We built a system that reads incoming deal memos, extracts key metrics, scores them against firm-specific criteria, and surfaces the top candidates. Here is how it works.
The problem
A mid-market PE firm receives between 30 and 50 deal memos per week. Each memo is 15 to 40 pages. An associate spends roughly 45 minutes per memo on initial screening: reading the executive summary, pulling out key financials, checking sector fit, and writing a one-paragraph summary for the partners.
That is 25 to 35 hours per week on a task that has a 90% rejection rate. Most memos do not fit the firm's investment criteria, and an experienced associate can tell within the first two pages.
The system
The pipeline has four stages. First, document ingestion. We parse PDFs, extract text, and normalize formatting. This handles the fact that every bank and broker uses a different template.
Second, structured extraction. The LLM pulls out specific fields: revenue, EBITDA, growth rate, sector, geography, deal size, management team background, and key risks. Each field has a confidence score.
Third, scoring. The extracted fields are scored against the firm's criteria. This is not AI. It is a simple rules engine that the partners can modify without engineering support. Revenue below $10M? Reject. EBITDA margins below 15%? Flag for review.
Fourth, presentation. The top candidates get a one-page summary with the extracted metrics, the score, and a link to the full memo. Partners review 5 to 8 memos per week instead of 40.
What we learned
The hardest part was not the AI. It was getting the scoring criteria out of the partners' heads. They had years of pattern matching that they could not articulate as rules. We spent two weeks sitting in deal review meetings before we could build the scoring engine.
The second hardest part was trust. The team needed to verify the system against their own judgments for six weeks before they stopped double-checking every recommendation.