Personal quarterly commentary for forty thousand households

The problem

Quarterly client reporting had settled into an unhappy compromise. The top few hundred relationships received genuinely personalized commentary written by their advisor. Everyone else received a templated market summary that said nothing about their actual portfolio — the same paragraphs about rates and equities that every other client received, next to numbers that were specific to them.

Clients notice. The most common question advisors fielded after a statement went out was some version of "what does this mean for me?" — which is precisely what the commentary was supposed to answer and didn't.

The hypothesis

The facts needed for personal commentary already existed: performance attribution, contribution and withdrawal effects, allocation changes, security-level contributors. What was missing was the writing, and writing at scale is exactly what generative models do well — provided they are never allowed to produce a number.

The build

• Deterministic attribution engine — returns, contribution and withdrawal effects, allocation and selection attribution, and top contributors and detractors, all computed in code per household.
• Grounded generation with injected figures — the model writes the narrative from a structured fact set. Every number is inserted from the attribution output; the model is never asked to calculate, infer, or estimate one.
• Compliance rulebook as an automated first pass — prohibited constructions (guarantees, forward-looking promises, performance predictions, unsuitable comparisons) detected with the specific rule cited, plus required disclosure checks per account type and jurisdiction.
• Human review queue with edit tracking — advisors review and sign every piece of commentary. We measure edit distance as the primary quality signal: how much a human changes the draft tells you more than any fluency score.

Rollout

We piloted on five hundred households across three advisors, with every draft reviewed and edit distance tracked per section. The first pass revealed the model over-explaining market context and under-explaining the client's own allocation changes — a prompt and fact-set problem, not a model problem. Compliance ran a full review of the rulebook checks against a labeled set of past violations before we scaled.

Results

Every household now receives commentary about their own portfolio rather than a market template. The quarterly cycle compressed from days of writing to hours of reviewing, and compliance review concentrated on flagged items instead of reading everything. Advisors kept their signature on every piece, which is both a regulatory position and the right one.

What we'd tell you

• Never let a language model compute a number that a client will read as fact.
• Encode your compliance rulebook explicitly and cite the rule on every flag — reviewers need to know which line to check.
• Edit distance is your real quality metric. Track it per section and you will find exactly what to fix.
• Keep the human signature. It is what makes the whole thing defensible.