The Tool That Makes You Stop Thinking

AI decision making is making people worse at decisions. Not because the technology is bad, but because it is good enough to make thinking feel unnecessary. The person who should be deciding stops deciding: they paste a problem into a prompt, receive a confident answer, and act on it without examining a single assumption. The tool did not fail. The thinking did.

Roger Estall and I spent decades watching organisations adopt tools that promised to improve decisions. Risk registers, scoring matrices, enterprise risk management frameworks: each one arrived as an aid to thinking and was gradually promoted to a substitute for it. AI is following the same trajectory, with a shinier interface and a larger marketing budget. The output of any model, whether it runs on a spreadsheet or a neural network, is not the decision; it is information the Decider must weigh against purpose, context, and consequences that no algorithm can own.

AI decision making is the use of artificial intelligence to support, not replace, human judgement in consequential choices, by testing assumptions, structuring options, and challenging reasoning while leaving the decision with the person who owns the outcome.

The Model Is Not the Decision

Every model is a simplification built on assumptions, and the model's output is only as reliable as those assumptions are valid. That observation is Principle 14 of the Universal Decision-Making Method, and it was written about spreadsheets and actuarial tables long before anyone used the phrase "machine learning." The principle applies identically to AI decision making: a large language model processes statistical patterns in text; a predictive algorithm extracts trends from historical data.

Neither process includes the step where someone decides what matters, which trade-offs are acceptable, and what consequences they are prepared to own.

The market for AI decision maker tools reflects this gap: each processes inputs according to embedded assumptions, but none examines whether those assumptions are sound.

The question "can AI make decisions" contains its own answer: a decision is a commitment to act, made by someone who will own the outcome. No model, however capable, bears consequences.

The confusion between model output and decision has never been harmless. Before the 2007 global financial crisis, credit-ratings agencies relied on models assuming perpetual economic stability; the assumptions were never tested because the model's output looked authoritative enough to pass for a decision. That was not a failure of computing power but a failure of the thinking that should have surrounded it.

AI systems are more capable than a credit-risk spreadsheet, but capability does not change the structural problem. The model suggests; the Decider decides; and when that boundary dissolves, the consequences do not care whether the dissolving agent was an Excel formula or a neural network.

The same error appears wherever predictive analytics is treated as a decision engine rather than as one input among several. A forecast is a statement about probability; a decision is a commitment to action in the presence of uncertainty. The forecast can inform the decision, sharpen it, and help the Decider understand what they are betting on, but it cannot make the bet. When I hear someone say "the data told us to do X," I know that the data did nothing of the sort; a person decided, and they are attributing their judgement to a number because attribution feels safer than ownership. AI predictions are faster and sometimes more accurate, but they remain predictions, not decisions.

What a Useful AI Decision Tool Actually Does

A useful AI decision tool does not tell you what to decide; it helps you see what your decision depends on. That distinction separates tools that support judgement from tools that replace it, and it applies to everything from AI-assisted decision making to a purpose-built decision-making tool. The useful version asks questions: what is the purpose of this decision, what options are genuinely open, what assumptions carry the most weight? The harmful version skips those questions and hands you a recommendation decorated with a confidence score.

AI is replacing what consultants traditionally sold: the research and analysis. What it cannot replace is the decision those deliverables were supposed to inform.

The Walk is our implementation of this principle. It uses AI to structure a five-step decision process, not to produce an answer. The AI keeps the Decider moving through the method: framing the choice, developing real options, identifying assumptions, testing whether certainty is sufficient, and designing monitoring for after commitment. At no point does the system say "you should do X." That constraint is not a limitation; it is the entire point. An AI decision-making tool that tells you what to do is a consultant without accountability, and I have spent enough years watching that model fail to know where it leads.

Decision coaching has always worked this way: the coach returns the decision to the person who owns it. AI can do the same job if the system is built to ask rather than to answer. The moment the tool starts providing recommendations, the Decider's role shifts from thinking to approving, and approval is not judgement.

Take the AI recommendation sitting on your desk and name the three assumptions it depends on before you act on it. Start the Walk →

How to Use AI Without Outsourcing Judgement

Three uses of AI in decision making consistently produce value. First, assumption testing: describe your preferred option to an AI assistant and ask it to identify what has to be true for that option to work. A competent model will generate a list of dependencies you may not have articulated, and some of them will be assumptions you have been carrying without examination.

Second, option structuring: when a decision has become tangled, asking AI to separate the options into distinct paths often reveals that what appeared to be five options were really two options presented with different emphasis. Third, monitoring design: after a commitment, AI can help identify the signals that would indicate your key assumptions are failing, which matters because most organisations commit and then neglect to watch.

When organisations skip these steps, AI makes wrong business decisions not because the technology failed but because the process it was added to was already unsound.

I cover the practical detail of using AI to make decisions in the advice I give every client: where AI helps, where it does not, and where most teams unknowingly cross the line.

Three uses consistently destroy value. Ask-and-obey: pasting your problem into an AI system and acting on whatever it recommends. Consensus laundering: using AI output to justify a decision that has already been made, giving the conclusion an appearance of objectivity it does not deserve. Accountability diffusion: claiming that "the AI recommended it," which detaches the decision from the person who should own the consequences.

