A proof of concept at St. James's Place shows what happens when a financial services firm puts governed AI to work on the jobs nobody has time for.
Every wealth management firm has a backlog it doesn't talk about publicly. Not the client-facing stuff, the plumbing behind it: pipelines that need rewriting, survey data that needs reading, platform consolidation projects that keep losing their best engineers to firefighting instead of building.
St. James's Place, one of the UK's largest wealth management businesses, had exactly that problem. Two areas stood out. First, the firm was processing thousands of client survey responses by hand, a job that ate roughly 4,000 engineering hours a year just to extract sentiment and themes. Second, an ETL migration tied to platform consolidation was crawling, with pipelines that took days each to rewrite and validate manually.
Neither problem is unique to SJP. Most financial services firms are sitting on some version of this same backlog, and most are reluctant to hand it to AI. Regulated data, audit requirements, and a justified suspicion of vendor claims all make caution the default. So instead of a rollout, SJP ran a proof of concept with Maia, an AI Data Automation platform. Worth being direct about what that means: this is not a customer deployment, and SJP is not a paying Maia customer. It was a bounded test, designed to answer one question. Could agentic AI support real data workloads in a regulated financial-services environment while operating within the firm’s security and governance requirements?
The results, at least on the two problems tested, were hard to ignore. Sentiment analysis that had previously required around 4,000 hours of manual effort annually was completed in roughly 16 hours during the proof of concept. Maia described this as a 1,300% efficiency gain. The ETL migration effort dropped by around two-thirds, turning multi-day jobs into work measured in hours.
The two-thirds reduction in migration effort matters less for the time saved in isolation and more for what it lets the team do next. As Kelly Maggs at SJP put it: "We recognized that platform consolidation would help us to reinvest in the team, to enable us to build out more on the SAP and AI roadmap tomorrow."
That's the part worth pausing on. The productivity number is the headline, but the more interesting outcome is what the POC proved about trust. Financial services firms don't adopt AI tooling because it's fast. They adopt it because it can be governed, audited, and trusted not to introduce risk into a regulated process. SJP's test wasn't set up to chase a speed record. It was set up to check whether an AI system could operate inside the constraints that actually matter in this sector, and only then to see what speed followed from that.
This is where a lot of AI-in-finance coverage gets the emphasis backward. The interesting question isn't whether AI can move fast. Plenty of tools can move fast in a demo. The interesting question is whether it can move fast and stay inside the governance, auditability and control requirements expected in a regulated financial-services environment. A 1,300% efficiency gain means nothing to a CDAO if it comes with an unauditable black box. SJP's POC is a data point suggesting those two things aren't mutually exclusive, at least for the two workloads tested.
None of this makes the case that every wealth manager should hand its data backlog to an AI agent tomorrow. A proof of concept is a proof of concept: bounded scope, specific workloads, one firm. What it does suggest is that the backlog itself, the manual sentiment coding, the years-long ETL migrations, the platform consolidation projects that never quite finish, is a more tractable target for agentic AI than the industry's caution has assumed. The barrier was never really speed. It was whether the work could be trusted. SJP's test is one answer to that question, not the final one.