This is the origin story of the most quietly useful agent I run: the one that does takeoff analysis — reading construction documents and extracting the scope of work and materials before a human estimator spends a day doing it by highlighter.
Construction documents are a hostile environment for machine intelligence. Thirty-page drawing sets where the critical detail lives in a margin note. Revision clouds stacked on revision clouds. Abbreviations that mean different things on different sheets. A coffee stain that has been photocopied so many times it now appears to be a design element. Into this walked a language model with the confidence of a straight-A student on the first day of a job site.
What happened
First pass: it read the clean sheets beautifully and hallucinated politely on the messy ones — confidence unbudgeted by accuracy. So we did what you do with any promising estimator: gave it structure. Per-line confidence scores. A rule that "not legible" is a valid and honorable answer. Cross-checks between the drawings and the spec book, because when they disagree — and they disagree — that's not an error, that's a finding, and findings make money.
Second pass, with a human estimator reviewing every line: hours of highlighter work compressed to a review that takes minutes, plus a bonus nobody priced — the agent flags scope gaps between drawings and specs that a tired human skims past. It once caught a door schedule referencing hardware the spec never included. That single catch paid for the experiment.
The moral
AI didn't replace the estimator. It replaced the highlighter. The estimator got promoted to judge — which was always the valuable part of the job. And the model's only complaint, expressed through confidence scores, is our industry's handwriting. Fair. It's earned that note.
Grade from the robot: scope, extracted; quantities, itemized; handwriting, 2/10, "illegible but confident." The robot fits right in here.