Ask an executive why an AI pilot failed and you will usually get a technical answer. The model hallucinated. The data was not ready. The vendor oversold it. The integration was harder than anyone expected.
Those answers are usually true. None of them is the reason.
The reason is almost always something that happened months earlier, in a room where nobody believed a decision was being made. A use case got picked because it would demonstrate well. Nobody said out loud who would own the result after the project team dissolved. Nobody agreed what would count as success, so every outcome could later be described as one. Nobody was positioned to stop it, so it never stopped. It just got quieter.
Here is the part that should worry a leadership team more. The pilots that succeed frequently produce nothing either. The demonstration works. The hours-saved number is real. The room nods. A year later the workflow is unchanged, the budget is unchanged, and nobody can say where the value went.
Success and failure turn out to have the same cause. In both cases, the decisions that determined the outcome were never made deliberately. They were made by default, late, and usually by whoever happened to be in the room when the deadline arrived.
The Evidence does Not Agree with Itself, and That is Informative
Any leader who has tried to build an evidence-based recommendation on AI has run into the same wall. The research does not line up.
A 2025 report from MIT's Project NANDA described most enterprise generative AI initiatives in its sample as producing no measurable financial return. Wharton, looking at a different population, reported broad adoption, rising budgets, and positive returns from most respondents. RAND, examining AI and machine-learning projects more generally, documented the recurring root causes behind failed deployments: unclear problem definition, weak data, inadequate infrastructure, and technology chosen before the underlying user problem was understood.
Read carelessly, those studies look like a fight. Read carefully, they are measuring different layers of the same adoption curve. One is measuring individual use. One is measuring executive sentiment and budget behavior. One is measuring whether integrated, task-specific initiatives produced sustained financial impact. They can all be correct at once.
The user-level evidence points the same direction. When Anthropic invited people to describe how they actually used AI, more than eighty thousand participants across one hundred fifty-nine countries responded during a single week in December 2025. People reported meaningful productivity gains. About one in five also voiced concern about economic displacement. Senior professionals reported greater personal benefit than early-career workers. Many described AI less as a speed tool and more as a way to expand what they could do at all.
Then there is the part most organizations have not measured. A late-2025 industry survey found that forty-nine percent of employees admitted using AI tools without employer approval. A separate study put the number higher still, with roughly eight in ten employees using unauthorized tools and sixty-eight percent of security leaders admitting to unauthorized use in their own daily work. The precise percentage moves with survey design. The operating fact does not. Employees did not wait for a formal program to begin.
Feasible is Not the Constraint
Put the studies together and the shape of the problem appears. Individual AI use is widespread and frequently ahead of formal organizational approval. Measurable value from integrated enterprise initiatives is far less common. Those two facts are not in conflict. They expose the same gap: people are finding value faster than organizations are learning how to govern, scale, and measure it.
This is what we mean by feasible, not achieved. Most of what an enterprise does is now technically within reach of AI. Very little of it has actually been converted into performance the business can see. The distance between those two numbers is not a technology gap. It is an organizational one.
Generative AI also inverted the order in which enterprise technology usually arrives. Earlier systems required the company to buy, install, and authorize something before an employee could use it. Now a single employee can open a browser tab, change how work is done, and create value or risk before procurement, technology, human resources, legal, or leadership has shaped a path around it.
That means adoption numbers alone prove very little. They may show curiosity. They may show workaround behavior. They do not show that an organization has managed the change.
The Question that Actually Decides the Outcome
The pattern is familiar to anyone who lived through earlier waves. Enterprise resource planning exposed processes companies had never truly redesigned. Customer systems exposed data that was duplicated, inconsistent, and ungoverned. Cloud let teams create infrastructure faster than leaders could assign ownership, cost control, or retirement rules. In every wave, the hardest problems appeared after the technology worked, when people had to change routines, managers had to change expectations, and leaders had to change how decisions were made.
AI makes that pattern faster and more personal, but the management lesson is unchanged. The hardest question is not what AI can do for your company. It is what your company is willing to stop doing, stop funding, and stop protecting to let something new take hold.
Before approving more pilots or buying more tools, a leadership team can answer three things and learn most of what it needs to know. What AI activity is already happening inside the organization, and can we separate individual use, experiments, pilots, deployed capabilities, and measured value? Which AI claims are we relying on, and do we know whether they are independently supported, company-reported, vendor-sponsored, or still assumptions? And before we scale anything, what must be true around the technology for the change to endure?
Most organizations cannot answer the first one. Until they can, every further investment is being made on top of an unmeasured base. Feasible is no longer the constraint. Achieved is, and achieving it is a leadership problem before it is a technology one.
Adapted from Feasible, Not Achieved: Why AI Transformation Stalls and the NINE Decisions That Finish It, by Weston A. Jones, MBA and Jacob C. Roxon, PhD (JRNine Press).
The Jones-Roxon Nine™ is the Operator’s System for AI Transformation.