The most dangerous AI skill isn't prompting. It's knowing when not to trust the answer.

The first twelve weeks of this series explored how AI is changing work, leadership and competitive advantage. This week starts the second half of the journey. And I want to begin with a question.

What is the most important AI skill for a business leader?

Many people would say prompting. Others would say automation. Some would say learning to use AI agents. I think they're all looking in the wrong place.

The most important AI skill isn't getting better answers. It's recognising when an answer shouldn't be trusted.

Finding an answer is becoming cheap. Evaluating the answer is becoming valuable.

AI is remarkably good. Until it isn't.

One of the most important AI studies published to date came from researchers at Harvard Business School and Boston Consulting Group. They gave 758 consultants access to GPT-4 and asked them to complete a series of realistic consulting tasks.

The results were extraordinary. On tasks that fell within the model's capability frontier, participants completed 12.2% more work, completed it 25.1% faster, and produced significantly higher-quality output. Lower-performing consultants improved 43%. Higher performers improved 17%.

Then the researchers changed one thing. They introduced a task that deliberately sat outside the model's capability frontier. Everything changed.

Consultants using AI were 19 percentage points less likely to reach the correct answer than consultants working without AI.

Read that again. The same people. Using the same tool. On the same day. The difference wasn't intelligence. It was the type of problem.

The most surprising finding wasn't that AI got it wrong

It was that people didn't reliably know when it was wrong. That is why the paper is called Navigating the Jagged Technological Frontier. The frontier isn't a line you can see. It shifts with the problem.

Inside the frontier, AI can dramatically improve performance. Outside it, the same confidence can become a liability.

The dangerous part is not that AI makes mistakes. Humans have always made mistakes. The difference is that AI often makes them with remarkable confidence. The dangerous part is believing you know which answers deserve your confidence when you don't.

Confidence is not competence

Some will say experienced people can spot the difference. They know their domain. Surely they can tell when AI has gone wrong.

Sometimes they can. Often they can't. And here is why.

Human judgement was trained on human failure. We learned to read the signs. The pause before an uncertain answer. The hedge. The shift in tone. The visible strain when someone reaches the edge of what they know. Those signals tell us when to lean in and check.

AI removes many of the cues we've learned to rely on. It fails without hesitation. It is wrong in the same fluent, confident voice it uses when it is right. The tells we spent careers learning to read simply aren't there.

Daniel Kahneman spent a career studying this. His conclusion was uncomfortable. Confidence is a feeling. It is not a measure of accuracy. We trust the answer that sounds coherent, not the answer that happens to be correct.

AI produces coherence on demand. That is what makes it so useful. It is also what makes it so easy to over-trust.

So the danger isn't that leaders lack judgement. Many have excellent judgement. The danger is that judgement about people does not automatically transfer to judgement about AI. They are different skills. One reads human uncertainty. The other has to detect a failure that shows almost none of the signals we've learned to recognise.

This changes the role of leadership

For years, leaders have been rewarded for finding better answers. AI increasingly provides those.

The leadership challenge is changing. It is becoming less about generating answers, and more about deciding which answers deserve further scrutiny. Which assumptions should be tested. Which conclusions need independent verification. Which decisions carry enough uncertainty that human judgement must take over.

That is not a prompting skill. It is a leadership skill.

AI changes the economics of certainty

Before AI, uncertainty often came from having too little information. Today, uncertainty increasingly comes from having too many plausible answers.

The bottleneck has shifted. Finding an answer is becoming cheap. Evaluating the answer is becoming valuable.

That is a very different problem. And it is one many organisations have not yet recognised.

Judgement has a new job

Week four argued that the AI skill gap is not prompting. It is judgement. I still believe that. But months of researching this series have sharpened the point.

Calibration is not simply judgement pointed at AI. It is a distinct capability. General judgement asks whether a decision is sound. Calibration asks a narrower, harder question: how much should I trust this particular answer, on this particular problem, right now?

Philip Tetlock found the same thing in a different field. The best forecasters were not the smartest people in the room. They were not the greatest experts. They were the best calibrated. They knew the difference between what they knew and what they merely felt sure about.

That is the skill AI now demands of every leader.

And here is the part most people miss. It gets harder as AI gets better.

When a system is wrong often, you stay alert. You check. When a system is right ninety-nine times, you stop checking on the hundredth. Reliability breeds complacency.

As AI becomes more capable, its failures become less frequent. The risk is that we become less likely to notice the occasions when it is wrong. Better models do not remove the calibration problem. They make it easier to forget that it exists.

A simple calibration habit

One practical habit has stayed with me while researching this series. Before acting on an important AI-generated recommendation, pause and ask three questions:

• Is this the kind of problem AI is generally good at solving, or have I pushed it beyond its likely capability?
• Does this answer sound more confident than the evidence behind it actually justifies?
• How would I verify this if AI hadn't produced the answer?

Calibration is not about distrusting AI. It is about knowing when verification creates more value than speed.

The bottom line

Every major technology changes the skills that matter most. AI is no different. Prompting will improve. Models will improve. Agents will improve.

The organisations that consistently make better decisions will not simply use better AI. They will develop better calibration. They will understand where AI is exceptionally strong, where it is uncertain, and where human judgement still creates the greatest value.

Because in an AI-enabled world, the competitive advantage isn't getting answers faster. It's knowing when confidence should give way to verification.

So here is the question I will leave you with. When your AI gives you an answer that sounds completely convincing, how do you know whether you've just reached the edge of its frontier?

---

Sources

Dell'Acqua, F., McFowland III, E., Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F. & Lakhani, K. R. (2026). "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality." Organization Science, 37(2), 403–423. doi.org/10.1287/orsc.2025.21838

Kahneman, D. (2011). Thinking, Fast and Slow.

Tetlock, P. E. & Gardner, D. (2015). Superforecasting: The Art and Science of Prediction.