87%
reduction in preliminary diligence time
per-target preliminary diligence, roughly three weeks to 48 hoursper target, steady statepreliminary screen; confirmatory diligence and committee judgement unchanged
THE SITUATION
Why this work mattered
For each potential acquisition, a team of three analysts spent about three weeks pulling financials from SEC filings, scanning news archives, searching court records for litigation history, and building competitive maps. Industry norms put preliminary diligence at two to four weeks, so a three-week manual cycle was in band and the binding constraint was capacity, not method: the firm screened 40-plus targets a year but could only deep-dive about 15. Analysts spent roughly 60% of their time gathering data and only 40% on the judgement work. The commercial stake was deal throughput: in competitive processes the team that builds conviction first wins, and preliminary-screen capacity was the limiter on how many targets the firm could pursue with a fixed team.
THE FAILURE MODE
What was breaking before us
The firm had tried to scale by adding analyst hours and by using generic research tools, and both stalled. More hours did not change the per-target cycle length; generic tools produced unsourced summaries an investment committee could not act on. An IC will not move on a brief it cannot trace to a filing or that does not flag its own uncertainty, and a confident, unsourced brief is a liability rather than an asset. Those attempts failed for the same reason the legal-AI tools failed in our contract-review work: speed without traceable grounding is not committee-grade, it just produces wrong answers faster.
THE BUILD
What we built
We mapped the analysts' research workflow in a single day and identified 12 distinct data-gathering tasks that followed repeatable patterns. The system is a multi-source research pipeline that pulls from SEC filings, financial databases, news APIs, court-record systems, and competitive-intelligence platforms. Gemini 3.5 Flash runs the broad sweeps, scanning hundreds of documents and flagging relevant sections; Claude Opus 4.8 then performs the structured analysis, extracting financial metrics, identifying risk factors, and writing the narrative sections. The output is a 40 to 60 page brief in the firm's existing template, with a source citation and a confidence score on every line-item finding.
HOW IT WORKS
How it actually works
The pipeline fans out across the source systems, retrieves into a Pinecone-indexed working set, and runs a two-stage model pass (Gemini 3.5 Flash for breadth, Claude Opus 4.8 for structured extraction and narrative) over a FastAPI service backed by PostgreSQL on AWS. Every finding carries the source it came from and a confidence score, which is what makes the brief committee-grade rather than merely fast. The human boundary is firm: analysts review and refine a draft, they do not approve a black box, and the citation-and-confidence layer exists precisely so a person can audit any line in minutes instead of redoing the research.
The system carries the volume. A person carries every judgement call.
THE OUTCOMES
The outcomes that held
Every number below carries its denominator, window, and scope. No claim a buyer with a calculator can break.
48h
target selection to completed brief
one targetfrom target selection to delivered draft briefstructured brief generation; analyst review and refinement follow
1 week
kickoff to production
single engagementkickoff to productionmulti-source pipeline build to the firm's existing brief template
40+
sources cross-referenced per target
distinct data sources per target briefper briefSEC filings, financial databases, news, court records, competitive intelligence; every line-item finding source-cited and confidence-scored
SECOND-ORDER EFFECTS
The firm moved from deep-diving about 15 targets a year toward evaluating closer to its full 40-plus pipeline with the same three-analyst team, because the binding constraint was preliminary-screen capacity and that is what the pipeline relieved. Analysts shifted from data gathering to the judgement work that actually differentiates a deal call. The per-finding citation and confidence layer changed the review itself: the committee debates the analysis instead of re-verifying the inputs.
This changed how we evaluate deals. We look at roughly twice as many targets now with the same team, and every line in the brief points back to a source we can check.
Managing DirectorMid-market private equity firm, roughly $2B AUM
RELATED WORK
More of this work
The same shared system, applied to four other regulated and high-volume problems.
Tell us the problem. We'll scope the path.
Tell us the problem, the constraint, and what success looks like. We'll tell you whether there's a credible path to production.