Most AI strategy work starts with capability: what can the model do? The better question is where. Every organization I've worked with has at least one decision quietly being handed to a model that was never built to make it — and at least one bounded, repeatable decision still eating up a person's time for no good reason. Getting this wrong in either direction is expensive. Getting it right is a test, not a guess.
01
Volume & Repetition
Is the decision high-volume and repeatable enough that a system gets meaningfully better with every pass — or is this closer to a one-of-a-kind call that won't repeat in quite this shape again?
02
Ground Truth
Does a clean, eventually knowable outcome exist? Will you find out, in a form a model could learn from, whether the call was right — or does it stay permanently contested?
03
Stability
Does the right answer hold steady no matter who's asking — or does it depend on a relationship, a history, a context that changes the calculation each time?
Where the answer to all three is yes, hand it to the model — that's not a compromise, it's the correct call. Where the answer to any one is no, the decision is still yours.
WHERE THE FRAMEWORK APPLIES
What This Looks Like
Fraud Detection
The clearest case for the model. High volume, a clean ground truth, a definition of success that holds steady no matter who's asking. The model surfaces the pattern — a person still decides what to do about it.
Diagnostic Imaging
AI can match or exceed a specialist at spotting a pattern in a scan. It has nothing to say about a patient's life beyond that image — the judgment call stays exactly where it's always been.
Logistics & Routing
Countless variables, one agreed number to optimize against — exactly the terrain a model was built to dominate. The moment the goal stops being a single number, the decision moves back to a person.