Choosing what to build.
WHAT TO ASSESS FOR EACH CANDIDATE
The value if it works The probability it works The data available The cost to build and run The cost to maintain Whether anyone will act on the output
WHY THE LAST POINT DECIDES MOST CASES
Predictions nobody acts on have no value at any accuracy.
WHAT TO PREFER EARLY
Problems with existing data Problems where a modest improvement is valuable Problems where the decision already exists
WHY THAT LAST ONE
A model improving an existing decision needs no new process.
WHAT TO DEFER
Problems requiring new data collection Problems requiring organisational change to act on Problems where success cannot be measured
WHAT TO DECLINE
Problems where a rule would work Problems with no clear definition of success
WHAT TO SEQUENCE FIRST
Something small that ships, establishing the path to production.
WHY
Every subsequent project becomes cheaper once that exists.
WHAT TO REVIEW PERIODICALLY
Whether deployed models are still used, still accurate, and still worth their cost.
WHAT TO RETIRE
Anything failing that review.
WHY
Unused models consume maintenance and erode trust when they degrade unnoticed.
