The wealth management industry is adopting artificial intelligence—carefully, and mostly at the edges. As Michael Kitces has described, the mainstream advisor remains something of a “collective dabbler,” adopting focused AI tools for specific tasks without yet integrating them into the core work of advice. The clearest example is meeting documentation, one of the fastest-growing categories in advisor technology. with roughly 18% of advisory teams now using AI meeting-note tools. Useful. Peripheral. And, for now, safe — because the output can be reviewed before it becomes part of the client record.
The trajectory, however, points somewhere else. The real promise of AI in wealth management lies at the intelligence layer: portfolio analysis, risk assessment, scenario modelling, tax-aware planning, and identifying inconsistencies between what clients say and what their financial lives reveal. These are no longer administrative tasks. They influence judgement. That is where the value is. It is also where the supervision problem begins.
Advisory compliance was built to oversee communication. The supervisory architecture at most firms — archiving, surveillance, review — assumes the thing being supervised is a message: an email, a chat, a piece of marketing. That architecture works when AI is drafting the email. It fails the moment AI is shaping the advice, because a recommendation produced with a model is not a message to be archived. It is a decision to be explained. Most firms have no infrastructure that can explain it.
A compliance officer must be able to answer several straightforward questions: What was recommended? To whom? On what information was that recommendation based? And why was that conclusion reached? When a human advisor makes that call, the reasoning lives in notes and in the advisor’s account of their judgment. When a model contributes to it, the reasoning has to be captured at the moment of inference — or it is gone. You cannot reliably explain in July why a system assigned greater weight to one risk factor than another in March if the relevant inputs, prompts, outputs, and review process were never recorded. That is the blind spot: not that AI makes mistakes, but that the industry is deploying AI into supervisory systems that were never designed to capture how those decisions were made.
Compliance officers are already recognising the challenge. In the 2025 Investment Management Compliance Testing Survey, AI use was the single most-cited concern in the industry. The unease is well founded — and it compounds because of how these tools enter a firm. They arrive feature by feature, each modest on its own: a smarter search here, an automated summary there, a risk flag that seems helpful. No single addition triggers a compliance review. But cumulatively, the locus of judgment migrates from people the firm can supervise toward systems it cannot yet see. The regulatory questions — is this suitable, is this in the client’s interest, can we defend this recommendation — do not change. The ability to answer them quietly erodes.
None of this is an argument against AI at the intelligence layer. In fact, it is where the profession should be heading. It is an argument that observability has to be designed in, not bolted on. Three capabilities separate a defensible deployment from a hopeful one. Auditability: every AI-influenced recommendation should leave a durable, timestamped record of inputs and outputs. Explainability: the reasoning must be reconstructable in terms a compliance officer, and ultimately a regulator, can follow. Supervisability: the firm must be able to set boundaries on what the system may do, and prove those boundaries held. A tool that offers none of these is not a productivity gain. It is an unpriced regulatory risk.
There is a reason the firms furthest along with AI tend to say the same thing: they considered compliance at the beginning of the process, not the end. That instinct mirrors how other high-stakes fields adopted decision-support AI — in medicine, in aviation, in credit, the question was never merely “does the model help?” but “can we account for what it did?” Wealth management, which holds a fiduciary duty at its center, has every reason to hold the same standard. The obligation to act in a client’s interest presupposes the ability to show that you did.
The firms that navigate this well will not be the ones adopting AI fastest, nor the ones avoiding it out of caution. They will be the ones treating supervision as a first-order design requirement — asking of every AI capability not just what it can do, but what it leaves behind for the record. AI at the inbox created no new duties. AI at the intelligence layer creates several. An industry can adopt a technology quickly, or it can adopt it blindly. The distance between the two is observability — and it is worth closing before the recommendations, not after the examination.
