Yann Magnan & Joe Zein — 73 Strings & Soal Labs | AI & Portfolio Monitoring in Private Equity

Private equity's messy portfolio data blocks AI monitoring

Private capital's portfolio monitoring problem is a data problem. Every company reports on its own template with its own vocabulary, and deal teams key the numbers by hand into systems that cannot audit or version them, so problems surface in the numbers long after they happen. The fix the guests describe is infrastructure. AI produces useful insight only from a data model that is structured, labeled and auditable.

Takeaways

• Portfolio data arrives fragmented and unlabeled. Portfolio companies report in different formats and cadences, and label the same metrics differently. One CFO calls it revenue, another net sales. As AUM flows into private capital, firms that do not fix the data model and process throw more people at the problem.
• AI fails without an auditable data model. Magnan says the industry needs 100% accuracy, not 99.5%, because the numbers reach LPs and the SEC. Portfolio monitoring is hard to build, and tools vibe-coded by associates fail because their builders cannot explain how they work.
• Governed models need rules and version history. Zein says a data model supports querying and writing, and needs auditability, version history, a record of who edited what, and rules on top. Excel allows none of it. Building one means reducing a firm's files to a canonical set and deduplicating entities.
• Implementation takes one to three quarters. Magnan says moving to an AI-native monitoring system is a change-management process. It most often takes one, two or three quarters, depending on the client's size, the portfolio and whether it spans multiple strategies and geographies.
• Clean data lets smaller models do more. With data structured, labeled and standardized, lighter open-source models such as Gemma can extract the first layer of intelligence at lower cost. Work that requires scouring the internet still needs a capable model, and 73 Strings uses its own machine learning where it beats an LLM.
• Buy a vendor tool before building. Zein says firms should buy rather than build, and that migrating off an existing monitoring system is hard in itself. Magnan says the first thing to think about is your data model.

Where portfolio data goes wrong

1. Portfolio companies send financials and data packs to the GP, each in its own format and cadence.
2. Portfolio teams forward the packs to deal teams, who key the numbers into a monitoring system or into Excel when the interface is poor.
3. Undocumented mapping rules, a forgotten zero or a mislabelled field put wrong numbers into the system.
4. Zein has seen the quarterly process take 45 days across a portfolio of 70 companies.