One-line takeaway: Not counting. Not diffing. Four specific trade checks: equipment-schedule cross-check, transition sweep, accuracy spot-check, cross-set inference. Stopwatch on.
Last week I wrote about the wrong measurement, quantity match, and the right one, elapsed hours from drawings-in to a takeoff you would stake your name on. On the hospital-set ductwork run that framed that piece, my review pass was 22 minutes. A reasonable next question, one estimators have asked me directly since the post went up, is what happens in those 22 minutes.
The honest answer is that most of it is boring, and that is what makes it work.
I open the AI output next to the equipment schedule first. Not the drawings. The schedule. Equipment schedules are where undrawn scope hides in ductwork, because the schedule tells you which two rooms have to connect for the mechanical system to function even when the drafter left the run itself off the sheet. I scan the AI’s counts for those room-pair connections and if a run is missing, I add it. On the hospital set that took about four minutes and surfaced three connections the AI could never have inferred.
Next I walk every duct transition on the drawings the AI did quantify. Transitions are where a size change forces a fitting whether or not the plan calls it out. If a 24-inch main steps down to a 16-inch branch and the AI counted zero fittings on that node, I add the transition fitting because I know the trade always installs one. On dense mechanical drawings a full transition sweep is about eight minutes. This is the part of the review that looks the most like counting, but the volume is small and the mental cost is low because I am pattern-matching on the same fitting family the whole way through.
Then I spot-check the AI on drawn scope where I know estimators, human and AI, get it wrong. Three places consistently: complex junctions where four or five runs meet at different elevations, curved runs where the AI’s centerline extraction can drift, and sheets with poor legibility from a bad scan. If the AI’s count on those areas looks clean, I sample one and hand-verify. If it looks off, I sample three. On the hospital set that was six minutes.
The remaining four minutes go to the parts of the drawing set the mechanical takeoff depends on but the mechanical sheets do not fully show. The structural set implies a beam I have to route around. The architectural set implies a drop soffit I have to run below. The equipment schedule implies a fitting I did not add during the transition sweep because the schedule column is on a different sheet. Cross-set inference is the last thing I do, because catching a cross-set implication is a five-minute save that saves two hours of downstream RFI resolution.
None of what I described is quantity diffing. I am not comparing my old-workflow number to the AI’s number and reasoning about the delta. I am looking at four specific places where I know experience adds value the AI cannot add on its own. Every one of those checks maps to a specific trade phenomenon: undrawn connections in the schedule, always-installed transition fittings, error-prone geometry categories, and cross-set implications. If a junior estimator on my team runs the same review pattern, they get the same output. If they diff quantities, they get twenty different opinions on whether the AI is any good.
The 22 minutes is not a lower bound. It is the shape. On a straightforward warehouse mechanical set the review pass is closer to 12 minutes because there are fewer transitions and the equipment schedule is trivial. On a full hospital set with three mechanical rooms and specialty exhaust it is closer to 45 minutes because the cross-set implications proliferate. What stays constant is the four buckets. The estimator’s job is those four buckets. The AI’s job is everything else on the drawings.
I get asked whether this review pass can be automated. The parts inside the drawing set can, in principle, get closer to full automation as the models improve. The cross-set inference part cannot, because it depends on trade knowledge about what the drawings imply but do not state. That is the layer where our value is stored, and it is the layer the AI hands back to me every run. The AI does not want that job. It does not have the priors to do that job. Twenty-two minutes of trade experience is what the takeoff cost, and it is where the trade earns its keep.
