The cure for AI de-skilling is a test, not a timeout
What BCG found
In "When Everyone Uses AI, Companies Risk Losing Critical Skills" (BCG, June 17, 2026), Sagar Goel, David Martin, and Charikleia Kaffe report on a survey of 70 C-suite leaders and senior executives. Half already see de-skilling in their organizations. More than 60 percent expect it to become a material threat within three to five years. Judgment and decision making carries the highest de-skilling risk score of any skill they measured.
The finding that matters most sits a few paragraphs further down. Almost 90 percent of the leaders cited overreliance on AI outputs without stress testing or challenge. One of them put it this way: "Teams start treating AI-generated work as 'good enough' and skip the critical thinking step entirely. Over time, nobody is poking holes in the analysis, which means errors and blind spots slip through unchecked."
The authors call the result "cognitive debt," and they frame the company-wide version as "distributed de-skilling," a design flaw in how work is organized rather than a failing of individual employees. They are right about that.
Where the article goes soft
BCG offers six strategies. The first one says companies should define "how output verification should be conducted," and then moves on without saying how. The strategies that get the vivid examples are the ones about stepping away from the tool: AI-off zones, a bank's monthly AI-free Friday, a three-day no-AI hackathon, "AI failure drills" that show employees a hallucination so they learn to distrust the machine.
Set those next to the 90 percent figure. Nearly every leader surveyed said the problem is people accepting AI answers with no way to check them. An AI-free Friday does nothing about that. On Monday the same manager carries the same unexamined recommendation into the same meeting. Rationing also gets weaker every quarter, because AI is going to be inside every tool these teams touch, and a strategy that depends on people choosing not to use it has a short life.
Developers hit this wall first
Software developers were the first profession to have capable AI pointed at the center of their work, and they ran into BCG's problem immediately. A developer who pastes AI-generated code into production without checking it has handed over the judgment and kept only the typing. That is de-skilling in exactly BCG's sense.
The developers who got good at working with AI did not use it less. They leaned on a habit their field already had. Before you trust a piece of code, you write a test: a plain statement of what a correct result must do, written before you look at the result. Then you run the code against it. If it fails, you fix the instructions or the inputs and run it again. The test is the gate the output has to pass through, and writing it forces you to understand the problem well enough to say what "right" means.
None of this is technical. It is the habit of deciding how you will judge an answer before you see it.
What a test looks like for a management decision
Take a decision that lands on an operations leader's desk every year: which customer accounts get a price increase. You could ask an AI to go through your order history and produce a list. Most managers would read the list, recognize the accounts they expected to see, and move on. That is what the 90 percent looks like in practice.
Instead, write the tests first. Four would do to start:
Four tests for a price-increase list
- No account on the list has disputed pricing or lodged a service complaint in the last twelve months.
- No account on the list is mid-contract on fixed pricing.
- The combined margin lift from the list clears the target you set before you asked.
- For every account on the list, the sales rep who owns it can be shown the reason it is there in one sentence.
Now run the AI's list against the tests. Suppose it fails the second one because the contract data lives in a system it cannot see. You add that data, or you add the rule, and run it again. Suppose the fourth test turns up an account that is on the list only because it grew last year. You tighten the instruction and run it again. Three or four passes later you have a list you can defend, and you know why.
Why building the test is the antidote
Writing those four tests used every skill BCG says is at risk. You had to frame the problem, reason about cause and effect, and decide what a good solution has to satisfy. You cannot write a test for a decision you do not understand, which is why the exercise protects the understanding.
It also answers the other symptoms in BCG's survey. Junior staff learn by drafting tests and having a senior colleague tear them apart, which is mentoring with a concrete object in the middle. The tests pile up into a written record of how your company decides things, a kind of institutional memory that an AI-free session will never produce. And the tests catch the AI when it is wrong, which is the one moment when judgment gets exercised.
BCG's remedies protect a skill by resting it. Testing protects it by working it where the stakes are real. Only one of those gets stronger with practice.
Frequently Asked Questions
BCG uses "distributed de-skilling" for the gradual erosion of judgment, problem framing, and analytical skill across an organization when employees routinely accept AI outputs without challenging them. Its June 2026 survey of 70 senior executives found half already observing it, and more than 60 percent expecting it to become a material threat within three to five years.
Only when the AI's answers are accepted without a way to check them, which is what almost 90 percent of BCG's respondents reported seeing. A manager who writes down the criteria a good decision must meet before consulting AI, then tests the output against those criteria, exercises more judgment per decision, not less.
A short list of checks, written before you see the recommendation, that a correct answer must pass. For a pricing decision the checks might cover contract status, recent complaints, a margin target, and whether each line can be explained to the account owner. You run the AI's output against the list, correct what fails, and run it again until it passes.