An analyst copilot that reads everything and decides nothing
Filings, transcripts, and broker research made searchable with citations — so analysts spend their hours forming a view rather than locating the paragraph.
The problem
An investment team covered several hundred names across two sectors. Ahead of earnings season each analyst faced the same grind: read the filing, read the transcript, compare guidance language to the prior quarter, cross-reference broker notes, and reconcile it against their own model. The reading was necessary and the locating was not — a great deal of the week went to finding the paragraph rather than thinking about it.
Generic AI tools had been tried and abandoned. They summarized fluently, occasionally invented a figure, and gave no way to check. In a business where a wrong number reaches a client note, fluency without provenance is worse than useless.
The hypothesis
The bottleneck is retrieval and comparison, not judgment. A system grounded strictly in the firm's own document corpus, that cites every claim to a page and refuses when the evidence is not there, would compress the mechanical hours without touching the analytical work — which is the part the firm is paid for.
The build
- A governed corpus — filings, transcripts, licensed broker research, and internal notes, indexed with entitlements enforced at retrieval so an analyst only ever sees what their firm is licensed and permitted to see.
- Structure-aware retrieval — financial documents chunked so a table stays intact and a guidance statement stays with its qualifier, combined with hybrid semantic and keyword search over instrument and period metadata.
- Citation-mandatory answering — every sentence in an answer links to the document, page, and passage. Claims that cannot be grounded are not produced.
- Period-over-period comparison — a purpose-built path for the most common question: what changed in the language between this quarter and last, with both passages shown side by side rather than characterized.
- An explicit refusal to forecast — the system does not produce price targets, recommendations, or estimates. Asked for a view, it returns the evidence and declines the conclusion.
The line we drew, and would draw again: we do not build alpha models, and we told the client so in the first meeting. An AI system that generates investment conclusions is a research process nobody can defend to an investment committee or a regulator, and it removes exactly the work that justifies the fee. The analyst forms the view; the copilot makes sure they have read everything relevant before they do.
Rollout
We built the evaluation set from real analyst questions with known correct answers, and measured two things above all: citation accuracy — does the cited passage actually support the claim — and refusal behavior on questions the corpus could not answer. Analysts trialed it through one full earnings season before it was made standard, and their corrections became the eval suite that now gates every release.
Results
Analysts recovered several hours a week each, coverage breadth improved because preparation cost less, and the citation requirement produced an unexpected benefit: junior analysts learned faster, because every answer showed them where in a filing the answer lives.
What we'd tell you
- Citation accuracy is the metric, not summary quality. Verify that the cited passage supports the claim.
- Enforce entitlements at retrieval. Licensed research has terms, and they are not negotiable.
- Design refusal deliberately. "The corpus does not answer this" is the most valuable output the system produces.
- Do not automate the conclusion. It is the part your clients are paying a human for.
More from the field.
Consumer lending, decided in hours
Early-warning credit risk
The 6 a.m. advisor brief
Bring us a hypothesis. Leave with a system.
Tell us what's eating your team's time. We'll give you an honest read on whether AI is the right tool — and if it is, a scoped v1 with a timeline and cost.