Insights · Research
AI agents cannot govern portfolios without the data layer beneath them
AI is succeeding in private equity wherever the task is bounded and the data is available. It is failing at cross-portfolio governance, because the connected record that evaluation needs does not exist in most firms' current stack.
AI agents in private equity work well on bounded tasks such as deal screening, document extraction and templated reporting, because the data is available and a human checks the output before it becomes a decision. They fail at cross-portfolio governance, judging whether a prior investment decision still holds, because no system connects the original decision, its evidence and the live operating data. That connection is a data layer, not a better model.
Where AI is working in private equity
AI is succeeding across private equity wherever the task is bounded and the data is available. The industry is buying accordingly: Accenture found that AI and machine-learning deal value in private equity more than tripled between 2023 and 2024, rising from $41.7 billion to $140.5 billion.
Deal sourcing and screening works because the inputs are external and structured: market data, company financials, sector trends, relationship graphs. The AI narrows a broad universe and a human evaluates the shortlist, so the boundary between what the AI does and what the human decides stays clear.
Document extraction in diligence works for the same reason. The task is well defined: read a data room full of PDFs, extract structured information, and surface it for review. The AI reads faster than a person; the person still validates and decides. LP reporting and management updates work because the output is template driven: take portfolio data, apply a reporting format, produce a draft.
| Task | Why AI succeeds here |
|---|---|
| Deal sourcing and screening | Structured, external inputs; a human reviews the shortlist |
| Document extraction in diligence | Bounded task; output is validated before it is used |
| LP reporting and updates | Template-driven; the format is fixed, the data changes |
| AI as IC challenger | A human committee still decides; AI adds structured dissent |
Some firms have gone further and added AI platforms as non-voting investment committee members that analyze deals and market data and challenge groupthink. That is a meaningful use of AI: structured dissent inside a process historically vulnerable to consensus bias. The pattern across each of these: AI succeeds when the inputs are available, the task boundary is clear, and a human evaluates the output before it becomes a decision.
Where AI is failing: cross-portfolio governance
AI is failing wherever the task requires connected context that does not live in any single system. Post-close portfolio monitoring is the clearest case. AI can scan news, flag events and summarize financial trends, but it cannot connect those events to the specific thesis, covenants or decision history of a given investment. The constraint is structural: the usefulness of a monitoring agent depends directly on the quality of the context it can reach, and today that context is scattered.
Cross-portfolio governance is the harder problem still. Very few systems can look across a portfolio and say which holdings have conditions that no longer support their original committee approval, what specific assumption broke in each case, how much time remains before the gap becomes material, and what each deal team must do to restore the basis.
- The original IC approval, with the specific assumptions that justified the decision
- Current operating data connected to those assumptions, not just to KPI thresholds
- The credit agreement or investment terms, read as structured data rather than a PDF
- The history of prior decisions and evaluations: what was flagged, what was decided, what happened
- A time-to-consequence estimate based on agreement terms, trajectory and deadline structure
None of those inputs sit in one system in the current private equity stack. The IC memo is on a shared drive. The KPIs are in a monitoring tool. The credit agreement is a PDF. The decision history is in email threads and meeting minutes. Without that connected record, an AI agent still depends on a person to remember the thesis, find the agreement, reconstruct the context and decide what matters. The agent makes the work faster. It does not change who carries the judgment.
The data architecture problem beneath the AI layer
The root cause sits beneath the model. Without a unified, governed data architecture, AI stays a surface enhancement, and private equity firms are starting to treat data engineering, once viewed as plumbing, as strategic infrastructure in its own right.
As AI improves, the need for clean, connected data goes up, not down. An AI agent cannot reason well over files that are a mess.
The inefficiency is not in the AI layer. It is in the data layer beneath it: manual data preparation, inconsistent formats, and reporting processes that were not built to connect a decision to the evidence behind it. The industry is spending on AI capability that sits on top of fragmented, disconnected infrastructure. The AI is fast. The data it can reach is incomplete.
What the data layer actually needs to be
The precondition for useful AI agents in portfolio governance is not a better model. It is a connected record.
