Insights · Guide
AI for portfolio monitoring in private equity: why dashboards aren't the missing layer
The pitch for AI portfolio monitoring is appealing: feed operator updates into a model and have it surface what matters. The pitch runs into the data architecture underneath it.
AI portfolio monitoring today is either an LLM bolted onto an existing monitoring platform, or a standalone layer that digests operator PDFs into a quarterly summary. Both save an analyst reading time. Neither tests this quarter's KPI against the entry case, because a typical portfolio company's data is split across five surfaces that do not share a schema. What closes the gap is reading each reporting pack to its covenant terms, with the figures cited and the headroom shown.
The state of AI in monitoring today
Two patterns dominate. The first is an LLM bolted onto an existing monitoring platform, such as iLEVEL, Allvue or Chronograph, that summarizes board packs and answers questions across the stored documents. The second is a standalone AI layer that ingests PDFs and emails from operators and produces a quarterly digest. Both are useful at the level of saving an analyst reading time per portfolio company per quarter.
The PwC 2024 AI productivity study put the range of AI productivity gains in knowledge work at 35-85%. Reading and synthesizing operator updates sits inside that range, so the productivity case is honest. The decision case is a separate question.
| Pattern | Example | Good for | Where it stops |
|---|---|---|---|
| LLM bolted onto a monitoring platform | iLEVEL, Allvue, Chronograph | Summarizing board packs; answering questions across stored documents. | Testing a KPI against the entry case the IC approved. |
| Standalone AI digest layer | Ingests operator PDFs and emails | A quarterly digest without a monitoring platform. | Same gap: no connection to the entry assumptions. |
The fragmented context problem
Inside a typical mid-market PE firm, a single portfolio company's quarterly state is split across at least five surfaces: the operator board pack (PDF), the management KPI sheet (spreadsheet), the accounting export (a CSV from the general ledger), the credit agreement's covenant schedule (a PDF in the data room), and the IC memo from entry (a document on a shared drive). None of these share a schema, and none of them know about the others.
When an LLM is asked to monitor the company, it operates across that fragmented context. It can summarize and extract figures. It cannot test the figures against the entry case, because the entry case is not in a form the model can test against; the IC memo is prose, not a typed record.
What a usable monitoring layer needs
The report arrives
Board pack, KPI sheet or accounting export lands.
Figures are spread and cited
Figures are traced to source, cited to the page where a locator resolves.
The figures are tested against the covenant terms on file
Not against a generic benchmark, against the terms the credit agreement sets.
Headroom and drift are shown on the next render
The gap between the current figure and the covenant line, and how it has moved.
Why the dominant vendors do not ship this
The dominant monitoring platforms were built around the document and the dashboard as the primary objects. Their data model has a node for the portfolio company, a node for the reporting period and a node for the KPI, but no node for the IC decision the firm made on the asset at entry. Adding that node after the fact means rebuilding the schema.
How Capital Refinery monitors the position
Capital Refinery does not replace the monitoring dashboard. It reads the reporting packs and statements the team already receives.
- Reads each reporting pack to its covenant terms, with headroom and drift shown on the next render.
- Spreads the figures behind the covenant test, traced to source and cited where a locator resolves.
- Grades risk signals against the firm's own policy thresholds; move a threshold and the memo re-grades on the next render.
- Exports the monitoring memo as DOCX, with its evidence fingerprint.
Numbers are read by deterministic parsers. Where candidates conflict, a local model chooses among the named candidates; it does not generate the figures. Some written outputs, such as narrative summaries, are drafted by a hosted model from Anthropic, which receives deal details and summary figures.

See the monitoring layer on a position you actually own
Bring us a position from your book. We will parse the entry pack and the most recent operator update and show the covenant test behind it, with headroom and drift.
The model is not the headline value in AI portfolio monitoring. Reading the report to the terms the firm already agreed to is.
Questions
What is the state of AI in portfolio monitoring today?
Mostly an LLM bolted onto an existing monitoring platform, or a standalone digest layer. Both summarize operator updates well. Neither tests a KPI against the entry case the IC approved, because that case is not stored in a testable form.
Why is a portfolio company's data fragmented across so many files?
A typical company's quarterly state spans the board pack, the KPI sheet, the accounting export, the covenant schedule and the IC memo from entry, usually in five different formats that do not share a schema.
What does Capital Refinery actually monitor?
Each reporting pack and statement a portfolio company sends, read to the covenant terms on file, with headroom and drift shown on the next render.
Does Capital Refinery replace my monitoring dashboard?
No. Capital Refinery does not integrate with the monitoring platform or the fund admin system. It reads the reporting packs and statements the team already receives.
Does AI generate the risk grade in Capital Refinery?
No. Risk grading runs against the firm's own policy thresholds. In the extraction step, a local model only chooses among deterministic candidates; it does not generate figures.