AI adoption in asset management is accelerating to boost productivity, improve compliance processes, and strengthen data verification, but moving from pilot to production is proving slower and harder than expected.
The gap from pilot to production was the focus of a recent asset management breakout lunch during Connect Boise. Industry leaders shared what the challenges were in the transition and what was proving successful.
Themes included AI accuracy and verification, financial services requirements, human-in-the-loop design, and data readiness in asset management.
Clearwater’s own agentic software development life cycle (SDLC) process starts with what may be the hardest step: defining the use case. Before any AI capability moves forward, the team aligns on exactly what problem they’re solving. As one speaker put it: “If we don’t solve it, AI is not going to solve it.”
From there, ideas move through validation, testing, and a rigorous pre-production review that always includes humans in the loop. People are checking the work at every stage. AI isn’t used to test its own output.
For banks and other regulated firms, the biggest hurdle is oversight, not technical capability. One panelist described an 18-month rollout of Microsoft Copilot that expanded to include additional models as accuracy improved. Tasks now range from email review to full portfolio analysis, but client-facing deployment is off the table.
Questions remain over how accuracy is verified. What disclosures are required? The SEC and OCC have not clarified how they view AI-assisted client communications. For these reasons, that panelist said the firm’s AI use stays internal, with roughly an 80/20 split between AI output and human review.
That 80/20 split came up repeatedly in the discussion. Across firms, human-in-the-loop isn’t a transitional phase. It’s treated as a permanent, table-stakes feature of any agentic workflow.
The fastest win across the room in using AI was personal productivity such as summarizing emails, transcribing meetings, and surfacing conversations. Several firms have also had success using AI for first-pass reconciliations, like flagging common breaks or checking them against historical patterns before a human signs off.
One firm has gone further, building its own internal agentic platform that spans multiple models and connects to its internal research and data systems, so client data never reaches external providers. That firm is still early in developing agents and is deliberately slowing down in the area of access control. Determining who can access HR data, client data, or personally identifiable information (PII) is being handled with real caution.
For most firms, data was the biggest issue. One panelist, drawing on prior experience at a custodian bank, noted that even before machine learning tools existed, 60–80% of project time went into simply sourcing and preparing data, not building models. That dynamic hasn’t gone away with AI; if anything, it’s been exposed more starkly.
AI performs best on data that’s already clean, reconciled, and governed, which several panelists identified as one of Clearwater’s structural advantages. Where firms are still struggling is upstream, using AI to help produce that clean data in the first place, whether that is from accrual systems or an OMS.
AI adoption also surfaced older, quieter problems like data vendor licensing. Rules that human analysts had internalized over years (which datasets they were and weren’t permitted to touch). Once AI entered the picture, these had to be re-codified and automated, since the models had no inherent sense of those boundaries.
The most vivid story about data accuracy came from a real estate finance professional who used AI to reconcile property-level and holding-company-level financials. A large, structured task the model handled quickly.
A manual double-check revealed simple addition errors buried in the AI’s output. The tool was roughly 95% accurate. The problem is in this line of work; the expectation is 100%. “The more we rely on AI and the less we check it,” she noted, “the more vulnerable it makes us.”
What resonated across the room was that although AI does the work faster, someone still has to verify it, which eats back much of the time saved.
Clearwater’s own experience converting unstructured data reflected the same pattern. Reaching 80% accuracy took weeks, but closing the gap to 98% took substantially longer and required continuous prompt refinement. Compounding the challenge, underlying models keep changing, meaning a workflow tuned for accuracy today may behave differently tomorrow.
A related and more fundamental issue is data legibility at the source, particularly converting PDFs to Excel. AI can perform the downstream math flawlessly, but if it misreads a table during OCR, every calculation built on that misread number is wrong. As one panelist put it, the model’s math isn’t the risk its reading comprehension of messy source documents.
A consistent theme emerged throughout the event. AI is no longer a question of if but of when.
For practitioners, using AI in verification, data quality, compliance, and access control was harder and slower than the initial pilot ever suggested. The firms furthest along aren’t the ones that removed humans from the loop; they’re the ones that built the loop deliberately.
Contact us to learn more about how Clearwater approaches AI in financial data