How CEOs Use AI for Decision Making
How do CEOs use AI for decision making? The ones who get real value use it as a structured thinking partner that challenges the decision, not as an oracle that hands one back. Two moves do the work. First, ground the model in the organisation's own verified data, so it reasons from your specifics rather than from the average of everything it has read. Second, flip it: instruct the model to argue against the recommendation, name the assumptions it rests on, and say what would have to be true for it to be wrong. Used as an adversary and challenger, AI improves the quality of a decision; used to offload the thinking, it degrades judgement, because a model's first instinct is to agree. On consequential calls the accountable human stays the decider, with AI kept advisory and a human in the loop. This is the Flip, applied to the 4-Lines the Havruta Methodology installs.
AI has moved to the centre of the decision
A few years ago AI sat at the edge of executive work, a productivity aid the function below the leader used to draft and summarise. It does not sit there now. In an independent survey of the C-suite, about 43 per cent of leaders named AI and technology the single top investment priority for 2026, ahead of every other call on the budget (The Conference Board, 2026). Ownership has risen with the stakes: the decision about how AI is used is increasingly made at the very top, by the chief executive, not delegated down.
Adoption is no longer the question either. The canonical independent index reports that organisational AI adoption is near-universal, with the overwhelming majority of organisations now using AI somewhere in the business (Stanford HAI, 2026). So the interesting question is no longer whether a CEO uses AI. It is how. Two leaders with the same model in front of them, asking it about the same decision, can get opposite outcomes, and the difference is method, not technology.
The move that separates good use from bad
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 use that degrades judgement: AI as a vending machine, dispensing an answer you then approve.
The leaders who get value do the opposite. They cast the model as an adversary. Instead of asking it to confirm the recommendation, they instruct it to attack it, to name the assumptions the call rests on and argue the strongest case that it is wrong. This is not a rhetorical flourish, it is the most evidence-backed move in the field. When researchers set up a conversational agent as a devil's advocate inside a group decision, a challenger rather than an answer machine, the groups it argued with reached higher-quality and more inclusive decisions, resisting the social pull towards a premature consensus (Lee et al., CHI 2025). That is the empirical mirror of the move we call the Flip: the AI questions the human before it answers.
The contrast is stark enough to hold in one frame.
Vending machine
Ask it for the answer
You describe the decision and ask what to do. The model agrees with the view your question implied and dresses it as analysis. The friction that catches an error is gone. Degrades judgement.
Thinking partner
Make it argue against you
You ground it in your own data, then instruct it to attack the recommendation, name its assumptions and find the gap. It pushes back in time to matter. Improves decision quality.
The decision loop in practice
What does the thinking-partner stance actually look like in the chair? It is a short loop the leader runs before the decision, not after. The leader poses the decision; the AI challenges it, the Flip; the leader supplies the Ground Truth the challenge exposes as missing; and a sharper decision comes out the other side. One checkpoint never moves: the named, accountable human stays in the loop and owns the call.
None of the steps asks the model to decide. Every one makes it do what the people around a senior leader too rarely do: push back hard, in time to matter. Run often enough, the loop compresses the decision cycle and sharpens it at the same time, the effect we name Decision Velocity.
Keep a human in the loop
A challenger is powerful precisely because it argues hard, which is also why the leader, not the model, has to remain the one who decides. This is not a matter of taste. The leading inter-governmental standard, the OECD AI Principles, requires human agency and human oversight as a condition of trustworthy AI, a defined human in the loop on consequential calls (OECD, 2024). For an executive that translates into a simple rule: on a decision that carries real consequence, AI stays advisory and adversarial. It tests the call, it does not take it.
The practical discipline is to separate two jobs that look similar and are not. The model's job is to challenge, ground and stress-test. The leader's job is to weigh the challenge and own the outcome. Collapse the two, let the fluent answer stand in for the hard decision, and you have handed accountability to a system that cannot hold it. Keep them apart, and AI becomes the most demanding adviser in the room without ever becoming the decider. For the contrast in full, see vending machine versus thinking partner.
Why most executive AI changes nothing
If the method is the whole game, the failures should look like method failures rather than technology failures. They do. The most-cited study of the year, from MIT NANDA, found that about 95 per cent of enterprise generative-AI initiatives showed no measurable return (MIT NANDA, 2025). The interesting part is the cause. The authors traced the divide not to model quality and not to a shortage of talent, but to a learning, integration and adaptation gap: organisations bought the tools and never changed how the work, or the decision, actually got made.
