If your AI can't sign an export, it can't defend the IC decision.
Generic AI tools can summarize and brainstorm. Private markets work has to survive the room — IC memos that hold up under scrutiny, figures that trace to source cells, decisions that stay defensible quarters after close. Chatbots do not produce that.
If your “AI transformation” is a chat window, you are still rebuilding models, rekeying numbers, and defending conclusions with screenshots. The polished output makes incomplete diligence look complete. That is not intelligence. That is faster confusion.
The five structural failures (and why prompts won’t fix them)
These are not prompt-engineering problems. They are structural problems with treating private markets work as ask-and-answer instead of as a governed lifecycle: underwriting → IC → portfolio → exit.
What goes wrong in production — and what fixes it.
| Capability | Failure mode · why it happens | Capital Refinery |
|---|---|---|
| Inconsistent KPI definitions | Same metric, different meaning across deals. Chat tools do not enforce schema or mapping rules. | Unified KPI model, lane-aware extraction, candidate trail to source cell |
| Evidence theater | Claims without traceable source support. LLMs generate language; they do not carry provenance. | Every figure clicks back to the source cell. Cryptographic fingerprint on every export. |
| The AI's role is invisible | You cannot tell which numbers a model wrote. Everything looks equally confident. | A method column on every exported number: deterministic_parse, adjudicated, operator_edit, or declared. The LLM's fingerprints are labeled, per figure. |
| Guessing instead of refusing | Ambiguity in, confident answer out. A password-protected file or a cross-document conflict becomes a hallucinated value. | The pipeline refuses: locked files and conflicting documents become review items a human resolves — pick a value, unlock with the password, or upload a corrected source. Audited. |
| Fragile repeatability | Every IC memo is a rewrite. No reusable structure; everything is one-off text. | Lane-aware memo dossier — same shape every deal, every quarter. |
| Governance gaps | No human approval checkpoint or accountability. Generic AI assumes single user, not regulated workflow. | IC anchor on approval. Decision basis locked. Override audit trail. |
| Lifecycle disconnect | Diligence output dies after close. Portfolio monitoring starts from scratch. | “Since IC” is the unit of work — every operator update tested against the original thesis. |
What survives the room
Private markets work is not “ask and answer.” It is extract → normalize → grade against firm policy → lock at IC → track drift → produce a defensible record when the LP asks why two years out.
What to require of any AI in your workflow
- Every figure has to click back to the source cell. No LLM-generated narrative pretending to be analysis.
- Every export has to carry an evidence fingerprint the LP can verify independently.
- Every IC approval has to anchor the decision basis. The memo two years from now has to render the same numbers it rendered the day of approval.
- Every operator update has to be tested against the original underwriting. The team finds out when the thesis breaks — with the time still left to act.
- Every override has to be auditable. Who changed what, when, and on what basis.
If the AI in your stack does not do those five things, it is an accelerator, not a decision system.
Want proof on your own documents?
Bring a real deal pack — CIM, credit agreement, financial model. Same-day diagnostic pack back, source-backed, graded against firm policy. No demo data.