McKinsey's 44% and MIT's 95% count different populations

The two numbers have different denominators. McKinsey counted organizations, and 44% of them say they are scaling AI across the enterprise ^[1]. MIT counted projects, and found 95% of generative AI pilots with no measurable impact on the bottom line ^[2]. No study here follows the same programmes from pilot into enterprise rollout. The two figures cannot be subtracted to produce a failure rate for scaled AI.

McKinsey's own number moved six points in a year, from 38% to 44%, and it reaches 54% at companies with at least $1 billion in annual revenue, ten points above the overall figure ^{[1][3][15][16]}.

Workflow redesign is the one item McKinsey placed next to a financial outcome: it found redesign has the strongest association with AI's EBIT impact ^[4]. An association across a survey population is not a controlled result. It is still the most useful sentence in the material for anyone writing a business case, because the thing it points the budget at is process change.

Greg Keith, founder of MGKgroup, calls the opposite instinct the "Shiny New Penny": he has watched organizations meet a tool at a conference or through industry enthusiasm and decide quickly that it will solve their problems, and he says AI has intensified that ^[19][7]. His explanation for the gap is governance. He argues organizations become unstable when governance and decision-making fail to scale at the same pace as the organization itself ^[8], and his Scaling Instability Curve marks the point where delivery velocity outpaces architectural maturity and operational control ^[5]. What follows, in the framework's terms, is a compounding phase of high costs, blurred team ownership and slower deployments ^[6]. Keith presents the curve as an observational framework drawn from recurring real-world experience over more than 25 years, and the write-up puts no measurements behind it ^[18][7]. The survey figures are silent on his account.

"I often return to one question as the test for whether a proposed change deserves to happen: 'Why are you doing this?'" Keith said ^[9]. On the tool itself he was blunter. "AI is a tool that should be used as such," he said ^[10]. His worry about junior teams is that AI lets them move faster while they lack the experience a critical production failure requires ^[11]. He compares that to commercial aviation. "That's what you're paying for," he said. "It's that kind of experience to know exactly what to do when there's no choice." ^[12]

Cost is where a team rolling this out has firm numbers. FinOps Foundation research in 2025, covering organizations responsible for more than $69 billion in public-cloud spending, found workload optimization and waste reduction were still the leading priorities, and that 63% of respondents were already managing AI spending ^[13][14]. That 63% sits 19 points above the 44% reporting enterprise-wide scaling, across two different survey populations ^[17].

For anyone deciding this quarter whether a pilot goes enterprise-wide, two questions come out of this material. Did a workflow change, or did the tool land on top of the workflow that was already there? And is there a bottom-line figure for this specific programme, measured before and after? If neither answer exists, the thing being scaled is the pilot, and MIT's 95% is the base rate to plan against ^[2].