Automating Investment Memo Prep with AI Agents
Turning raw GP data rooms into validated, audit-ready investment analysis with a multi-agent workflow.
Key Results
Memo Prep Time
Cost per Run
Data Accuracy
1Context
Through Lomita AI, I partnered with Selby Lane Capital, a fund-of-funds investor whose diligence process centered on a standardized "Master Template" workbook: one row per portfolio investment across every fund a GP has raised, plus fund-level performance data and a battery of quantitative analyses.
Building that workbook was entirely manual. Analysts combed through raw GP data rooms—Excel files and PDFs with inconsistent layouts, multi-row headers, and operating metrics scattered across dozens of columns—then hand-built the analysis on top. It took 4+ days of skilled analyst time per GP, and a single silent error (like a row offset mapping every company to its neighbor's data) could undermine the whole memo.
2Problem
Investment memo prep was slow, error-prone, and impossible to fully trust.
Every GP data room had a different structure—no two extractions were the same
Manual data entry introduced silent errors that were hard to catch downstream
Analysts spent their time transcribing numbers instead of forming investment judgment
Verifying the final workbook against source files was as much work as building it
The team needed diligence-grade accuracy—every number traceable to a source document—at a fraction of the time cost.
3Strategy
I designed the system around a core principle: AI does the tedious work, adversarial AI checks it, and humans stay in the loop at the decision points.
Divide and specialize
Break the workflow into narrow, single-purpose agents (extract, validate, fix, analyze) instead of one monolithic prompt.
Trust through adversarial validation
Every extraction is independently re-checked against the source files by a separate QA agent, with fix loops until the data is clean.
Human review at checkpoints, not everywhere
The investor reviews the workbook at two defined milestones rather than babysitting the whole run.
4Execution
Standardizing the Data Model
Before any automation, we needed a rigorous target schema.
Defined a Master sheet schema: one row per investment with entry/exit dates, invested capital, proceeds, IRR, MOIC, and detailed operating metrics
Defined a Fund Performance schema: one row per fund vintage with net returns and DPI
Codified the team's data rules—value splits, currency conventions, edge-case handling—so every workbook comes out identical regardless of the GP's source format
Multi-Agent Extraction and Validation
The heart of the system is a three-phase workflow of ten specialized agents.
Extraction agents read the raw data room and build the base workbook
Adversarial validator agents independently re-read the source files and cross-check every number, producing a structured issues list
A fixer agent corrects discrepancies by going back to the source data, then the validator re-checks—looping until the data is clean
Human review checkpoints gate each phase, so the investor signs off before analysis is built on top
Automated Quantitative Analysis
Once the data is validated, the workflow generates the full analytical package automatically.
14 analysis modules produce 20+ workbook tabs: MOIC distributions, sector and region breakdowns, partner attribution, value creation bridges, and more
Fund scorecards benchmark performance against Cambridge Associates top-quartile data by strategy and vintage
Every tab is formatted in the firm's house style, ready to drop into a memo
A System That Improves Itself
Each run makes the next one better.
A run logger tracks every phase, validation error, and fix cycle for a complete audit trail
After each run, a post-run analyst agent categorizes all errors, identifies systemic patterns, and suggests concrete improvements to the workflow
This feedback loop turned one-off automation into a compounding asset