The AI Leadership Gap That AI Is Making Worse: Strategy, Not Tools

The Evidence

Two studies, plus a caveat I’ll get to.

In February 2025, researchers from Microsoft Research and Carnegie Mellon published a survey of 319 knowledge workers describing 936 real, first-hand uses of generative AI at work. The headline finding is uncomfortable and specific: higher confidence in the AI tool predicted less critical thinking. Higher confidence in one’s own expertise predicted more.

Read that again, because the direction matters. It isn’t that AI users think less across the board. It’s that trust in the tool substitutes for scrutiny, and trust grows with familiarity. The more comfortable you get, the less you check. Which is precisely backwards from what every enterprise AI policy assumes.

The researchers also found the nature of the work shifting: from gathering information to verifying it, from solving problems to integrating someone else’s solution, from doing the task to supervising it. They describe a move toward passive oversight.

Then in June 2025, MIT Media Lab published “Your Brain on ChatGPT,” which put EEG caps on 54 participants writing essays under three conditions: LLM, search engine, and unaided. Brain connectivity scaled inversely with the amount of external support. The unaided group showed the strongest, most distributed networks. Search users came in the middle. LLM users showed the weakest coupling. In a fourth session, participants who’d been using the LLM and were then asked to work unaided underperformed. The authors called what accumulates “cognitive debt.”

The caveat, because truth first: these are early studies. The Microsoft/CMU work relies on self-reported effort, which is a soft measure. The MIT study is small, young, and focused on essay writing in an academic setting, not on executives making capital allocation decisions. Neither proves that AI use degrades leadership judgment.

But they point the same direction, and they point at a mechanism that anyone who’s run a training organization already recognizes: capability that isn’t practiced doesn’t hold. That isn’t a new finding about AI. It’s an old finding about humans, showing up in a new context. A National Bureau of Economic Research study of call center agents using AI assistants found newer workers gained the most immediate productivity — and showed the least skill growth over time. The easier the tool makes the work, the smaller the developmental footprint it leaves behind.

Now Connect It to the Thing That’s Actually Failing

MIT’s Project NANDA studied enterprise generative AI adoption in 2025: 300-plus publicly disclosed initiatives, interviews across 52 organizations, surveys of senior leaders. The finding that made the rounds: roughly 95% of enterprise AI pilots produced no measurable impact on the P&L. Only about 5% were generating real value.

The authors were clear the failure wasn’t the models. It was integration, workflow, and misaligned priorities. Budget flowed toward visible functions like sales and marketing while the clearest returns showed up in unglamorous back-office work. Companies spent where the excitement was, not where the payoff was.

Gallup’s 2026 workplace data lands on the same conclusion from a different angle: AI isn’t the bottleneck. Human systems are. Their research found that when employees believe their manager actively supports the team’s use of AI, they’re 8.7 times more likely to say AI has meaningfully changed how much work gets done. And yet fewer than a third of U.S. employees in AI-implementing organizations strongly agree their manager does that.

So here’s the shape of the problem.

Ninety-five percent of AI initiatives are failing for reasons that have nothing to do with technology and everything to do with judgment: which problem to point it at, what “working” would look like, what to stop doing, what to protect, what to tell people. Every one of those is a hard-thinking question.

And the evidence suggests the tool is quietly eroding the appetite for exactly that kind of thinking, most severely in the people who trust it most — which, in most organizations, is the executives who’ve been loudest about adopting it.

That’s the gap. Not a tools gap. Not a literacy gap. A thinking gap, appearing right when the thinking gets hardest.

What A Leader Can Actually Do About It

I’m not going to tell anyone to use AI less. That’s both unrealistic and wrong. The productivity gains are real, and I use these tools constantly. The question isn’t whether. It’s where in the sequence.

Do the hard thinking first. Then bring in the tool. Form your own view of the problem before you ask for one. The Microsoft/CMU finding cuts both ways: self-confidence in your own expertise predicted more critical engagement. Arriving with a position makes you a better editor of what comes back.

Use it as an adversary, not an author. “Write our AI strategy” produces something fluent, generic, and unfalsifiable — you’ll like it, which is the problem. “Here’s my strategy. Find the three assumptions most likely to be wrong” produces something useful. Prompt for friction, not agreement.

Protect one decision a quarter from the tool entirely. Pick something consequential and think it all the way through unaided. Not out of nostalgia — as a diagnostic. If you can’t do it anymore, you’ve learned something worth knowing.

Ask the questions AI can’t answer for you. What are we willing to be worse at? Which jobs change, and what do we owe the people in them? Where is a wrong answer catastrophic rather than merely costly? What do we do when this fails publicly? These are judgment calls loaded with values and context. A model will generate plausible prose about all of them. None of that prose is a decision.

Make it a leadership development priority, not an IT rollout. The organizations in that 5% weren’t the ones with better models. They picked one real problem, integrated it into an actual workflow, and had leaders who owned the outcome. That’s a leadership capability, and it’s trainable.

The Short Version

We’re handing our leaders a tool that makes thinking feel unnecessary, at the exact moment we’re asking them to think harder than they have in a decade.

The gap isn’t between companies with AI and companies without it. It’s between leaders who use it to sharpen their judgment and leaders who use it to avoid exercising theirs.

That second group will be fast, articulate, and confidently wrong. And they will not see it coming, because the output will read beautifully the whole way down.