Stop asking "can AI do this?" Ask what a wrong answer costs
The capability question is almost always yes, eventually. The question that actually separates good AI product bets from bad ones is the price of being wrong.
On this page 2 sections
In almost every AI brainstorm, the first filter is capability: can the model do this? With current models the answer is usually “mostly”, which means the filter lets everything through and the roadmap fills up with demos.
A better first filter is the cost of a wrong answer.
Two axes, four kinds of bets#
Plot each idea on two questions: how often will it be wrong, and what does one wrong answer cost the user or the business?
| Cheap mistakes | Expensive mistakes | |
|---|---|---|
| Rarely wrong | Ship it, automate fully | Ship with review and audit trail |
| Often wrong | Ship as a suggestion, user picks | Don’t ship yet. Fix the data or narrow the scope |
Drafting a first-pass email lives top-left even when it’s often wrong, because a bad draft costs ten seconds. Telling a customer their loan eligibility lives on the right even at high accuracy, because one wrong answer is a complaint, a compliance issue, or a lost customer.
The interesting part is that the same idea can move between boxes depending on design. “Answer customer billing questions” is bottom-right as an autonomous bot. Reframed as “draft an answer with the source invoice line attached, for an agent to send”, it moves to the left column and becomes shippable this quarter.
The question to bring to your next review#
Next time someone pitches an AI feature, skip “can it do this?” and ask: when it’s wrong, who notices, how fast, and what does it cost them?
If nobody in the room can answer, that’s the work. Not the prompt.