The 2025 and 2026 evidence puts the gap in how the work was rearranged around the tool. That changes who an enterprise should put in charge.
In brief
- Eight in ten respondents say AI improved their own productivity. About six in a hundred attribute 5% or more of EBIT to it, unchanged from a year ago.
- MIT NANDA’s study says its divide does not seem driven by model quality or regulation. McKinsey finds high performers redesigned workflows at three times the rate of everyone else.
- Read another way, the talent that closes the gap works across the technical and the social. No study tests that directly. It is my reading of what the studies measure.
Where the money is lost
The data shows scaling rising and return standing still. In McKinsey’s 2026 survey (1,719 respondents, fielded May to June), 80% say AI improved their individual productivity and 44% say it is scaling across the enterprise, up from 38% a year earlier. Yet 37% attribute any EBIT impact to AI, about the same share as last year. The high performers, those attributing at least 5% of EBIT to AI, remain at about 6% of respondents (Exhibit 1).
Eight in ten feel the gain. About six in a hundred see it clearly in EBIT.
Individual productivity is widely reported; enterprise return is not.
Source: McKinsey Global Survey, “The state of AI in 2026” (Aug 25, 2026; n=1,719). Self-reported.
MIT NANDA’s 2025 study, built on interviews and a survey of leaders, found the same shape: 95% of organizations getting zero return, and 5% of integrated pilots extracting millions in value. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls.
Why a better model is not the fix
NANDA says the divide “does not seem to be driven by model quality or regulation.” It points instead to a learning gap: most systems do not retain feedback, adapt to context or improve over time. The buyers who crossed the divide sourced initiatives from frontline managers, not central labs. Partnerships with customized tools reached deployment about 67% of the time, against about 33% for internal builds, though the report calls these figures self-reported and does not claim causation.
McKinsey shows the same split inside the enterprise. Nearly three-quarters of high performers fundamentally redesigned workflows because of AI. One quarter of other respondents did.
Feeling is a poor guide to any of this. In METR’s 2025 trial, 16 experienced developers expected a 24% speed-up, took 19% longer, and afterwards believed they had been 20% faster. METR now calls those results out of date. Its February 2026 update says developers are likely more sped up today, but rates its newer data as an unreliable signal. The narrower point survives: what users report and what is measured can sit far apart, and only one of them reaches EBIT.
Read another way
Seen through how the parts fit together, these studies measure one arrangement from different sides. That is my reading, not their finding. Workflow redesign is process. Frontline sourcing against central labs is structure. Sentiment against operating results is metrics. The learning gap is technology, and what a tool does after a quarter of use depends on what the organization lets it retain. Who does the redesign, and whether they can see how a team works as well as what a model does, is talent.
No study tests that last link directly, and NANDA’s own explanation sits on the tool side. The evidence does show that high performers changed the work. If the return is made there, the person put in charge of an AI program needs to read a model’s failure modes and also watch how a team works around it. That is a technical profile with social-science training on top. It is an appointment decision before it is a tooling decision.
The outcome is made along the links, not inside the tool.
The studies above point to losses at several of them.
Source: Mentor Cana analysis. A reading, not a measurement.
What to do about it
Executive Team, next quarter. Ask every AI initiative owner to name the workflow that changed and the line manager who signed it off, and review the operating measure it should move beside the satisfaction score. Pause funding where neither can be named. The basis is the 80% against 37% in Exhibit 1 and the redesign split, an association, not proof.
Executive Team, next 12 months. Before any agent goes live, decide who reviews its actions, who may override it and who answers when it is wrong. Gartner names inadequate risk controls among its causes of cancellation. The design is my reading.
Transformation Leaders. Now, let the manager who runs the process write the requirement and hold sign-off, with the central team in support. NANDA reports frontline sourcing among buyers who crossed the divide; causation is open. Add one test to every tool trial: after a quarter of real use, what has the tool retained, and what changed in its behavior as a result? That targets NANDA’s learning gap. Over the next 12 months, when you select the program lead, ask candidates to explain a model’s failure mode and then how a team works around it. That follows from my reading above, not from a study.
In your organization, who owns the redesign of the work around the tool, and what do they report to you?
Sources
- McKinsey Global Survey, “The state of AI in 2026,” Aug 25, 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- MIT NANDA, “The GenAI Divide: State of AI in Business 2025,” July 2025 (copy hosted by AI News; pp. 3, 19, 20). https://www.artificialintelligence-news.com/wp-content/uploads/2025/08/ai_report_2025.pdf
- Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity,” July 10, 2025. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- METR, update on developer productivity, Feb 24, 2026. https://metr.org/blog/2026-02-24-uplift-update/
Leave a Reply