AI portfolio monitoring: retrieval is not substantiation
By Derwin Lucas ·
Summary
AI portfolio monitoring makes reported figures faster to find, rank, summarize and explain. Those are real gains. They do not establish that the figures being compared were measured alike, and a fluent answer can make an unsupported comparison more persuasive than a spreadsheet ever did. For a platform built by acquisition, the value of an AI layer is bounded by what its inputs carry: the definition behind each figure, the population it covers and the evidence supporting it. Where those are absent, the model has nothing to qualify its answer with.
What AI adds to portfolio monitoring
Language models and related tools change how investment and finance teams reach portfolio information. With AI search for private equity, a question asked in plain language replaces a search through board packs, monthly submissions and management presentations. Variance commentary can be drafted in a fraction of the time it once took. AI investment data analysis can flag unusual movements across many companies at once.
None of this is in dispute. The question this paper addresses is narrower: what happens when those capabilities are applied to figures whose comparability has not been established.
The answer inherits the input
A model asked to compare portfolio companies works from the figures it is given. If those figures share a label but not a definition, the comparison it returns is arithmetically correct and analytically mixed, in the same way a consolidated report is. The difference is in presentation. A report shows a table; a model gives an answer, often with a reason attached.
Exhibit 1. The same question, answered two ways (illustrative)
| Answer from reported figures | Answer with the basis attached | |
|---|---|---|
| Question | Which company has the highest gross margin? | Which company has the highest gross margin? |
| Answer | Company C, at 62.0%. | Company C reports 62.0%, stated before installation labor and sales commissions. On the portfolio basis, the same P&L gives 29.9%. The comparison is not supported as reported. |
| What the reader learns | A ranking. | A ranking cannot yet be made, and why. |
The first answer is not wrong about what was reported. It is wrong about what the reader will take it to mean.
Four ways an AI layer can mislead without making an error
- Ranking across definitions. A request for the best or worst performer produces an ordering. If close rates are built on demonstrations at some companies and appointments issued at others, the ordering reflects denominators. In one illustrative month, the same records produce close rates of 33%, 27%, 20% and 13%. See the arithmetic →
- Explaining a definitional gap as an operating one. Asked why one company’s margin exceeds another’s, a model can supply a plausible operating reason: pricing, crew productivity, material costs. When the actual cause is where a cost was booked, the explanation is fluent and unfounded.
- Presenting a partial total as complete. Asked for portfolio revenue, a model sums what it has. A total built from 22 of 25 companies is returned as the portfolio’s unless the coverage travels with the figure. See the reporting coverage case →
- Detaching the figure from its basis. An answer quoted in an email or a board pack loses the period, population and definition it was drawn from. The number travels; the qualifications do not.
In each case, the model has done what it was asked. The exposure lies in what it was not given.
The questions most exposed
The questions finance teams most readily put to an AI layer are the ones most sensitive to definition. Each has a condition that must hold before the answer can be relied on.
Exhibit 2. Common questions and what each depends on (illustrative)
| Question asked | What the answer depends on |
|---|---|
| Which company is most profitable? | Every company carrying shared-service costs on one basis. In an illustrative 25-company portfolio, eight companies that leave out about $150,000 each report $1.2 million more EBITDA between them than they would on that basis. |
| Are we on plan year to date? | Like months compared with like months, and any company that has not reported named rather than filled in. |
| What drove the change in margin? | The same cost classification in both periods. A reclassification produces a movement with no operating cause. |
| Which add-backs are at risk? | Each adjustment reviewed against documented criteria. At 8.0×, $600,000 of recurring costs presented as adjustments puts $4.8 million of implied enterprise value at issue. |
A model can answer each question in seconds. Whether the answer holds depends on conditions the question does not mention.
What an AI layer needs from its inputs
An AI layer can only qualify an answer with information that exists alongside the figure. For a comparison across acquired companies, that means each figure carries:
- the definition applied, and where it departs from the portfolio’s;
- the companies, periods and records it covers, and what it excludes;
- whether it traces to source records, or rests on a submitted total;
- a ruling on whether a given comparison is supported, supported with a stated limitation, or not supported.
With those attached, the second answer in Exhibit 1 becomes possible, because the qualification is in the data rather than in an analyst’s memory. Without them, no prompt can produce it reliably.
Where AI is most useful, and where judgment remains
AI is well suited to locating documents, drafting commentary from governed figures, and flagging movements that warrant review. A sudden step change in one company’s close rate is worth surfacing; it may indicate a change in how the rate is calculated rather than in how the company sells. KPI tracking for private equity → Definition drift in portfolio reporting →
Deciding whether that step change is definitional, whether an adjustment qualifies, or whether two companies’ figures may be compared remains a matter of evidence and approval. Those decisions belong to the people accountable for the figures, and the record of who made them is part of what makes a figure defensible.
What this paper does not claim
- This is not an argument against AI in portfolio monitoring. The gains in speed and access are real, and they compound once the inputs are governed.
- The problem predates AI. Dashboards and consolidated reports carry the same exposure. AI raises the stakes because its answers are faster, more conversational and more readily repeated.
- Not every question needs this rigor. Locating a document or summarizing a board discussion does not. Comparisons that move capital, test a covenant or support a valuation do.
Implication
AI makes the comparison question more urgent, not less. A platform that adds an AI layer to unexamined figures gains speed in reaching conclusions the evidence may not support. A platform that first establishes what each figure measures gains an AI layer whose answers can carry their own qualifications. The second is achieved one consequential measure at a time, beginning with the figures leadership already relies on.
In Goverisyn, every figure carries an evidence state: Substantiated, Qualified, Limited or Not Available. What each state means → For the broader distinction between monitoring and comparability, see Private equity portfolio monitoring and the comparability question.
Goverisyn’s Portfolio Measurement Review establishes this for the measures leadership relies on most. The first step is a free 60-minute Portfolio Review, which needs no data in advance.
Start with a Portfolio Review.
60 minutes. No data needed in advance.
By Derwin Lucas, founder of Goverisyn, with nearly three decades in management and information systems in home services. Published . About Goverisyn →