You have probably met both of these numbers by now. The first ran as a headline almost everywhere last year: 95 percent of AI pilots fail. The second circulates in private equity reading piles: 95 percent of funds with AI programs in production report initiatives performing at or above their business case. The first traces back to an MIT study; the second comes from FTI Consulting’s 2026 Private Equity AI Radar. Same technology, same industry conversation – and the two most quoted statistics about it point in opposite directions.
So which one is wrong? Neither. They are two ends of the same funnel. Roughly 95 percent of what enters the funnel dies before production. Roughly 95 percent of what reaches production pays. Put those together and the conclusion is uncomfortable for anyone selling tools and useful for everyone else: the value of AI in a portfolio is mostly decided before anything launches, in the selection of which company, which use case, which way, and whether at all.
That claim needs defending – starting with the suspicion both numbers deserve and rarely get.
What each study measured
“95 percent of AI pilots fail” is the popular press’s paraphrase, not the study’s finding. The study – MIT’s Project NANDA, published July 2025 – reported that 95 percent of organizations are getting zero return from generative AI, against an estimated $30–40 billion of enterprise investment. The useful part is the funnel behind that number. For task-specific AI tools, about 60 percent of organizations evaluated them, 20 percent piloted them, and 5 percent got them into production. Everything else stalled somewhere between the demo and the workflow.
And it is a small early study: 52 interviews and 153 surveyed leaders, findings the authors themselves call preliminary. Their own limitations section concedes that grading success only six months after a pilot “may be insufficient,” potentially understating success rates. Remember the six-month window.
Now the optimistic number. FTI surveyed 200 senior private equity decision makers – managing partners, operating partners, principals – in December 2025, all at funds of $1 billion in assets and up, so read it as describing funds a notch above most of the lower middle market. Among funds with AI programs in production, 95 percent report initiatives at or above the original business case: 17 percent significantly exceeded it, 48 percent exceeded it, 30 percent met it. Note the unit carefully. This is not “95 percent of AI initiatives succeed.” It is funds that got programs into production, reporting on those programs – the survivors of the funnel, grading the survivors.
Neither number survives diligence at face value
A skeptic can pick both apart, and should. MIT’s study is small, interview-based, and graded P&L impact on a six-month window its own authors admit is short. FTI’s number is self-reported performance against self-set targets – and the targets are modest. In FTI’s own data, 63 to 65 percent of initiatives aim for improvement in the 5-to-10-percent range, below the 15 percent cross-industry benchmark FTI itself cites from Gartner. Most of its funds expect time-to-value at 7 to 24 months (69 percent of respondents). Which means FTI’s respondents would flunk MIT’s six-month test, and MIT’s dead pilots were never graded on FTI’s clock. Different moment, different ruler.
Both publishers also have something to sell – FTI a transformation practice, the MIT group an agentic-infrastructure research agenda. Every famous number arrives with a sponsor.
That is how research gets funded – and why the fine print repays more attention than the press release.
So neither number survives diligence at face value. Both directions do. Larger, sturdier surveys point the same two ways: McKinsey, across 1,993 respondents, found 88 percent of organizations using AI regularly but only 39 percent attributing any EBIT impact to it; a BCG survey of 1,803 executives found roughly 60 percent getting little material value. High mortality before the P&L. Real payoff for a disciplined minority. The funnel shape holds.
It is also, not incidentally, the shape of an answer that survives a limited partner’s follow-up question. AI has made its way into due diligence questionnaires, and the reconciliation above – both numbers, fine print shown – does better in that conversation than either number quoted alone.
What kills the 95 percent that die
Not the models. The consistent finding across all of this research is that failure concentrates in how AI meets the organization, not in what the technology can do. BCG’s writers on private equity call the failure pattern Deploy without Reshape: handing out AI tool licenses while changing nothing about how the work gets done. Deployed that way, the productivity never reaches the P&L, the pilot produces no measurable value, and the initiative quietly dies – which is a fair description of most of the pilot graveyard this industry accumulated between 2023 and 2025.
