AI for Private Equity Modeling

AI for private equity modeling has matured into a real product category. Endex, Eilla, Hebbia, BlueFlame and a handful of others have made the analyst's job of pulling structured financials out of a CIM and assembling a draft model dramatically faster. The productivity gain is honest. The category is real. And it is not the headline problem in private equity decision-making.

What AI for PE modeling actually does well

The state of the art handles three jobs at near-analyst quality: extraction of historical financials from a CIM or audited statement, scaffolding of a three-statement model with reasonable drivers and capital structure, and generation of a first-pass returns analysis with sensitivity tables. None of these is a small win — they collectively compress what used to be a week of analyst work into something closer to an afternoon.

The PwC 2024 AI productivity study put the range for AI gains in knowledge work at 35–85%. Modeling work sits squarely inside that range, and for early-stage diligence — where the deal team is trying to decide whether the asset is worth a deeper look — having a draft model in three hours instead of three days is a structural improvement to throughput.

The dominant blocker to extracting value from AI in private capital is not model capability — it is the absence of a unified data architecture that the model can operate on continuously, post-close.

Where the AI modeling category stops

The model file is a snapshot. It is the deal team's view of the world at the moment IC is asked to approve. Eighteen months later, the operator data has moved, the assumptions have shifted, the capital structure has been amended at least once — and the model file is sitting in a folder with the version date in the filename. The AI tool that produced it is not in the loop anymore. It was never designed to be.

This is the architectural choice the AI-for-modeling vendors made, and it is a defensible one for the job they set out to do. The model file is the unit of work. The deal team is the customer. The handoff is at IC. After IC, the AI vendor is gone and the model becomes a static artifact. That works for productivity. It does not work for decision validity, because decision validity is a property of the position over time — not of the model file at the moment of approval.

What needs to exist around the model

A typed, structured investment record at IC. The model's key assumptions — the EBITDA growth rate, the margin recovery curve, the working capital normalization, the exit multiple — get pulled out of the spreadsheet and bound to the IC decision as testable clauses. From the moment of approval, those clauses become the reference points against which operator data is tested.

A deterministic-first extraction kernel for the source documents, so that every figure in the model has a candidate ID and a provenance trail back to the page it came from. When the assumption gets re-tested in year three, the team can trace it back to the line in the CIM where it was sourced. Without that trail, the assumption is unanchored — and the AI's draft becomes the only reason it is in the model at all.

A continuous re-test loop that compares the live operator data against the model assumptions, quarter after quarter, for the life of the position. When the actual KPI drifts beyond the band the model assumed, the platform surfaces it — with the assumption that broke and the time-to-consequence ranking against the rest of the portfolio.

How Capital Refinery does this

• →Three-pass extraction kernel: regex sweepers identify candidate figures, deterministic constructors assemble the structured record, and a small local LLM adjudicates conflicts. Every number has a method, candidate ID, and source-page provenance.
• →Structured investment record at IC: thesis, assumptions, conditions, KPIs are bound at approval as typed objects the platform can re-test continuously.
• →Operator data binding: each accounting feed and KPI report is mapped to the model assumption it tests. The connection is a join, not a generation.
• →Decision validity scoring: when conditions change enough that the original IC decision is no longer defensible, the platform surfaces it — with the assumption that broke and how long the team has to act.

AI for modeling is a starting point. It compresses the analyst work and shortens the time to a draft. It does not produce a decision system, because the model file is not a decision system. The model is the input layer. The structured investment record is the decision layer. Capital Refinery is built around the second one.