Autonomy fear looks just like distrust

You shipped the AI feature months ago. The satisfaction survey came back warm: people call it useful, they’d recommend it, they’re glad it exists. The usage dashboard tells a different story. They log in, look at it, and route the real work around it. Every adoption playbook you know reads that gap the same way. They don’t trust it yet. Show them it’s accurate, add citations, make the reasoning visible, and usage will follow trust.

For a lot of your users, that playbook is exactly right. Research published in 2026, including work from Workday, the enterprise software maker, documents how a distrusted feature quietly cancels its own value. People who doubt the output double-check it, re-run the numbers by hand, or avoid it and go back to the old way [verified]. Build the trust and the double-checking stops. Aimed at that user, the trust toolkit hits the target.

There’s a second kind of non-user, and the same toolkit does nothing for them. The Decision Lab, a behavioral research group, surveyed 1,200 workers and named a pattern it calls acceptance-with-avoidance: people who fully endorse the tool, believe it works, cleared every onboarding step, and still use it about half as often as their peers on the highest-value work [reported]. They aren’t doubting a thing. When the researchers dug into why, the reason had nothing to do with accuracy. It was autonomy [reported].

A stalled AI feature has two causes that look identical on your dashboard, and the trust playbook only fixes one of them. One person avoids the tool because they don’t believe it. The other avoids it because they believe it completely, and can feel it eating into the judgment their job is built on.

Sit with the second one, because it’s the half nobody designs for. The Decision Lab and Harvard Business Review trace the mechanism to a fear of atrophied judgment and displaced competence, not a fear of wrong answers [reported]. Ship more explainability at that person and you’re answering a question they never asked. The economic stake is large: WRITER, an enterprise AI company, reports that 95 percent of companies see zero return on their AI spending even after deploying it widely [reported]. And the distribution runs backwards from what you’d want. Praditus, a workplace research firm, finds the workers with the longest careers ahead of them carry the heaviest version of this and use AI the least [reported]. The cohort you most need compounding is the one pulling back hardest.

The obvious objection is that these are one problem at two depths. Trust gets people in the door; autonomy governs how far in they go. Two gates in sequence, no contradiction. That holds until you check what both gates are measured against. The number that matters was never logins. It’s reliance on the tool for real decisions, the strategic work it was bought to handle. That’s the exact surface where the endorser pulls back, for reasons trust never touches. These aren’t stacked gates on one hallway. They’re two different doors into the same room, and the dashboard can’t tell you which one a given user is standing at.

That’s the useful part: the two failures share a symptom and split on cause. Both show up as a flat usage line on a working, provisioned feature. The doubter gives themselves away through verification. They use the tool and check it, re-run it, keep a hand on the wheel because they don’t trust the wheel. The endorser shows the opposite shape: warm words, cold usage on exactly the strategic tasks you built the feature for. Same flat line on the chart. You read the difference from the behavior around it, never from the line itself.

This is where a quarter of engineering gets spent well or wasted. Every remedy in the trust toolkit, the calibration signals and citations and failure-mode transparency, is aimed correctly at the doubter and inert on the endorser. Push explainability at an acceptance-with-avoidance cohort and you’ll spend three months reassuring people who were never doubting you, while the real barrier sits untouched. Their fix lives somewhere else entirely: redesigning the work so the tool carries the load without hollowing out the judgment, so the person stays the decision-maker the AI assists [reported for the mechanism; the specific redesign is still unproven].

One caution, because it sets how hard you should lean. The trust findings are verified. The autonomy findings, so far, come from a single cluster of surveys, The Decision Lab, HBR, and WRITER, with no independent replication. The researchers also concede they measured people already using AI, so whether this autonomy debt blocks adoption from the start or only sets in after rollout is still open [unverified]. That’s enough to stop treating trust as the whole story. It isn’t yet enough to crown autonomy the bigger half.

So change the first question you ask about a stalled feature. For two years it’s been “how do we build trust.” Make it this instead: are these people doubting the tool, or protecting themselves from it? If they use it and double-check every output, you have a trust problem and your onboarding work is aimed right. If they tell you it’s great and won’t go near the strategic work, no amount of explainability will move them, and the thing worth finding is the signal in your own telemetry that separates the two before you’ve committed the roadmap. That’s a question you can actually answer. You just have to run the diagnosis before you reach for the cure.