Auditing Causal Claims with AI

The one idea: An AI assistant can list every reason a business claim might be wrong, but only you know which of those reasons actually applies to your company. Use it to generate suspects fast — you stay the judge.

⚠️ The trap: "I'll just ask the AI whether the claim is causal and go with its answer." That feels right and it's wrong. Ask it that and it will politely extend your framing; the AI's job is to generate suspects fast — confounders, rival DAGs, cheap tests — while you stay the judge who decides which ones actually apply to your business.

An AI assistant is the perfect tireless skeptic. It has read every confounding story ever written. It never gets bored on the eleventh alternative explanation, and never worries about offending the deck's author. It also has no idea which of those stories applies to your business. That division is the whole operating manual: breadth is the AI's; judgment is yours.

Used well, AI turns a thirty-second gut check into a five-minute pressure test. Used badly, it produces confident agreement with whatever framing you handed it. The difference is in the prompts: five patterns, a worked audit to copy, the anti-patterns that waste the exercise, and the rule for when to stop prompting and call a specialist.

The five prompt patterns

Pattern 1: Steelman-then-kill. Force the model to argue both sides at full strength.

Here's a claim from a strategy deck: [claim + how the data was gathered].
First, make the strongest possible case that this relationship is causal.
Then make the strongest possible case that it isn't — reverse causation,
confounders, selection. Argue…