AI for the Investment Committee in Private Equity

The investment committee is the highest-leverage decision surface in any private equity firm. It is also the surface where AI is being applied most superficially — and the gap between "AI in the IC" as a marketing phrase and as an actual workflow has gotten wide enough to be its own problem.

What "AI for the IC" usually means today

Walk into any AI-for-PE pitch and the demo is almost always the same. A chat window over the data room. A summarization of the CIM. A draft IC memo generated from a document corpus. An assistant that can answer "what is the EBITDA growth rate in the model." The vendor calls this "AI in the investment committee" — and it is a useful productivity layer for the work that precedes the IC meeting.

But the IC meeting itself does not need a chat window. It needs a decision. And the work that comes after the IC meeting — the eighteen months of monitoring the position, the twelve quarterly reviews, the three operator-update cycles where the team has to decide whether the original case is still working — has nothing to do with summarization.

With $3.6 trillion of unrealized value sitting on GP balance sheets and extended exit windows, the cost of getting the post-IC decision wrong is structurally higher than it has ever been.

What the IC actually needs from AI

The IC needs three things, and they are not the things the dominant vendors are shipping:

A structured decision record at approval.Not a draft of the memo — the memo as a typed object. The thesis as a structured field. The assumptions as separate, individually-testable rows. The conditions the team relied on as logical clauses the platform can evaluate against operator data eighteen months from now. None of this is what a chat window over the data room produces.

A live binding from operator data to the assumption it tests. The KPI on this quarter's board pack is interesting only because of what it tests about the entry case. AI for the IC has to make that connection — and the connection is structural, not generative. It is a join, not a summary.

A decision validity score that moves with the data. When conditions change enough that the original IC decision is no longer defensible, the IC has to find out — with the time still left to act. That is the headline value of AI in the IC, and it is the one capability nobody is shipping.

Why the dominant vendors do not ship this

The reason is structural, not strategic. Building a structured decision record requires three things the dominant AI-for-PE vendors do not have. First, a deterministic-first extraction kernel — most are LLM-first, which means they cannot guarantee provenance. Second, a domain data model with the IC decision as the primary object — most are built around the document as the primary object. Third, a continuous operator data binding loop — most operate inside the diligence window and walk away at close.

None of those are easy retrofits. They require designing the system around the decision instead of around the document, and that is the architectural choice no incumbent has made.

What changes when the IC has the layer

• →Every IC approval lands in a structured record with thesis, assumptions, and conditions — not a slide deck headed for a shared drive.
• →Every quarterly review is informed by which assumptions are still holding and which are not — without rebuilding the model from scratch.
• →The IC sees, on every position, how much runway is left before a structural break — and reorders attention accordingly.
• →Institutional memory survives team turnover. The next principal inherits the structured reasoning, not a folder of PowerPoint files.
• →LP scrutiny on decision discipline gets answered with data, not narrative — the same structured record exports cleanly into LP reporting.

Capital Refinery is built around exactly this premise. It is not a chat window over the data room. It is a structured record of every IC decision the firm has made, alive across the life of the position, and continuously re-tested against the operator data the firm is already collecting. The IC does not need a faster memo drafter. It needs a memo that does not go stale.