That is worth saying plainly, because the graveyard usually gets blamed on the wrong people. Dead pilots indict the deployment model, not the people who ran them. MIT’s strangest finding says as much: while only about 40 percent of the companies it surveyed had bought an official AI subscription, employees at more than 90 percent of them were already using personal AI tools for their work, regularly and by their own account productively. The capability and the appetite were in the teams all along. What failed was the way the official initiative met the day-to-day work.
FTI’s conclusion could have been written for this argument. The funds outperforming with AI do so “regardless of fund size, industry or AI spend,” the report notes, and then lands the sentence that matters: “What sets these funds apart is not how much they invest in AI. It is where and how they deploy it.” Where and how. Placement, in other words – decided before launch, when the initiative is still a line in a plan and the money is still in the budget.
The constraint set a notch below the samples
Both studies look upmarket from where most lower-middle-market portfolios live. FTI’s floor is $1 billion in fund assets; MIT’s sample leans enterprise. A notch below, at founder-built companies doing $20–150 million in revenue, the placement question gets harder and more concrete, for a reason neither report dwells on: the data.
Plenty of companies at that size run finance on QuickBooks with a part-time controller. There is no warehouse and no data team; the systems are the ones that grew up with the business and did their job – they were built to run the company, not to feed a model. I spent part of my career as VP of analytics inside a roughly $1B private-equity-owned telecom, building the data function from scratch, and even at that scale, establishing an honest inventory of what the data could support was a project in its own right. A placement decision that ignores what the data can carry is not a plan. It is how the next dead pilot gets commissioned.
Speed cuts the other way, in the smaller company’s favor. Korn Ferry’s practitioners report lower-middle-market PE expecting proof-of-concepts in two to four weeks – an anecdote from the field rather than a survey, but it matches what MIT found: top mid-market performers went from pilot to implementation in about 90 days, against nine months or longer at large enterprises. Smaller companies clear the funnel faster when the placement is right. They simply have fewer places where it is right, and less slack to absorb the attempts that were not.
A defensible front gate
If selection is where the value is decided, then selection deserves a method, not an instinct. The scaffolding I find most useful is BCG’s own ladder – Deploy, Reshape, Invent – used as triage rather than as ambition. For each company the question is what the honest ceiling is this year: Deploy (tools and hygiene, cheap but rarely material on its own), Reshape (a specific workflow rebuilt so the gain lands in the P&L – where the funnel’s survivors mostly live), or Invent (AI in the product itself – board-level work with no playbook, and the right answer for very few).
A placement decision that can be defended to an investment committee has to weigh at least five things: whether the company can absorb change this year at all; whether the data can carry the use case; whether a named owner exists who wants the win; whether there are enough related use cases that the second one gets cheaper because the first one was built; and whether the lever touches revenue or only cost. None of that requires a single license to be purchased first. All of it determines whether the license ever pays.
And the gate has to be allowed to say no. “Not here, not yet” is the verdict that makes the rest of the list credible – to the partners who have to defend the spend, and just as much to the company that would otherwise be asked to attempt AI it cannot yet support, at the cost of a quarter it cannot get back. If every company in a portfolio gets a green light, nobody did the selection. A no-go list is not pessimism. It is the evidence that someone looked.
The game is played before launch
Back to the two numbers. Ninety-five percent of what enters the funnel dies before production – and the deaths trace to placement, not to the people or the technology. Ninety-five percent of what reaches production pays – as reported by the funds that placed it, against targets they set on clocks they chose. Neither statistic tells you AI works or does not work. Together they tell you where the outcome is decided: before launch, at the front gate. That selection is the work.
The front gate is also a bounded piece of work – a diagnosis, not a transformation program. That is what The Portfolio AI Baseline is for: a placement read across a portfolio, done for free, costing the partner who commissions it under three hours of their own time. Free is not a favor: the baseline is how I choose where to work, and I only make money if the fund decides to continue past it. If the funnel argument above holds, the first rational spend on AI is not a tool or a hire. It is the selection.