95% of AI Pilots Fail the Same Way: Tools-First

MIT found almost no company with measurable returns from generative AI. The models were not the problem. The order of change was.

The number

In 2025, MIT’s Project NANDA reported that 95% of organizations see no business return from generative AI, after an estimated 30 to 40 billion dollars of enterprise spending.

Treat that number with care. The study rests on interviews, surveys and a review of about 300 public AI initiatives, and critics point out that it only counts effects within six months. It shows a direction, not a law of nature.

Not a model problem

The study does not blame the models. Its central explanation is a learning gap in both the tools and the organizations using them.

Two more findings point the same way. Of all organizations, 60% evaluated enterprise AI systems, 20% reached a pilot and 5% reached production. Meanwhile, only 40% of companies bought an official subscription, but employees at more than 90% of them use personal AI tools for work.

Read those together. The people crossed the divide on their own. The organizations bought tools and waited for adoption.

Three orders of change

Every transformation changes people, processes and tools. The order is a choice, and each order fits a different group. This is a comparison, not a ranking.

Order	Fits	How it runs
People-first	Builders and trailblazers	People pick and shape the tools, the process follows
Process-first	Case work and everyone who prefers clear direction	Define the process, buy the tools, staff the roles that operate them
Tools-first	Specialists and domain experts	A tool adds one specific new capability

Tools-first is the right call for a specialist. Applied to the whole organization, it stays IT modernization. The typical failed pilot does exactly that: buy a tool, roll it out to everyone, count the logins.

One strategy for everyone is the mistake

In my keynote model, every organization has three groups: builders at roughly 0.1%, a capability-augmented group at roughly 10%, and commodity users at roughly 90%. These are working estimates, not survey data.

The mistake is not picking the wrong group. It is running one strategy for all three.

Competence has three parts

Transformation competence is skill, capability and willingness. All three must be present, or nothing changes.

A tool rollout addresses none of them directly. Willingness in particular cannot be purchased. It comes from being involved in how the work changes. That is why the Souveränitäts-Kompass measures transformation competence as its 14th axis.

The people who know where a process really breaks are often the ones closest to retirement. I wrote about that in Your Most Valuable Context Retires by 2039.

Measure the order, not the tool

Pilot-to-production rate per group. How many pilots reach production, split by builders, augmented and commodity users.
Time to first production use case per group. Weeks, not quarters.
Share of use cases owned by line managers. The MIT study names empowering line managers, not only central AI labs, as a success factor.

Logins and licenses are not on this list. Adoption is irrelevant if it isn’t measurable.

The order matters

If you transform people-first, people solve problems. If you transform process-first, people satisfy processes. If you transform tools-first, you modernize IT and call it transformation.