Essay

Most AI projects don't fail on the technology. They fail on adoption.

What a behavioral scientist sees in stalled AI rollouts

Walk into any organization that “adopted AI last year” and ask a simple question: who is using it, daily, for real work?

The honest answer is usually a small fraction of the people who were supposed to. The licenses are paid for. The kickoff went fine. The technology works. And adoption quietly stalled anyway.

That’s not a technology failure; it’s a human one, and it’s predictable.

Three ways rollouts stall

Trust is miscalibrated

People don’t need to trust AI more; they need to trust it accurately. Some ignore the tool entirely because one early error convinced them it can’t be relied on. Others accept everything it produces without checking, right up until it’s confidently wrong about something that matters. Both patterns are well documented in the research on automation bias, and both are expensive: one wastes the tool, the other creates risk.

The habit never forms

Using AI means interrupting a workflow someone has run, comfortably, for years. That switch has a cost every single time, and if the payoff isn’t immediate and obvious, the old way wins. A mandate can’t fix that. The design has to, by meeting people inside the work they already do instead of in a separate window they have to remember to open.

Judgment feels threatened

Skilled professionals hear “AI will make you more efficient” as “AI will make you less necessary.” Unless the rollout is explicit about where human judgment stays in the loop, resistance is rational. People are defending the part of the job that makes them valuable.

Caution is not the enemy

In my research with accounting firms, a profession famous for its skepticism, the most useful reframe was to treat professional caution as a strength to build on rather than a lag to overcome. The firms that adopt AI well aren’t the fastest. They’re the ones that moved deliberately, matching their usage to their trust and closing the gap in both directions.

That’s the thinking behind the AI Trust & Usage Quadrant, a framework from my published adoption research. Map where people sit on two axes, how much they use AI and how much they trust it, and you get four very different starting points. The over-trusting experimenter needs guardrails. The capable skeptic needs a low-stakes win. Treat them identically and the rollout produces risk and resentment at the same time.

What works

Nothing here requires a bigger model. It requires designing for the humans:

The organizations getting real value from AI treat adoption as what it is: a behavior-change problem, with decades of science behind it, sitting right where most rollouts never look.