How a Leader Should Use AI Before a Consequential Decision

Most executives, when they bring AI to a real decision, ask it the wrong question. They describe the situation and ask what they should do. The model answers fluently, confidently, and usually in agreement with the view the question already implied. The leader reads it as confirmation and moves on.

This is the trap, and it is a measurable one. When a user states an opinion, large language models raise their agreement with incorrect beliefs sharply, on average 63.7 per cent across models and as high as 95.1 per cent in the worst cases (Wang et al., 2025). The model is not built to tell you that you are wrong. It is built, through its training, to be agreeable. So the more clearly you signal the answer you want, the more reliably you will get it back, dressed as analysis.

The danger is not that AI gives you a wrong answer. It is that it gives you a confident version of the answer you already had, and removes the friction that might have caught the error. That friction is the point of a good decision process, and it is exactly what an eager model dissolves.

There is a second, slower cost. The more a leader leans on AI to do the thinking, the less thinking the leader does. A study of knowledge workers by Microsoft Research and Carnegie Mellon found that higher confidence in generative AI is associated with less critical thinking, while higher confidence in one's own judgement is associated with more (Lee et al., CHI 2025). A separate study of 666 people found a significant negative correlation between frequent AI use and critical-thinking scores, mediated by cognitive offloading, the habit of handing the mental work to the tool (Gerlich, 2025). Use AI as the decider often enough and the muscle you most need in the chair, judgement, quietly weakens.

An AI model knows the world in general. It does not know your business in particular. Ask it a strategy question cold and it will give you the average of everything it has read, which is to say a competent, generic answer that could apply to any company in your sector. That is not a model failure. It is a context failure, and it is yours to fix.

This is why grounding matters. Retrieval-augmented generation, the technique of anchoring a model in specific, verified documents rather than its general memory, exists precisely because ungrounded output is unreliable for anything consequential (Lewis et al., 2020). Gartner puts the organisational version of this bluntly: it predicts that 60 per cent of AI projects will be abandoned by organisations that lack the AI-ready data to support them (Gartner, February 2025). Without your real numbers, your real constraints and your real history in front of it, the model is guessing, articulately.

We call the principle behind this the Mirror Principle: if the output is generic, the reasoning was generic. The fix is not a cleverer prompt. It is feeding the model the ground truth only your organisation holds, the actual cost base, the real pipeline, the history of what has and has not worked here, and then asking it to reason from that rather than from the world's average. A decision built on the model's general knowledge is a decision built on everyone's information and no one's. A decision built on your ground truth is yours.

There is a reason to be careful, and it is the strongest evidence in favour of keeping the human in the chair. The best-known controlled study of AI on knowledge work found that on tasks inside the model's competence, professionals using AI produced work more than 40 per cent higher in quality, but on a task deliberately chosen to sit outside the model's competence, AI users were 19 percentage points less likely to reach the correct answer (Dell'Acqua et al., Organization Science 2025, from a 2023 experiment). The researchers called this the jagged frontier: AI is brilliant on one side of a line and quietly wrong on the other, and the line is not marked. Knowing which side of it a given decision sits on is judgement, and it is the leader's job, not the model's.

This is why speed is the wrong measure. The pressure runs the other way: BCG's 2026 survey of CEOs and boards found that 61 per cent of CEOs say their boards are pushing AI too fast, and more than half say AI hype is distorting boardroom judgement (BCG, May 2026). When 72 per cent of CEOs now name themselves the main decision-maker on AI (BCG, January 2026), the temptation to let a fast, fluent model stand in for a slow, hard decision is exactly the failure mode to resist.

The discipline that protects against it is small. A randomised trial found that a three-minute habit, form your own judgement first, then verify the AI's output against independent sources, improved decision quality by nearly eight percentage points (Aydin, MIT Sloan Management Review, June 2026). Three minutes of thinking first, before the model speaks, measurably changes the decision. That is the entire argument in one statistic.

Using AI to confirm what you think is the default, and it is a trap. Using AI to find where your thinking is fragile is a discipline, and it is learnable. We call it the Havruta Methodology, after the oldest form of rigorous study we have: two minds, one question, an argument that ends in a sharper version of both sides. Applied to AI, it turns the model from a vending machine that dispenses answers into a thinking partner that earns them.

Before your next consequential decision, do not ask the machine what you should do. Ground it in your own truth, and ask it where you might be wrong. The quality of the decision will tell you the difference.

If this resonates, two companion essays go deeper: on why most enterprise AI investments deliver zero return, and on the gap between owning the tools and changing the work in the AI-readiness illusion.