Resistance Is Rational.

The standard explanation for why AI adoption stalls: people resist change. The standard prescription: more training, better communication, stronger change management. The standard outcome: nothing changes.

I keep seeing the same pattern. Organizations invest in change programs, internal comms campaigns, training sessions. Adoption numbers barely move. The conclusion is always the same: "our people aren't ready" or "our culture isn't there yet."

I think the diagnosis is wrong. People are not resisting change. They are responding rationally to the structural position they have been placed in.

It is a role design problem

When an organization deploys AI into a function, something happens that nobody names explicitly: the human is placed in a structural relationship with the AI. That relationship is not determined by the training program or the internal memo. It is determined by how the role is designed, how performance is measured, and what happens to the people who cooperate.

In most deployments I see, the human and the AI do the same work. Output is compared. The implicit question is: which one is cheaper? The human knows this. They may not articulate it, but they feel it in every interaction with the tool. They are being benchmarked against their replacement.

In that structural position, withholding is not resistance. It is self-preservation. Training does not fix it because the problem is not a skill gap. Communication does not fix it because the problem is not a misunderstanding. The problem is that the role makes cooperation against the person's interest.

You cannot train someone into cooperating with their own displacement. The incentive runs the wrong way.

Three roles, three different dynamics

Not every AI deployment creates this problem. The dynamic depends entirely on how the human role is designed relative to the AI. I see three structural positions, and each produces a fundamentally different adoption pattern.

The progression from Competitor to Principal to Architect is not a maturity ladder. These are different structural realities. An organization can have all three happening simultaneously in different functions. The diagnostic question is: which role has each function designed, and does the incentive architecture match?

Why training fails

Training solves a skill problem. Most AI resistance is not a skill problem.

A Competitor who receives better AI training becomes a more efficient Competitor - but the structural incentive to withhold does not change. A Principal who receives no training but whose role is well-designed will figure out the tool because their interest and the tool's purpose are aligned.

The training-first approach assumes the bottleneck is knowledge. In most organizations I work with, the bottleneck is structural: people understand the AI fine. They do not trust that using it well is in their interest. And in many cases, they are right not to trust that.

The metric trap

Here is where it gets worse. Most organizations measure AI adoption by usage: logins, session time, queries per user. This creates exactly the wrong incentive. People learn to perform adoption without doing it. They log in, generate an output, then do their work the old way.

When I overlay actual usage data with self-reported adoption surveys, the gap is often 30 percentage points or more. 85% say they use the tool weekly. 40% actually do. The organization has not achieved adoption. It has achieved AI theater.

The fix: measure the outcome the AI enables, not the usage of the AI. A Principal whose output improves does not need to be tracked on how often they opened the tool. An Architect whose system runs well does not need a usage dashboard. Measure AI usage and you get AI theater. Measure the outcome the AI enables and you get genuine adoption.

What honest looks like

There is something most consultancies will not say, so I will. For some roles - where the AI simply does the work better, faster, and cheaper, and there is no higher-value frontier for the human to move to - no incentive design makes cooperation rational. The role is going away.

Pretending otherwise is what produces the worst outcomes: people sense the dishonesty, trust collapses, and even the roles that should cooperate stop cooperating because the organization has lost credibility.

The honest lever for those roles is transition fairness: a clean exit with terms that make a thorough handover beat sandbagging. Generous timeline, real support, cooperation-contingent terms. The people who remain calibrate their own cooperation on how the leavers were treated. Treat them badly, and adoption poisons through the survivors, not the leavers.

The organizations that are getting AI adoption right are not the ones with the best training programs. They are the ones that designed the role before deploying the tool. They asked: what is the human's structural position relative to this AI? Is it Competitor, Principal, or Architect? And does our incentive architecture make cooperation rational?

If you have not asked those questions, you are not facing a change management problem. You are facing a design problem. And no amount of training will solve it.