Two failure modes dominate AI-augmented decision tools.
The calculator failure produces a clean output the user cannot defend. A meeting opens with "the model says we should ship feature X" and immediately stalls — nobody on the team can reconstruct the assumptions behind the score. The output cannot survive the discussion that follows.
The vanilla failure produces an output the user can defend, but the inputs are still being guessed without scaffolding. Confidence ratings cluster at 80% by default; the dimension meant to reduce hubris reinforces it. The output is defensible but no better than what the team would have produced on their own.
The middle path: AI as thinking aid, not calculator. The AI sits inside the user's reasoning — asking the questions a senior colleague would ask, surfacing comparable examples, checking overconfident inputs — but the inputs and the score belong to the user.
Three design principles for tools in this register:
- Each input gets a dedicated prompting layer. Generic "tell me your assumptions" prompts get skipped. Structural questions about the specific dimension — segment definition, behavioral change, evidence quality — get answered.
- The output reconstructs the reasoning. Show the inputs and the assumptions behind them, not just the final number. A discussion can rebuild the argument from the artifact alone — no one has to ask "where did this come from?"
- Strategic narrative, separate from scoring. Generated independently from the math, so the math stays auditable. The narrative explains what the score means in context; the math is reconstructable line by line.
Calculator AIs are faster. Thinking-aid AIs produce decisions that hold up after the demo.