When should a founder not use AI to make a decision?

When Founders Should Keep AI Out of the Call

Keep AI out of one-way-door decisions, value-laden calls, and choices where you cannot verify the inputs. Bezos's reversible/irreversible test is the filter: use AI to widen options on two-way doors, but own irreversible bets yourself. Wharton's 2026 work found 73% of people surrender their judgment to a confident model rather than check it.

I treat AI like a brilliant, fast, and slightly unreliable advisor who has never once felt the consequences of being wrong. That framing decides when I let it near a decision and when I don't. The cleanest filter I know comes from Jeff Bezos's 2016 shareholder letter: separate one-way doors from two-way doors. A two-way door is reversible. If the call is wrong, you walk back through it cheaply. A one-way door is irreversible. Once you step through, the cost of returning is enormous. I let AI run wild on two-way doors and I keep it firmly on the other side of one-way ones.

Why the asymmetry? Because models like Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro are extraordinary at generating options, surfacing considerations I missed, and pressure-testing my logic. That is exactly what a reversible decision needs: more shots on goal at low stakes. Naval Ravikant, in the Almanack, argues that judgment, the ability to know which problems are worth solving, is the rare and compounding skill. AI can hand you a hundred plausible problems. It cannot tell you which one is yours. On irreversible bets, naming a co-founder, taking the wrong term sheet, betting the roadmap on a single architecture, that judgment is the whole job, and outsourcing it quietly hollows out the muscle you most need to keep.

The second category I keep human is anything value-laden. Whether to lay off a team to extend runway, whether to take money from a particular investor, how to handle a partner who is underperforming but loyal, these are not optimization problems. Paul Bloom's work on psychology and morality is a useful reminder that our moral intuitions are messy, contextual, and bound up with empathy a model only imitates. When I ask an AI a values question, it will produce a fluent, balanced, confident answer. The confidence is the trap. It has no skin in the outcome and no relationship with the people involved.

The third case is the one founders underrate: decisions where you cannot verify the inputs. A Stanford HAI study found general-purpose models hallucinated on 58 to 82 percent of legal research queries, and even retrieval-grounded legal tools were wrong more than 17 percent of the time. A 2024 Deloitte survey found 38 percent of executives admitted to making a wrong call on hallucinated output. If you cannot check the load-bearing fact yourself, you are not using AI to decide, you are gambling on its fluency. Kahneman's Thinking, Fast and Slow names the mechanism precisely. A confident answer recruits your fast, intuitive System 1 and quietly switches off the slow, effortful System 2 you would normally use to scrutinize it. Wharton researchers put a number on this in 2026: on flawed AI responses, people "surrendered" their judgment rather than checking it at a nearly four-to-one rate, 73 percent versus 19 percent. That is automation bias, and founders are not immune. We are arguably worse, because we are busy and predisposed to delegate.

As a coach, I see a fourth situation that has nothing to do with the model's accuracy. Sometimes the point of a decision is that you metabolize it. John Whitmore's GROW model, the backbone of executive coaching, works because the person arrives at the answer themselves and therefore owns it. If I hand a founder a finished recommendation, I have robbed them of the reasoning that makes them act with conviction and adjust intelligently when reality pushes back. An AI that gives you the conclusion does the same damage. This is also why I think learning decisions deserve protection: the struggle is the product. I have written about why exploration beats premature optimization, and a model trained to predict the consensus next token is structurally biased toward the obvious path, not the original one.

So the practical rule I give clients is a two-question gate before you let AI make any call. First: is this reversible? If yes, use the model aggressively to widen and stress-test your options, then decide. Second: can I independently verify the facts it is relying on, and is this mine to own, judgment, values, or relationships? If the answer to that is no, the AI's role ends at framing the question. You still walk through the door yourself, with your name on it. Used that way, AI does not make you a worse decider. It clears the easy decisions so your scarce, fallible, irreplaceable judgment is spent only where it actually counts.