Not the routine ones. Fifty years of skill-retention research points the other way: cognitive, accuracy-dependent, open-ended work decays faster under non-use than procedural or physical work — which means judgment goes before execution. And it goes quietly, because the system that replaced the practice also absorbs the evidence that the practice has stopped.
Most organizations assume the sequence runs the other way: the tool takes the mechanical work, the human keeps the thinking. The retention literature says thinking is the more perishable of the two, and that it perishes without producing a signal.
Which capabilities fade first, and why that order?
Three properties predict how fast a capability degrades once it stops being exercised. Arthur and colleagues pooled 189 effect sizes from 53 studies of trained skills after periods of non-use. Loss was negligible immediately after training and large — an effect size near −1.4 — after a year without practice. The moderators matter more than the range. Physical, naturalistic, and speed-based tasks proved durable. Cognitive, accuracy-based tasks decayed further. Open-loop tasks — those with no immediate signal of whether the performance was correct — decayed more than closed-loop tasks, the opposite of what the authors predicted.
Read together, those moderators describe one kind of work: cognitive, judged on correctness rather than speed, with no built-in feedback telling you at the time whether you got it right. That is professional judgment. Framing an ambiguous problem, deciding which evidence is load-bearing, choosing what not to do — all of it sits in the fastest-decaying quadrant.
A 2026 survey of C-suite and senior executives corroborates that ordering from a different direction: roughly half believe they are already seeing de-skilling, and they named judgment and decision-making, problem framing, and creative thinking as most at risk. The convergence is worth something — but the meta-analysis measured decay, while the survey recorded belief about decay. That difference is the subject of the next section.
Why does the decline stay invisible while it happens?
Because the ordinary evidence of declining competence is exactly what an assistant is designed to absorb.
Under normal conditions, a weakening capability announces itself. Someone hesitates. A draft comes back thin. An analysis misses the obvious objection. These are not failures of the system; they are its instrumentation.
Insert a capable assistant upstream and every one of those signals is intercepted before it reaches an observer. The hesitation is resolved in the tool. The thin draft is thickened. The missed objection is supplied. Output arrives at the same quality or better, on a shorter cycle, and nothing in the record indicates that the capability behind it has changed — because the record only ever contained the output.
An organization can therefore hold an accurate view of its performance and a badly inaccurate view of its capability at the same time. The strategic consequence of that gap — and what to measure instead of assisted output — is a separate argument. The mechanism is what matters here: the decline is not concealed by anyone. It is consumed at the point of origin.
How fast does this actually happen?
Faster than the review cycles built to catch it, on the one direct measurement available.
Budzyń and colleagues examined four Polish endoscopy centers that had introduced AI polyp detection, comparing colonoscopies performed without AI in the three months before and after. Across 19 endoscopists, each with more than 2,000 procedures behind them, the unassisted adenoma detection rate fell from 28.4% to 22.4% — a 6.0 percentage-point absolute decline, significant at p = 0.0089.
Three months. Thousands of repetitions each. A measurable drop.
The limitations need stating plainly, because one study is carrying weight here. It is retrospective and observational, not randomized on exposure, covering one clinical task in one country within one trial network. The capability is perceptual and diagnostic, not strategic judgment — and while the retention literature above would predict faster decay for the latter, that is an inference, not a finding.
Treat the number as an existence proof rather than a rate. What it establishes is that meaningful decay can occur inside a quarter, in expert hands, in a domain with objective measurement. An organization assuming its capability is stable across an annual planning cycle is assuming something this evidence does not support.
When does the loss finally become visible?
At handover — and handovers are rare, unplanned, and tend to coincide with the moment the system has already failed.
Endsley and Kiris established the pattern in 1995, studying operators of an automated navigation system. Deficits appeared specifically at manual takeover after an automation failure, and tracked with degraded situation awareness. The detail that matters for AI adoption is which layer degraded: operators kept accurate awareness of the raw data in front of them while losing comprehension of what it meant for their goals. They were not disengaged. They felt informed — and they were, at the level of facts rather than meaning.
That is the disclosure condition. Capability loss does not surface through gradual awareness, because confidence holds while comprehension erodes. It surfaces when someone must act unassisted — and in most organizations that happens only during an outage, a vendor change, or an incident review. The true state of capability is learned at the worst available moment, under adversarial conditions.
Which is why the Budzyń study is worth reading twice. The decline was detectable only because unassisted procedures were still being scheduled. Had every colonoscopy been AI-assisted, the drop would have been just as real and entirely unobservable.
Why is this harder to detect in some markets than others?
The obstacle differs by region, and none of the three versions is technical.
In the United States, substitution outpaces the review calendar: if capability shifts materially inside a quarter, an annual talent review samples at the wrong frequency and reports a stability that no longer exists. In the Gulf, AI deployment often runs alongside rapid headcount expansion, so aggregate output rises for reasons unrelated to per-person capability — growth is an effective mask, and the numbers are all genuinely good. In Central and Eastern Europe the constraint inverts: benches are thin after long cost discipline, few people ever work unassisted, and the sample that would reveal decay is never generated.
Three mechanisms, one outcome — the decline runs without an observer.
Frequently asked questions
Is the skill actually gone, or just dormant?
Decay is not deletion, and relearning is typically faster than original acquisition. The operational risk is not permanence but availability: a capability that takes six weeks to restore is unavailable during the incident that required it.
Doesn’t AI make some of these capabilities genuinely unnecessary?
Sometimes, and the distinction is verification cost. Where a wrong output is caught immediately and cheaply, decay in the underlying skill is a tolerable trade. Where errors surface late, expensively, or never, the same decay is a liability carried off the balance sheet. The mistake is applying one policy to both cases.
Isn’t three months too short for real skill loss?
It is shorter than most people expect, which is why it is worth reporting. One observational study in one perceptual domain is not a general clock and should not be quoted as one. The meta-analytic evidence supports the direction and the ordering; the speed in any specific setting remains an open question.
Who is more exposed — a senior who is forgetting, or a junior who never learned?
Different failure modes. Forgetting is recoverable and, under the right conditions, observable. Never acquiring produces someone with no internal reference for what competent unassisted work looks like, and therefore no basis for noticing the gap. Worth noting: the endoscopists above each had over 2,000 procedures of prior mastery. It did not prevent the decline.