This is the resolution of an apparent paradox. Adoption is near-universal (Stanford HAI, 2026), yet most of it returns nothing, because adoption and method are different things. Putting a model on every desk changes the toolkit. It does not, on its own, change a single decision. The right test of AI at the executive level is therefore not "did it give me something" but "did it change what I was about to do". Most of the time the honest answer is no, and that is a leadership problem, the kind that lives in how a team reasons, explored further in High-Speed Waste.
The method has a name
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.
The mechanics are compact. Every consequential interaction opens with the 4-Lines, Persona, Goal, the Flip and Sequence, the universal opening that sets the model up as an adversary rather than an oracle. The Flip is the line that does the work. Ground Truth is what keeps the challenge honest. Decision Velocity is what you get back: the decision cycle compressed threefold to fivefold, and sharper, not merely faster.
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 companion essay on how a leader should use AI before a consequential decision works the same loop in finer detail. The fastest way to feel the difference is to run one disciplined conversation against a live decision and watch the call change.
Do not ask the machine what you should do. Ground it in your own truth, and ask it where you might be wrong.
Frequently asked questions
How do CEOs use AI for decision making?
The CEOs who get value use AI to improve the quality of a decision, not to make it. They ground the model in the company's own verified data, then instruct it to challenge the recommendation, name its assumptions and argue the opposite case. AI used as a structured challenger sharpens judgement; AI used as an oracle that hands back an answer degrades it, because a model's first instinct is to agree. The accountable human stays the decider and keeps a human in the loop on consequential calls.
Does AI improve decision quality or just speed?
Both are possible, but they pull in opposite directions. Used as a challenger, AI raises decision quality: in controlled work, an AI cast as a devil's advocate improved the quality and inclusiveness of group decisions (Lee et al., CHI 2025). Used to offload the thinking, AI mainly raises speed while quietly weakening judgement. Decision Velocity, the compression of a decision cycle, is only worth having when the decision that comes out is sharper, not merely faster.
Should AI make business decisions on its own?
For consequential decisions, no. The leading inter-governmental standard, the OECD AI Principles, requires human agency and human oversight, a defined human in the loop (OECD, 2024). The evidence reinforces it: most enterprise AI fails to move outcomes because it produces output without changing how decisions are made (MIT NANDA, 2025). Keep AI advisory and adversarial on the calls that matter, and keep the named, accountable human as the one who decides.
How do you stop AI from just agreeing with you?
Instruct it to disagree. The move that separates good executive AI use from bad is the Flip: you tell the model to question and challenge you before it answers. Ask it to list the assumptions the recommendation depends on, make the strongest case that the decision is wrong, and name the one fact that would most change it. An AI set up as a structured adversary surfaces the counter-case the agreeable first answer hid (Lee et al., CHI 2025).
Why does most executive AI fail to change decisions?
Because it is a leadership and method problem, not a technology one. MIT NANDA found that about 95 per cent of enterprise generative-AI initiatives showed no return, and traced the cause to a learning, integration and adaptation gap rather than to models or talent (MIT NANDA, 2025). Organisational adoption is near-universal and the overwhelming majority of organisations now use AI (Stanford HAI, 2026), yet adoption without method changes nothing. The decision process has to change, not just the toolkit.
What is the first step a CEO should take with AI for decisions?
Ground it. Give the model your real numbers, constraints and history to reason from, rather than letting it answer from its general knowledge. We call this Ground Truth. An ungrounded model returns the average of everything it has read, a competent generic answer that fits any company in your sector and yours in particular none too well. Once the model is anchored in what only your organisation holds, you can flip it and put it to work as a challenger.
References
- The Conference Board. "AI and the C-Suite: Implications for CEO Strategy in 2026." The Conference Board 2026 C-Suite Outlook Survey, 2026.
- Stanford Institute for Human-Centered AI. "The 2026 AI Index Report." Stanford HAI, 2026.
- Lee, S., Hwang, S., Kim, D., Lee, K. "Conversational Agents as Catalysts for Critical Thinking: Challenging Social Influence in Group Decision-making." CHI, 2025.
- OECD. "OECD AI Principles: Human-Centred Values and Human Oversight." OECD, 2024.
- MIT NANDA (Challapally, A., Pease, C., Raskar, R., Chari, P.). "The GenAI Divide: State of AI in Business 2025." MIT NANDA, 2025.