AI Meeting Notes for Wealth Management: What Works and What Hallucinates

There is a specific moment, about ten seconds into reviewing an AI-generated meeting note, where the advisor's eyes narrow. They have just spotted a recommendation in the draft that was not actually made in the meeting - a 60/40 rebalance the client asked about, that the assistant has confidently transcribed as a recommendation the advisor agreed to. The advisor strikes the line, mutters "that is not what I said," and the trust-tax begins. Hallucination in wealth-management notes is not a quirky AI failure mode. It is the difference between a tool that gets adopted and one that gets uninstalled within a quarter.

Picture the screen most advisors are looking at when they review these drafts: a split view with the meeting transcript on the left and a note-draft sidebar on the right, each claim in the draft tied back to a citation tooltip pointing at the moment in the conversation it came from. That citation pattern is the single biggest difference between AI notes that work in this vertical and AI notes that do not. This piece walks through what we have seen work, what hallucinates, and the design choices that separate the two.

What works: structured extraction over free-form summarization

Generic AI meeting notes - the ones built for sales calls or product reviews - tend to summarize. They produce a tidy three-paragraph recap with action items at the bottom. That format is wrong for wealth management for two reasons. First, examiners read by field, not by paragraph; a Reg BI reviewer wants to find the recommendation field, the rationale field, and the conflicts field, in that order. Second, summarization is where hallucinations breed. The model fills in narrative connective tissue that the conversation did not contain.

Structured extraction is the alternative that has held up across our pilot. Instead of asking the model for a summary, the system prompts it to fill specific fields - investment profile changes, recommendations made, alternatives considered, conflicts disclosed, action items - and to leave any field blank if the conversation did not address it. Empty fields are a feature, not a bug. They tell the advisor what the meeting did not cover, which often surfaces a follow-up that should happen.

What works: per-claim citation back to the transcript

Every claim in the draft note should carry a citation pointing back to the moment in the transcript it came from. Per claim, not per section. The advisors in our pilot adopted this pattern faster than any other product feature, because it changed the review task from "is this draft correct" to "is this specific sentence supported by what was said." Citation tooltips also turn out to be the cheapest way to catch hallucinations. If a sentence has no citation, or the citation does not actually support the claim, the advisor strikes it.

What hallucinates: numbers, names, and certainty

Three categories of content drift consistently across LLM-generated wealth-management notes:

• Numbers. Dollar amounts, percentages, and account balances. The model is fluent in the shape of these numbers and will invent plausible ones when the conversation referenced them indirectly. The fix is to disable numeric inference and require explicit transcript anchoring for any quantity in the draft.
• Names. Specific fund tickers, share classes, and product names. "The Vanguard fund" in a transcript becomes a specific ticker in the draft. The fix is to keep the draft at the level of specificity the conversation reached.
• Certainty. The conversation said "thinking about," the draft says "decided to." The conversation said "if it makes sense," the draft says "will proceed with." Certainty inflation is the most subtle hallucination because each sentence reads as plausible.

The two-strike rule we use during pilots. If an advisor flags two hallucinations in the same note, the draft is not edited - it is regenerated with stricter extraction prompts. Editing around drift is how drift sneaks into the record.

What works: keeping the suitability field human-only

The Care Obligation under Reg BI and the fiduciary standard under the Advisers Act both ask whether the recommendation was in the client's best interest given the profile. That is a judgment, not an extraction. Within our pilot, the suitability rationale field stays empty in the AI draft and is filled in by the advisor as the last step of the review. This is the single design choice that has the most defensive value during examinations - the record cannot be challenged as "AI-generated suitability" because the suitability narrative is human-authored and timestamped separately.

What still needs work

Two areas where the current generation of AI meeting notes still falls short for wealth-management use. First, multi-speaker disambiguation when more than two people are in the room - couples meetings and family-meeting transcripts produce more attribution errors than single-client meetings. Second, jargon density. "Roth conversion ladder," "529 superfund," "step-up in basis," and the dozens of similar terms specific to U.S. wealth planning are well-handled by the better models but not uniformly. The pragmatic answer for now is a per-firm glossary the model is grounded against.

Productivity, with the asterisk

Across our early advisor cohort, mean documentation time fell from 52 minutes per meeting to roughly 8 minutes of editing per meeting after a four-week ramp. The asterisk is that the eight minutes of editing has to actually happen. A draft that goes into the CRM unedited is not a productivity gain; it is a compliance liability. The advisors who treat the AI draft as a starting position, not a finishing one, get the time back. The advisors who try to bypass review do not last in any pilot we have run.

Source notes

• SEC Regulation Best Interest Care Obligation guidance on suitability documentation.
• FINRA Notice 4530 supervisory record expectations for client-facing communications.
• NIST AI Risk Management Framework, Generative AI Profile, on hallucination mitigation patterns.
• Cerulli U.S. Advisor Metrics 2025 on advisor time allocation across documentation and client service.
• T3/Inside Information 2025 Advisor Software Survey on advisor adoption of AI note tooling.