went into drafting summaries rather than talking to clients. Combined with broader AI-enabled digitization work across the firm, straight-through processing, transactions that complete without manual intervention, has nearly doubled over the comparison period the bank disclosed on its earnings call.
Those figures earn their place in a board deck, and they also reveal where the real leverage in this kind of project sits. The gains trace back to removing mechanical work that sat in front of judgment calls: formatting records, reconciling mismatched fields, summarizing documents into a usable brief. Analysts still make the actual decisions. CIOs evaluating their own agent pilots should study that pattern closely rather than fixate on the headline percentage. Targeting the preparatory grind ahead of a decision, instead of the decision itself, is what made this ROI case easy to defend to a board and a regulator at the same time.
Governance was the design brief, not an afterthought
O'Grady described the onboarding agent on Northern Trust's Q3 2026 earnings call as combining "advanced AI with enterprise-grade privacy, security and governance to make the process faster for clients." That ordering is deliberate. Analysts retain the ability to review and approve the agent's work and to inspect its reasoning before anything is finalized, and the system is built as what Northern Trust calls a governed intelligence layer that preserves onboarding knowledge so the agent accumulates expertise over time rather than starting from zero on every client.
Wealth and custody banking runs on exactly the kind of client data that regulators and clients alike expect to see handled with a documented chain of custody, which makes that governance structure especially valuable here. Shipping an agent without a built-in audit trail carries both a technical risk and a regulatory one, since examiners will eventually ask who approved what and on what basis. Northern Trust's choice to treat oversight as a design constraint from the start, rather than a control bolted on after a pilot succeeds, is the part of this story that travels well to other regulated industries, independent of which vendor stack a given CIO happens to be running.
Why the Microsoft forward deployed model fits this problem
Northern Trust built the agent with Microsoft's forward deployed engineering teams rather than through a standard software license handed off to a systems integrator once the contract was signed. Bill Borden, Microsoft's corporate vice president for worldwide financial services, said the approach "points to a broader opportunity across regulated industries," where off-the-shelf AI tooling rarely fits the compliance and data-handling requirements that firms like Northern Trust operate under every single day, regardless of how polished the vendor's demo environment looks during a sales cycle.
Forward deployed engineering puts vendor staff directly inside a client's build process, closer to a co-development arrangement than a traditional software implementation handled at arm's length through a systems integrator. For a regulated bank, that model shortens the distance between what the business actually needs and what gets shipped, because the people writing the integration code sit in the room when onboarding requirements get clarified, rather than receiving a requirements document weeks after the conversation that produced it. This approach costs more per project than licensing a packaged agent off the shelf. For a workflow handling sensitive client financial data under active regulatory scrutiny, that extra cost buys a level of fit and accountability packaged software rarely delivers on its own.
The sector context: this is becoming standard practice
Northern Trust is moving alongside a wider shift in financial services, not ahead of an industry that has yet to catch up. A recent Evident survey found that more than two-thirds of global financial firms have already added formal controls for AI systems, which puts the governance posture Northern Trust is advertising much closer to table stakes than to a genuine competitive edge. Execution speed and the depth of those controls, rather than their mere existence on paper, now separate the firms getting real value from the ones producing compliance documentation that nobody outside the audit team ever reads closely.
Northern Trust has also rolled out an internal AI assistant called Byron, which supports wealth managers preparing tailored client insights, a sign that the onboarding agent is one piece of a broader internal AI program rather than an isolated showcase project built for a single earnings call. Several connected agents, each targeting one specific workflow, describe how enterprise AI actually gets deployed inside a large regulated firm far more accurately than the single flagship launch that most vendor marketing prefers to highlight. The unglamorous, incremental version is the one worth studying.
What CTOs and CIOs should take from this
The lesson for technology leaders outside financial services centers on sequencing rather than on the specific Azure and Microsoft Foundry tooling involved. Pick the highest-friction, most mechanical step in a regulated workflow first. Build governance into the agent's architecture from day one instead of layering it on after a successful pilot. Measure the result in hours saved per transaction and in audit trails preserved, rather than in the kind of abstract productivity scores that rarely survive contact with a skeptical board.
Northern Trust's 83 percent figure will get cited in plenty of AI strategy decks over the next few months, often stripped of the governance context that made it defensible to a board and a regulator in the first place. The number itself is real and worth noting. The operating model underneath it, human review retained at every step, reasoning kept auditable, and institutional knowledge accumulated across clients instead of discarded after each engagement, is the part that actually transfers cleanly to other regulated industries trying to make the same case internally.



