A hundred rowers pulling harder. The ship doesn't move.

You see 90% of your teams using AI every day, and you conclude the transformation is underway. That's the trap, and it's measurable.

MIT studied 300 enterprise AI deployments. The finding is brutal: despite 30 to 40 billion dollars invested, 95% of projects show no measurable return on the bottom line. Only 5% extract real value. The report names the fault line: high adoption, weak transformation. In plain terms, the tools make the individual faster, but that gain never crosses over to the collective level.

In Switzerland, the 2026 Data & AI Observatory says the same thing differently: individual adoption is climbing fast, yet only 11% of organizations have redesigned roles around human-AI collaboration. The rest handed out tools. They didn't redesign the work.

Why does the gain stay stuck? Goldratt explained it forty years ago with the theory of constraints. A system's throughput isn't set by the speed of its fastest parts, but by its bottleneck. Speeding up everything that isn't the bottleneck doesn't raise throughput: it simply piles work up in front of the obstacle. An organization full of faster individuals can stay a perfectly motionless organization. That's the difference between a sum of local optima and an optimal system.

The individual gain is real and valuable, but on its own it stays the prisoner of whoever earned it. Three conditions make it cross to the collective level.

1. Find the constraint, don't speed up the fast. Look for the bottleneck before deploying broadly. Speed gained anywhere else is a cost dressed up as progress.
2. Redesign the roles, don't just equip the people. This is what the 11% do. The gain only crosses over if the work is rebuilt around it, not if a tool is laid over the old organization.
3. Measure the system's throughput, not the speed of individuals. Change the dashboard: move past usage rates to track end-to-end lead time, quality, value delivered.

So the right question for your next board meeting isn't "how many of our people use AI?" Nearly all of them, and it proves nothing. It's: where does the individual gain stay stuck, for lack of redesigning the work around it?