This is the third piece in our series on operational drag. It takes apart the most quoted statistic in enterprise AI. For the full framework, start with The Hidden Cost of Operational Drag.
The 5 percent did not have better AI. They had better maps.
In the summer of 2025, an MIT research report put a number on a problem enterprise AI programs were already encountering: 95 percent of enterprise generative AI pilots showed no measurable profit-and-loss impact.
The number traveled faster than the finding.
It was widely repeated as evidence that enterprise AI was failing. That is not what the study found. Adoption was high. Employees were using the tools. Organizations were running pilots.
What was missing was measurable financial impact.
The researchers documented problems beyond the technology itself. AI tools struggled to integrate with existing workflows, retain context, and adapt to feedback.
The pilot was new. The friction underneath it was not.
What the Study Actually Measured
The report, The GenAI Divide: State of AI in Business 2025, was published by MIT's Project NANDA.
Researchers combined interviews with representatives from 52 organizations, survey responses from 153 senior leaders, and an analysis of more than 300 public AI deployments. Against an estimated $30–40 billion in enterprise generative AI spending, the researchers found that 95 percent of organizations in their sample were seeing no measurable P&L return from their GenAI investments. Only about 5 percent of integrated pilots were extracting significant value.
The finding was not that the technology failed or that employees rejected AI.
More than 80 percent of organizations had piloted tools such as ChatGPT or Copilot, and close to 40 percent reported some form of deployment.
People were using AI. Organizations were struggling to turn that usage into measurable financial results.
MIT called it the GenAI Divide.
Is the 95 Percent Number Real?
There are legitimate reasons to be careful with the headline.
The interview sample was modest, and measurable P&L impact is a demanding standard. Some initiatives may also require more time to produce financial results than the pilots had been given.
There is another problem with measuring financial impact: a pilot cannot demonstrate improvement against a baseline that never existed.
That does not make the MIT findings irrelevant. It makes the definition of failure important.
Subsequent research has continued to show a large gap between AI adoption and enterprise-level value, although the studies use different methodologies and should not be treated as measuring precisely the same thing.
BCG surveyed 1,250 companies and found roughly 60 percent generating no material value from AI, while about 5 percent were creating substantial value at scale.
McKinsey reported that 88 percent of organizations were using AI in at least one business function, while more than 80 percent reported no meaningful impact on enterprise-wide earnings. Roughly 6 percent qualified as AI high performers.
S&P Global Market Intelligence reported an increase in organizations abandoning AI projects. RAND found AI projects failing at a higher rate than conventional IT projects. Gartner forecast substantial cancellation of agentic AI initiatives through 2027.
These studies do not independently prove MIT's 95 percent figure. They point to the same broader problem.
AI adoption is outpacing organizations' ability to produce measurable value from it.
Why Pilots Fail: Four Mechanisms
The useful part of the MIT report is not the headline number. It is what sits underneath it.
Four patterns help explain the gap.
The money goes where AI demos well.
More than half of enterprise generative AI budgets flowed toward sales and marketing.
The strongest reported returns appeared in less visible operational work: reducing outsourced processes, agency costs, and administrative overhead.
The visible use cases attracted investment.
The operational ones produced value.
The tools do not learn.
The report describes a persistent learning gap. In the authors' words, most GenAI systems "do not retain feedback, adapt to context, or improve over time."
That becomes a problem when AI moves beyond isolated tasks and into existing workflows.
A tool that needs the same context explained every session creates another layer of work. It may accelerate a task without improving the system around it.
Integration becomes the project.
Externally sourced tools reached deployment roughly twice as often as internal builds: about two-thirds of the time versus one-third.
The difficult part was not simply generating an intelligent response. It was connecting the technology to the systems, exceptions, approvals, dependencies, and undocumented practices surrounding the work.
Internal teams often discovered that complexity after the pilot had started.
The report describes an enterprise paradox: the largest firms ran the most pilots while completing proportionally fewer deployments.
More organizational seams mean more places for integration to fail.
There was no baseline underneath.
A pilot can only demonstrate measurable improvement if the organization understands what happened before the pilot.
How long did the process take?
Where did work stop?
How many people touched it?
What did those delays cost?
Where did errors or rework occur?
Without those answers, AI can land on workflows that were never fully mapped, timed, or costed. Even when the technology improves something, there may be no credible baseline against which to measure the change.
The inability to prove ROI may not be a reporting problem alone. It can reveal how little the organization understood about the operation before the pilot began.
That unmapped friction is operational drag.
AI did not create it.
The pilot exposed it.
The Workforce Crossed the Divide Anyway
The report found another gap.
Only about 40 percent of companies had official LLM subscriptions, while roughly 90 percent of employees at surveyed companies reported using personal AI tools for work.
Employees were adopting AI independently, outside sanctioned enterprise programs.
That creates legitimate governance and security problems. It also creates a useful signal.
Employees may be reaching for tools where existing processes are slow, repetitive, or unnecessarily manual. Their behavior can help identify where friction is concentrated.
Organizations can treat shadow AI exclusively as something to suppress. Or they can study where it appears and ask why employees needed the workaround in the first place.
They voted with their browser tabs.
What the 5 Percent Do Differently
The successful organizations described across the research do not appear to have access to fundamentally different AI.
They approach deployment differently.
They narrow the problem instead of spreading pilots across the organization.
They integrate tools into existing workflows rather than placing another interface beside them.
They put adoption close to the people who understand the work.
They target operational problems where cost and friction already exist.
They favor systems capable of retaining context and improving through feedback.
Those behaviors depend on something more basic.
They understand the process before they automate it.
They know where work slows down, what the existing process costs, what exceptions matter, and what success can be measured against.
That understanding informs which problem to address, which tool to use, how to integrate it, and how to measure the result.
The 5 percent did not have better AI.
They had better maps.
That map is buildable.
The Operational Drag Audit is a free, 25-question self-assessment that scores your company across the five types of operational drag and identifies where friction is concentrated.
Run it before your next pilot, not after. Take the Operational Drag Audit
Next in the series (Coming Soon): The Verification Tax — Why AI Can Make Your Best People Slower.
FAQ
Is the MIT statistic that 95% of AI pilots fail accurate?
MIT's Project NANDA reported that 95 percent of organizations in its sample saw no measurable P&L return from their generative AI investments. That is narrower than saying 95 percent of AI technology failed. Other research has identified a similar gap between adoption and measurable financial impact, although the studies use different methodologies.
What did the MIT GenAI Divide study measure?
The study combined interviews with representatives from 52 organizations, survey responses from 153 senior leaders, and an analysis of more than 300 public AI deployments. It examined whether enterprise GenAI investments were producing measurable P&L impact.
Why do enterprise AI pilots fail?
The research identifies several recurring problems: investment directed toward visible rather than operational use cases, tools that cannot retain context or learn from feedback, and difficult integration. Missing process baselines can make it harder to demonstrate whether a pilot improved the underlying operation.
What do successful AI deployments do differently?
Successful deployments tend to address narrower operational problems, integrate into existing workflows, involve the people who understand the work, and establish a baseline against which improvement can be measured.
What is the shadow AI economy?
Shadow AI refers to employees using AI tools for work outside officially sanctioned company systems. Beyond its governance implications, that usage can reveal where employees are encountering enough operational friction to seek their own tools.