Boards are seeing AI productivity gains across every function. Engineers ship more. Analysts produce more. Decision cycles accelerate at the edge. The dashboards turn green. EBITDA does not move.

That is not an AI problem. It is an operating model problem.

The productivity gains of AI are captured by individuals. The losses are absorbed by the organization. And in most mid-cap companies, that organization was architected for a pre-AI information flow that no longer exists.

The mechanism

Legacy organizations were built for slow information. Layers existed to filter and route. Approval cycles existed because the system could not synchronize itself. Coordination meetings existed because that was the only way to align decisions across functions.

AI changes the speed at the edge. It does not change the architecture at the core. A developer with AI assistance writes materially faster. The code still goes through the same review process, the same approval cycle, the same release window. The bottleneck moves from production to coordination. The team accelerates. The enterprise does not.

The atomic unit now operates at AI speed. The whole still operates at pre-AI speed. The asymmetry produces friction, fatigue, and disappointment, but rarely EBITDA.

The wrong response

The boards that try to solve this asymmetry by reducing headcount are solving the wrong problem. Headcount reduction takes out the individuals whose productivity gains were real. It leaves intact the organizational architecture that was dissipating those gains.

The result is fewer people doing the same effective work, with productivity still trapped at the individual level. The visible savings appear in Q3. The invisible cost appears in the EBITDA that does not move.

The right reframing

The right question in 2026 is not how many people can be reduced given AI productivity gains. The right question is what operating model would let the company actually capture the gains that already exist at the individual level.

That work is harder. It requires redesigning the connective tissue of the organization, not just the headcount lines. Removing layers of approval that no longer make sense. Shortening the cycles between insight and decision. Giving operating teams the authority to act on AI-generated information without escalating it through legacy validation paths.

This is difficult work because it is invisible, politically expensive, and structurally disruptive. It is also the work that determines whether AI productivity gains compound into a multiple at exit, or evaporate quarter after quarter in coordination overhead.

The compounding question

AI is not the bottleneck. Your operating model is. The companies that defend their multiples at exit will be the ones whose boards understood the difference early enough to act on it.

Productivity your operating model cannot absorb is not a gain. It is leakage.