Estimate and de-risk an ambiguous initiative

WHAT THIS TESTS The interviewer wants intellectual honesty under ambiguity. A large, fuzzy initiative cannot be estimated as a single date without lying. They are checking whether you decompose to learn, sequence to retire risk, and communicate uncertainty so stakeholders can plan rather than be surprised.

A GOOD ANSWER COVERS Decompose the initiative into phases ordered by risk, not by feature appeal. Phase one is a discovery spike or a thin end-to-end slice that answers the scariest unknowns: do we have the data, does the model beat a simple baseline, can we serve it within latency. Each phase produces learning that sharpens the next estimate, so the cone of uncertainty narrows over time. Communicate estimates as ranges or confidence levels, not false-precision single numbers, and state the assumptions that, if wrong, blow up the range. Tie delivery to milestones and decision points where you may continue, pivot, or stop. This turns a scary blank check into a series of bounded, learnable bets.

COMMON WRONG ANSWERS Giving one confident date for work nobody understands yet. Padding the estimate silently rather than naming the risks. Estimating the whole thing up front as if scope were known, then treating overruns as failures rather than learning.

LIKELY FOLLOW-UPS What is the cone of uncertainty? How do you decide what goes in the first spike? How do you communicate a range to a leader who demands a single date?

ONE CONCRETE EXAMPLE For AI-powered recommendations, phase one is a two-week spike: assess data quality and test whether a simple model beats the current heuristic offline. You report it as 'two weeks to a go or no-go decision, then a six-to-twelve-week build if green, contingent on data being usable.' After the spike confirms the data, you re-estimate the build at eight to ten weeks with higher confidence, having retired the biggest risk first.