Push AI adoption harder and some workers dig in

When AI adoption stalls, the playbook points one direction. Up. You look at governance, at incentives, at whether leadership is actually behind it. You add a mandate. You wire usage into the quarterly goals, build a dashboard, put an executive’s name on the whole effort. The premise under every one of those moves is that low adoption is a coordination problem. People aren’t using the tool because the organization hasn’t yet arranged for them to. Arrange it better and the numbers rise.

For most of what blocks enterprise AI, that premise is sound. The gap really is coordination more often than it’s capability. But a survey that landed this summer describes a slice of the workforce the premise gets exactly backwards.

The Decision Lab, a behavioral-science research firm, surveyed 1,200 employees across industries and found a resistance with an odd shape. These people agree the AI tool works. They’ve watched it save time, they’ll say so out loud, and they still keep it at arm’s length on the work that matters. The firm calls what they’re carrying psychological debt and breaks it into six costs. Two of them do the heavy lifting: reduced autonomy and diminished competence. The tool quietly threatens the employee’s sense that their own judgment still counts for something. So they use it for the small stuff and route the real decisions around it.

That’s a different failure from the one the playbook is built to fix. The standard cures for reluctance are trust cures. Make the model more explainable, more transparent, more accurate, prove it’s safe enough to lean on. None of those reach this. The employee who accepts the tool and avoids it anyway has no capability doubt left to answer. What they’re guarding is their standing in their own work, their sense that they’re still the one whose judgment matters.

And the levers the playbook reaches for when adoption is low are, for this cohort, the very thing that manufactures the resistance.

Amazon put a name on the mechanism without meaning to. Earlier this year the company ran an internal leaderboard, reported as KiroRank, that ranked employees by how many AI tokens they burned, tied to a goal of 80% weekly adoption. People gamed it. They handed agents busywork to climb the rankings, a habit staff called tokenmaxxing. One project using Claude for a menial coding task ran 860% over budget, roughly $1.8 million, and nobody caught it for five months. Two others overran by $541,000 and $134,000. Amazon’s own senior vice president, Dave Treadwell, pointed at the leaderboard’s good intentions as the direct cause, killed it, and swapped raw usage tracking for a metric aimed at useful output. All of this is reported from internal figures, not independently audited, but multiple outlets carry the same account.

Now read that leaderboard back through the survey. A ranking that forces usage and ties it to a target is, from the employee’s chair, a direct hit on the two costs the survey named. It strips autonomy, because you no longer choose when the tool helps. And it tells you plainly that you’re not trusted to reach for it on your own; you have to be made to. Amazon built an organizational lever of precisely the kind the playbook prescribes for low adoption, and it produced the individual resistance the survey measures. The survey shows the resistance sitting in people. Amazon shows the lever that generates it. They’re the same collision seen from opposite ends.

So the playbook’s most confident move inverts on this cohort. Adoption is low, so you mandate it. You mandate it, and for the employee already protecting their autonomy, you’ve deepened the debt. The debt deepens, the avoidance hardens, the dashboard still reads low, and the prescribed response is to push harder. The lever is the load. Pull it to raise adoption and, for these people, you lower it.

None of this sinks the organizational view in general, and it shouldn’t. Most of the enterprise AI gap really is coordination: governance nobody built, incentives pointed the wrong way, leadership that never showed up. For the employee who isn’t using the tool because no one arranged for them to, the organizational fix is the right fix, and it works. The debt cohort is a specific slice, not the whole payroll. The trouble is that it’s a slice the standard playbook can’t see, because every instrument the playbook carries reads low adoption as a coordination gap and answers with more coordination.

Which points at a test the playbook has never had to run. Name one organizational lever that reaches an employee who already agrees the tool works and avoids it to protect their sense of competence. Governance, incentives, mandates, dashboards, orchestration layers. Every one of them is a push, and for this cohort the push is what generates the debt. If the honest answer is that there isn’t one, then the binding constraint has moved somewhere no amount of org design can touch without inflaming it.

There’s one thing worth settling before you decide how much this bites, and the survey couldn’t answer it. Does psychological debt block adoption from the start, or does it accrue after rollout, in proportion to how hard the rollout was pushed? That timing is still unverified. But if it’s the second, the debt isn’t a curiosity at the margins. It’s how pushing harder turns into the very ceiling you were pushing to break. The governance-mature org that can’t move its numbers isn’t stalled despite pulling every lever. For these people, it’s stalled because it did.

When the rollout sticks, the reflex is to look up, at governance, incentives, the org chart. For a real and maybe large part of your team, the thing to look at is down and in: whether your rollout is asking people to trade away their sense of their own judgment for a tool they’ve already told you is good. The playbook only ever points you up. On this, that’s the wrong way to look.