An adoption score can't tell caution from erosion

There’s an employee in your org who uses the AI assistant every day for the small stuff. Drafting the email, cleaning up the notes, summarizing the doc. Then a real decision lands. Which bet to make, which candidate to hire, which number in the plan is wrong. And they close the tab and think it through themselves.

By the most talked-about framework in enterprise AI this month, that employee is the problem you need to solve.

The framework comes from The Decision Lab, a behavioral-science consultancy, and it has a name for what that employee is carrying: psychological debt. Their research puts numbers on it. Employees high in this “debt” score a 36 out of 100 on an AI-adoption scale, against a 60 for everyone else, even when they say out loud that AI is valuable. The reason, the firm reports, is that they confine AI to simple, low-risk tasks and keep it away from strategic decisions. Six psychological costs drive it: cognitive offloading (losing trust in your own thinking), reduced autonomy, diminished competence, weakened social connection, credibility loss, and identity threat. The pitch is clean. This debt caps how sophisticated your AI use can get across the org, so reduce the debt and the ceiling lifts.

It’s an easy story to buy, because it fits what most managers already believe about AI at work. The technology is ready. The drag is human reluctance. The holdouts who keep the tool at arm’s length are the bottleneck, and lowering their resistance is how value follows.

But look at what that 36 is actually counting. High “debt” is defined by low adoption, by employees who restrict AI to safe tasks and refuse to route the strategic calls through it. The firm has taken one specific caution, the refusal to hand your judgment to the model, measured it, and named it a disease. And the peer-reviewed evidence says that caution is the thing keeping your best people from getting quietly worse.

Start with what the number is made of. The whole construct hangs on an adoption score. High debt means low adoption: the employee uses AI less, and reserves it for tasks where being wrong is cheap. That’s the measurement. So “reduce psychological debt” and “get employees to push AI deeper into consequential decisions” are the same instruction written twice. The behavior the firm wants to dissolve is the behavior of keeping your own reasoning on the calls that matter.

It’s worth being precise about where this comes from. The Decision Lab is a commercial firm, and the product it’s describing is the debt-reduction framework itself. The finding rests on a single survey of 1,200 employees, with no peer review. A companion claim floating in the same neighborhood, that 95% of enterprise AI projects return nothing, gets pinned on MIT Media Lab, but the primary source is missing, so treat it as unverified. None of that makes the pattern fake. But notice the shape of the incentive. A firm whose measure of health is adoption, selling a program to raise adoption, has every reason to file restraint as pathology. Restraint is the one thing standing between its number and the ceiling.

Give the framework its due, because part of it is real. Cognitive offloading is a genuine harm. Lean on the model long enough and you lose the thread of your own thinking. If an employee has quietly stopped reasoning and now avoids AI out of vague anxiety rather than judgment, pushing them back in is the right call. Some fraction of that low-adoption group is exactly that: people who’d be better off using AI more, on more things.

Here’s what the framework can’t see. There’s a peer-reviewed study that measured metacognitive accuracy, meaning how well people judge whether their own answers are actually right. Fernandes and colleagues, writing in Computers in Human Behavior this year, ran it on 246 people and replicated it on 452. The finding runs against the intuition. Higher AI literacy correlated with lower accuracy. The people most fluent and comfortable with AI grew more confident in their self-assessments while getting less correct. Confidence went up. Calibration went down.

Now lay the two profiles on top of each other. The employee The Decision Lab wants you to manufacture has low debt, an adoption score near 60, comfort routing strategic decisions through AI, no competence anxiety, no identity threat. Trait for trait, that’s the profile the peer-reviewed study flags as miscalibrated. Low debt is high confidence in delegated reasoning. And high confidence in delegated reasoning is the exact thing that predicts getting worse at knowing when you’re wrong. The program pushes your workforce toward the profile the stronger evidence ties to getting worse without noticing it.

Weigh the two sources against each other for a second. A single vendor survey points one way. A replicated, peer-reviewed study points the other. A cheap survey pointing the opposite way from a replicated study is worse than weak. It’s pointing you backwards.

Which is the thing to sort out before you fund any of this. The vendor’s number can’t tell two opposite employees apart. Of everyone restricting AI to low-risk work, some are eroding: genuinely offloaded, lost the thread, better served using AI more. Others are discriminating: correctly reserving judgment for the calls where the calibration gap would bite. One should trust the model more. The other the model would make overconfident. Both show up as a 36. A single adoption score adds them together and reports the sum as a ceiling to clear.

So before you raise the friction or lower it, before you push AI into the strategic decisions or pull it back out, answer the question the survey skipped. On this decision, for this person, is holding back the debt or the safeguard? Until you can tell those two apart, “raise the adoption score” is a plan to fix your most careful employees for being careful.