The decision
What was approved, by whom, under what assumptions, with what risk accepted.
The evidence
The diligence findings, financial model, credit agreement and management representations behind it.
The ongoing state
Current KPIs, covenant compliance and operating metrics, connected to the assumptions they test.
The evaluation history
Each later assessment of whether the basis still holds, and what was decided in response.
The agreement structure
Covenant definitions, compliance thresholds and deadlines, read as structured data, not stored as PDFs.
Without the layer
- An agent reasons over files and a dashboard, not the decision
- The thesis lives in a deck and in people's memory
- Time to consequence is a guess, not a calculation
- Analysis is fast; reconnecting it to the decision is manual
With the layer
- An agent evaluates a position against its own decision history
- The thesis, its assumptions and its conditions are structured data
- Time to consequence is estimated from the agreement's own terms
- The record, not a person's memory, carries the connection forward
That is not a data warehouse. It is a decision record: a structured layer that connects what was decided to why it was decided to whether it is still defensible.
The agentic future depends on the present architecture
Where this is heading is clear enough. Industry research describes AI agents that will scan markets, model scenarios, flag diligence risks as a deal moves, and keep supporting the deal team after signing, with diligence shifting from a static snapshot to a living model. Vista Equity Partners built what it calls an agentic factory to build AI elements for its portfolio companies, and at one Apollo portfolio company, AI deployment cut costs by 40% in some content production work, 15% to 20% through automated lead generation, 15% in customer care, and 10% to 15% in software development.
Those are real operational gains inside individual portfolio companies. The next frontier, agents that operate across the portfolio at the decision layer, evaluating thesis validity, surfacing time to consequence, and spotting cross-portfolio patterns, needs a different foundation: not a better model deployed company by company, but a connected decision record that gives the model the structured context it needs to reach a conclusion rather than produce more analysis.
A firm's own decision history, connected to live evidence and agreement terms, is proprietary data that a general-purpose model cannot replicate. That is the moat.
What Capital Refinery connects today
Capital Refinery does not run cross-portfolio AI governance. What it builds today is a piece of the record that kind of evaluation would need. It classifies the data room against a taxonomy of 41 document families, from financial statements and bank statements to the quality-of-earnings add-back schedule, the credit agreement and the compliance certificate, and spreads the financials with figures traced to source where a locator resolves.
- Ten advisory reconciliation checks run on the spread financial statements; they flag inconsistencies rather than gate the analysis
- The team records a position on each add-back in the EBITDA bridge, accepted, rejected or negotiated
- Five house stress scenarios, from a 300bp rate rise to a combined 25% downside, show the covenants under pressure, with the first-breach quarter modeled
- Risk signals are graded against the firm's own configured policy thresholds
- Memo sections for the financials, the risks and the terms are assembled from that sourced evidence; the narrative stays with the deal team
- Open conditions and covenant obligations sit on the same deal record that monitoring reads from

Where an LLM is used, it adjudicates among candidates extracted from the documents. It does not generate the figures.
Questions
Why can't AI agents just monitor the portfolio directly?
Monitoring agents can scan news and KPI feeds, but they cannot connect what they find to a specific thesis, covenant or decision history unless that record already exists in structured form. Without it, the agent produces analysis, not a governance conclusion.
What is a data layer in this context?
It is a structured record connecting the original decision, its evidence, the live operating data that tests it, and the history of later evaluations. It is a decision record, not a data warehouse.
Where is AI already working well in private equity?
In bounded tasks with available, structured data: deal sourcing and screening, document extraction in diligence, and template-driven LP reporting. A human still evaluates the output before it becomes a decision.
Does Capital Refinery run AI agents across a portfolio?
No. Capital Refinery spreads financials traced to source, runs advisory reconciliation checks, grades risk against the firm's own thresholds, and assembles memo sections from that evidence. That record is a foundation a future agent would need, not a cross-portfolio agent itself.
What is time to consequence?
It is an estimate, discussed across the industry, of how long a portfolio company has before a broken assumption becomes irreversible. It depends on reading the agreement as structured data rather than a PDF.