TRD-03 · Estimating under uncertainty

Compare AI-assisted estimates against your shop's flat-rate guide and its actual job history. Where does the model run over, where under, and which of those errors is the one that costs you a customer?

At least twelve completed jobs with actual hours and parts recorded, identifying details removed. The shop's flat-rate or labor guide and its historical actuals are the two comparators — the model is the third estimate, never the reference. An AI tool the school permits, named with version. Estimates are produced from the same job description a service writer would have had, with the actual outcome withheld until scoring.

Showing the model the outcome, which turns the whole study into a transcription exercise. Reporting mean error and hiding the spread — an estimator that is right on average and wrong by six hours either way is useless. Treating over and under as equivalent when they land on different people. Comparing to flat rate alone while ignoring what the job actually took.

An estimating comparison: all three estimates per job against actuals; error distribution reported with spread, not just an average, and over- and under-estimates broken out separately; a written account of the customer-communication consequence of each direction — the surprise upcharge versus the unprofitable job; and a recommendation on where, if anywhere, AI assist belongs in the estimating workflow, including "nowhere yet" if that is what the numbers say.

CT: reading a distribution rather than an average. CR: designing a blind comparison. CO: a peer checks the actuals arithmetic on five jobs. CM: explaining an estimate range to a customer without hedging into meaninglessness. CZ: who eats the difference in each direction.

A service writer, estimator, or shop owner reviews the comparison design before estimates are generated. Standing instruction: reject any design where the outcome is visible at estimate time. School-supervised.

R1: one trade, one job family, twelve-plus jobs. R2: blind procedure documented and replayable. R3: flat rate and historical actuals both present. R4: spread reported; directions separated. R5: recommendation is actionable. R6: willing to conclude the model does not belong here.

(a) A single average error percentage, with no spread and no direction. (b) The model sees the actual, and the study measures nothing.

Lane A immediately (center capstone, or a business or marketing cross-listing). No verified credit claim.