Sit in enough quarterly boards and you will eventually watch this scene. The CEO of a portfolio company presents the quarter’s AI progress – a working tool, a faster process, numbers he is proud of – and the partner chairing the board nods, thanks him, and later asks someone privately: “Is that real?”
The industry now has a number for that scene. In AlixPartners’ most recent private equity leadership survey – one of the few whose sample reaches well below the mega-funds – portfolio-company leaders were more than twice as likely as their investors to say they are satisfied with the results of AI. Same companies, same results – more than twice as likely to call them satisfying.
Satisfaction is a blunt instrument – it measures sentiment, not P&L – but a gap that size, about the same facts, is the point.
The tempting explanation is the cynical one: the companies are inflating, and the sponsors have caught them at it. I want to argue something different – that almost nobody in this picture is lying, that both sides are reasoning correctly from where they sit, and that the gap persists because the structure gives neither side a way to produce evidence the other can believe. That last part matters, because it means the gap is closeable. Until it closes, it costs real money in both chairs.
The view from the board seat
Start with the sponsor’s skepticism, since it is the better documented. Skepticism is close to a professional credential in private equity – the quality-of-earnings review exists because seller-framed numbers are assumed dressed until proven otherwise, and the person who takes an add-back at face value is the one who ends up looking naive.
So a partner who suspects a company’s reported AI wins are running ahead of its measured ones is not being paranoid. She is noticing three structural facts about how those wins reach her.
First, the evidence chain is short and interested. The success metrics for most pilots were written by the vendor that sold the pilot.
A vendor’s post-pilot report concluding that the pilot failed is a genuinely rare document.
Second, the measurement was designed after the fact. Almost no operating company baselines a workflow the week before it changes the workflow. The team had a day job; the pilot ran on top of it; by the time anyone asks what the change saved, the before-state exists only in memory. That is not chicanery. It is what measurement looks like at a company that never had a reason to instrument itself for this question – the missing number indicts nobody. It is simply missing.
Third, the board seat is too far from the work. AI results live in the weekly texture of how things get done; what reaches the board deck is a summary of a summary, quarterly. A good chair can ask sharper questions, but interrogating a summary harder does not produce an independent read of the thing summarized.
So the sponsor discounts. Quietly, and usually without saying so.
What the discount does to the company
Now sit in the other chair, because the CEO knows exactly what is happening.
He presented progress he is proud of – often real progress, built by a team with no spare capacity that did it anyway – and he watched it get the polite reception and the quiet follow-up questions. Nothing was said outright. Nothing needed to be. A leader who has run a company under a sponsor’s operating cadence can read a board’s temperature to the degree.
The fund side tends to miss what the discount costs: it is expensive for the company even when – especially when – its wins are real. A win that is not believed does not compound. It does not build the equity story the CEO’s own points ride on. It does not appear in the sponsor’s telling to its investors, so the team that did the work gets no credit for having done it. And at exit, when a buyer’s diligence team probes the claimed wins with QoE-grade skepticism, “we believe it worked but never verified it” prices like doubt – because it is doubt.
The fund is not winning this exchange either. It cannot answer its LPs with numbers it privately discounts. It cannot translate a working play to the rest of the portfolio, because it does not know which of the reported plays are real. And it may be underpricing its best-run company, which is an expensive form of caution.
Both sides lose. Not because either did something wrong, but because the structure contains no reader both sides trust.
The gap is about to be priced
For a while, this could remain a private discomfort. I do not think it can much longer, for two reasons.
The LPs started asking. AI questions now arrive inside due-diligence questionnaires during the raise, and the accounting and advisory firm CLA reports that LPs are explicitly screening for “AI-washing” – claimed capability with nothing verifiable behind it. A DDQ answer built on wins the fund itself does not fully believe is not an answer. It is a liability with a delay on it.
The buyers started asking too. FTI Consulting’s research on exit readiness finds buyers probing data infrastructure, AI-enabled compliance, and AI-driven forecasting as part of pricing a company. (FTI surveys funds a notch above this bracket – which, if anything, means the questions arrive here with less preparation on either side of the table.) At that point the satisfaction gap stops being a mood in the boardroom and becomes a line in the diligence findings – and a finding of “claimed but unverified” moves a price in only one direction.
The upshot: “we never verified it” is becoming an answer given in public – to the people who set the fund’s cost of capital and the company’s exit price.
What a believable read would take
If the problem is that no reader exists whom both sides can trust, the specification for fixing it is not mysterious. I think it has three parts.
The first is structural permission to say “not here, not yet.” A read that can only conclude yes is marketing, whatever it calls itself; the reviewer has to be free to find that a company’s AI results are real and modest, or real and larger than reported, or not yet real – and that freedom has to be structural, not promised. Nobody sitting on a payroll who needs the answer to be yes.
The second is checkability. Independence is a claim; method is evidence. A read both sides can believe itemizes how it reached each conclusion, attributes every outside number to a named source, and keeps its claims small enough that a skeptical reader – an IC, a CFO, a buyer’s diligence team – can verify them in one sitting. I have tried to practice this in public: when I put the industry’s two most famous AI statistics against their own primary documents, neither survived at face value, and both directions of the story they tell did. Verification worth the name has to be allowed to embarrass the verifier’s own argument.
The third part is the one the fund side usually gets wrong: the read has to serve both chairs, and mean it. Independent verification keeps getting described as the sponsor’s check on its companies – which guarantees the companies experience it as an audit. (Nobody volunteers context to an audit.) The truer description is symmetrical. The same independence that catches a shallow win is the only thing that makes a real one fully bankable. A verified win stops being the CEO’s claim and becomes the company’s asset: it goes into the board deck without the quiet follow-up questions, into the DDQ answer without the AI-washing exposure, into the CIM without flinching at diligence. The CEO whose progress is real – and the AlixPartners numbers say that describes a great many of them – has more to gain from an independent read than anyone else at the table. He is the one currently paying a discount on true claims.
I say this as someone who has sat in the company-side chair while the numbers were being doubted. I spent years running analytics inside a roughly $1B PE-owned telecom, building measurement the leadership had been asking for; when the company later went through a balance-sheet restructuring, the metrics my team had built grounded the five-year plan the lenders approved. Lenders in a restructuring are the least generous readers a number will ever meet. What got those numbers believed was not seniority, and it was not insistence – it was that every figure traced to a method, and the method had been built before anyone needed it to flatter a conclusion.
Two chairs, one missing instrument
The CEO presenting real progress and the partner quietly discounting it are not adversaries. They are two people being failed by the same missing instrument, each paying a different bill for it – the partner in answers she cannot give her LPs and her IC, the CEO in credit he cannot collect on work his team did. The gap will not close because either side insists harder. Insistence is not evidence. What closes it is a read neither side wrote and both sides can check.
That is the work The Portfolio AI Baseline exists to do: an independent, board-ready read of AI reality across a portfolio – an independent read the IC can lean on, written so that a company’s real wins get counted, not just its gaps. It costs the partner under three hours: a working intake and a playback; the data request goes to an analyst, and the desk work happens off your calendar. It is free for a reason I would rather state than have guessed at: the baseline is how I choose the work, and money only changes hands if the fund decides to continue past it – which is exactly why the read can afford to say “not here, not yet,” and sometimes does.