A small but important pattern keeps showing up in recent workplace AI research. More people have access to AI. More people use it at least occasionally. Leaders expect bigger gains. And yet the actual change in how work gets done is much less dramatic than the slide decks suggest. Gallup reported in April 2026 that half of employed Americans now use AI in their role at least a few times a year, but also that the evidence for deep organizational change is still limited. Deloitte's 2026 enterprise AI report points to a similar tension: many companies capture productivity gains, while far fewer truly redesign the business around the new capability. That gap is usually described as an adoption problem. I think that is only partly true.
The convenient story is too easy
The comfortable explanation is that employees are slow, afraid, or stuck in old habits. Sometimes that is accurate. People can defend familiar routines long after those routines stopped making sense. But treating every hesitant employee as a change-management obstacle is intellectually lazy. It turns a diagnostic problem into a moral judgment. If someone avoids a tool that creates ambiguous accountability, exposes them to compliance risk, produces work that their manager may not trust, or adds review work without changing priorities, the hesitation may be completely rational.
Access is not safety
Access is not the same as permission. Permission is also not the same as safety. A company can buy licenses, publish an AI policy, run a webinar, and still leave people guessing about what actually happens when AI-assisted work goes wrong. Who owns the error? Is the employee expected to disclose every use of AI? Does using AI make the work look less valuable? Will the team reward experimentation, or only punish visible mistakes? These are not philosophical questions. They are practical questions about status, accountability, and risk. People notice the answers even when leaders avoid saying them out loud.
Training does not solve the deeper calculation
This is where many AI adoption conversations become too shallow. They assume the blocker is knowledge: employees need training, examples, prompt libraries, maybe a few internal champions. Those things help, but they do not resolve the deeper calculation. A person can understand how to use AI and still decide that using it is not worth the social or professional exposure. In that sense, resistance is not always ignorance. Sometimes it is a reasonably accurate reading of the environment.
The manager is the real interface
Gallup's data is useful here because it points away from pure enthusiasm and toward context. Employees use AI more often when it fits existing workflows, when managers actively support its use, and when expectations are clear. That sounds obvious, but obvious things are often the ones organizations skip because they are less exciting than announcing another platform rollout. Workflow fit is not a feature of the model. It is a feature of the organization. Manager support is not a motivational poster. It is a signal about what is rewarded, tolerated, questioned, and protected.
The manager becomes the real interface between the employee and AI. Not because managers should approve every prompt, but because they translate abstract strategy into lived risk. If a manager says, explicitly or implicitly, that AI is welcome only when it produces perfect work without extra discussion, people will use it cautiously or privately. If a manager treats AI-assisted drafts as normal inputs that require judgment, revision, and ownership, the tool becomes easier to use responsibly. The difference is subtle, but it matters. One environment makes AI feel like a trap. The other makes it feel like part of the work.
The product enters a social system
There is also a product lesson here. Many enterprise AI products still assume that adoption is a user-interface problem. Make the tool easier, put it where people already work, reduce friction. Sensible, but incomplete. The product is entering a social system. It changes who appears competent, who carries risk, who gets credit, and who has to verify the output. If that system is not understood, the product may be technically good and still emotionally expensive to use.
Resistance is signal before obstacle
I do not think every form of AI resistance deserves endless patience. Some people will protect comfort and call it principle. Some teams will hide behind ethics when the real issue is fear of being less special. But the opposite error is more common in ambitious organizations: dismissing hesitation before asking what it is detecting. A good signal can arrive in an inconvenient tone. The point is not to romanticize resistance. The point is to inspect it before trying to overpower it.
The question, then, is not only how to make employees use AI more often. That is too narrow. A better question is: what would have to be true for a reasonable person to use it without feeling foolish, exposed, or complicit in lower standards? If the answer is unclear, the adoption problem may not be in the employee. It may be in the environment the employee is reading quite well.