All three have the same structural defect: they use the tool to avoid the part of deciding that is genuinely difficult, which is taking responsibility for a judgement under uncertainty.

The structured decision process treats AI as one source of information among several, never as the decision itself. When software follows this principle, it gives the Decider a better basis for commitment; when software violates it, it gives the Decider a comfortable way to stop thinking entirely.

When the Tool Gets Promoted Beyond Its Capability

I have watched this pattern repeat across industries and decades. A tool arrives with a genuine use; organisations promote it beyond that use because promotion is cheaper than thinking; the tool fails at the job it was never designed to do; and the organisation blames the tool rather than the promotion.

The Australian Banking Royal Commission found that every major bank had risk committees, chief risk officers, risk appetite statements, and formal reporting structures. The full apparatus was present, but it had been promoted from a thinking aid to a substitute for institutional honesty. The commission uncovered conduct including charging customers for services never provided, and no amount of risk-management machinery had prevented it because the machinery was never designed to replace ethical judgement.

The real cases of AI in decision making follow the same pattern: from autonomous vehicle fatalities to sepsis prediction tools, the difference between success and failure is whether a human stayed in the decision.

AI decision making is now following the same trajectory. The tool is genuinely useful for pattern recognition, text analysis, and information retrieval; when organisations start treating its output as the decision rather than as input to the decision, they repeat the structural error that has produced failures from financial crises to engineering disasters.

The assumption that enough data eliminates the need for judgement was wrong when it was applied to big data, and it remains wrong when applied to AI. In Deciding, Roger Estall and I documented how tools promoted beyond their capability produced not merely ineffective decisions but actively harmful ones, because the illusion of rigour suppressed the questioning that would otherwise have caught the problem.

The pattern is not limited to financial institutions. When COVID-19 arrived, almost no organisation reached for its risk register, risk appetite statement, or scenario-planning matrix. They went straight to the decision, which is what people do when the situation is real and the tools they have built are ornamental. AI tools built on the same structural premises will meet the same fate when the pressure arrives.

ChatGPT, Claude, and the Decision You Still Own

The commercially available AI assistants are, at their best, thinking partners. I use these tools in my own work: to stress-test draft reasoning, to generate alternative framings of a problem, and to identify assumptions I have not yet articulated.

What I do not do is ask them what I should decide, because no language model has access to my purpose, my context, my tolerance for uncertainty, or my accountability. Those four elements determine whether a decision is right for the person making it, and they live outside the model's reach entirely. Sufficient certainty is personal; what is sufficient for one Decider might not be sufficient for another, and no algorithm can calculate that threshold on your behalf.

The calibration data is striking: ChatGPT for decision making produces false confidence because it states near-certainty whether right or wrong, and the people most influenced are those least confident in their own judgement.

The practical question is not whether to use these tools but how to use them without outsourcing the judgement they cannot carry. Prompt an AI with "what should I do about X?" and you will receive a plausible, confident answer that sounds like a decision but is not one.

Prompt it with "what are the three strongest objections to this course of action?" and you will receive something genuinely useful, because you have kept the deciding with yourself and used the tool for what it does well: generating structured challenge from a large information base. That is the difference between a decision partner and a decision substitute, and it is not a subtle one once you have seen the consequences of confusing them.

A practical test: after using any AI tool in a decision process, ask yourself whether you could explain your reasoning without referencing the AI's output. If the answer is yes, you have used the tool as a thinking partner; the reasoning is yours, and the AI sharpened it. If the answer is no, you have outsourced the judgement, and whatever confidence you feel about the decision is borrowed certainty. Borrowed certainty is indistinguishable from real certainty until the assumptions beneath it fail, at which point the person who borrowed it discovers they have no reasoning to fall back on.

If AI Is the Answer, What Was Your Question?

Roger Estall and I used to ask the same question about risk management: "If risk management is the answer, what was your question?" The response was usually genuine puzzlement, because the question had never been asked; the tool had arrived as a solution before anyone had defined the problem it was solving. AI decision making is heading toward the same dead end. Organisations are adopting AI tools for their decision processes without first asking what is wrong with their current process, or indeed whether they have a decision process at all. The technology is answering a question that most of its buyers have never bothered to articulate.

AI decision making ethics is not primarily a question of algorithmic bias, although bias is real and worth addressing. The deeper problem is abdication: the willingness of the person who should be deciding to hand the judgement to a system that cannot be held accountable for the consequences.

When a board accepts an AI recommendation without interrogating the assumptions beneath it, the board has not made a decision; it has outsourced one while retaining the appearance of governance. That is not a technology problem but a judgement problem, and it predates AI by centuries.

The technology merely makes the abdication easier to perform and harder to detect, which is why the accountability conversation deserves more attention than the conversation about cognitive biases alone. Bias distorts thinking, but abdication abandons it, and it is the abdication that AI makes trivially easy to perform.

---

Grant Purdy is the co-author, with Roger Estall, of Deciding (2020), and the architect of the Universal Decision-Making Method.