AI for financial analysis is genuinely useful inside a regulated institution — but its real value is narrower and more specific than the marketing suggests. In practice it does four things well: variance analysis, anomaly detection, document extraction, and the analytical layer of reporting. None of those replace the controller. They shorten the distance between a question and a defensible answer.
Start with variance analysis and anomaly detection, because that is where the return is most immediate. A model can read a NAV file and flag the position that moved more than its peers, the fee accrual that no longer reconciles, the share class whose performance has drifted from its siblings. It does not decide that any of these is wrong; it surfaces the candidates a senior analyst would otherwise hunt for by eye. The work shifts from finding the exceptions to judging them.
Document extraction is the second use case, and the one most teams underestimate. Subscription documents, KYC packs, custodian statements, and counterparty contracts all carry structured information trapped in unstructured form. A model can pull the relevant fields into a reviewable table in seconds. The honest caveat is that extraction is probabilistic: it needs a confidence threshold and a human checkpoint above it, especially where the field feeds a regulated process. Used that way, it removes transcription, not responsibility.
The third use case is the analytical layer of reporting, and it is where we have the most concrete evidence. Across 39+ funds at a leading Zurich investment foundation, the heavy lifting of factsheet and performance reporting is now near-automated — and the role AI plays is specific. It drafts the commentary that explains a month's performance, it reconciles the exceptions a calculation throws off, and it flags the figures that warrant a second look. The core calculation stays deterministic and auditable. AI sits around it, not inside it.
This is also where the limits matter. AI for financial analysis does not interpret a FINMA expectation, it does not own a regulatory judgment, and it does not carry accountability for a number that goes to investors. The controller still signs. What changes is what the controller spends time on: less manual reconciliation and transcription, more of the judgment that actually requires a qualified person. A tool that blurs that line — that makes the sign-off feel automatic — is a liability, not an upgrade.
The pattern that works is unglamorous. Build the analytical layer inside your own environment, on the Microsoft stack you already license, with logging and review hooks a compliance officer can actually use. Start with one workflow where you can measure the baseline today, so the improvement is provable rather than asserted. Treat the model as augmentation for people who already understand the numbers — not as a substitute for understanding them. Done that way, AI for financial analysis stops being a slogan and becomes another reliable part of the operating stack.