Thinking in Bets by Annie Duke

Judge decisions by their quality, not their outcomes, and put a number on your beliefs.

Three takeaways for AI PMs

• Grade the decision, not the result: Every retro, launch review and quarterly recap tempts you to score choices by what happened next.
• Put a number on every belief: Vague confidence hides disagreement; a percentage exposes it.
• Treat eval scores and demos as samples, not verdicts: One strong eval run, one impressive demo and one happy pilot customer are all single draws from a distribution.

Do this Monday

• In your next prioritisation call, replace "I think this will work" with a number: "I'm at 70% this improves activation." Forcing the number surfaces the disagreement that vague confidence hides.
• Before your next AI feature launch, run a pre-mortem with kill criteria: "It's six weeks post-launch and this failed. Why?" Then commit, in writing, to the signals that would make you stop. Duke is blunt that a pre-mortem without kill criteria is theatre.
• When something ships and wins, hold the review anyway. Ask "did we get a good outcome, or did we make a good decision?" Write down one thing you'd repeat and one thing that was luck.