Testing AI take-off on real facade drawings
Anonymised example. This is based on our own October 2026 tests of the QuantRead reading engine for a specialist sealant contractor in the East of England. The contractor, the projects and the drawing numbers are not named. All figures below come from the test log; nothing is invented or rounded up.
The job
Sealant work is priced by the linear metre: frame perimeters around windows and doors, movement joints, panel joints. Measuring those off multi-storey elevations by hand takes hours per sheet. The test question was simple: can an AI reader produce the counts and metres from the drawings the contractor actually receives, and say honestly where it is unsure?
What we tested
Real tender and construction PDFs for several trades, in October 2026: facade elevations, floor plans, electrical and mechanical layouts. The sealant results below are from three facade elevations.
What matched
- On one elevation with tagged openings, the window, door and louvre counts matched the tags exactly: 19 + 19 windows, 1 door, 7 louvres.
- On a second tagged elevation, one successful run matched 34 windows and 5 doors exactly. (A later run on a different build moved six balcony windows into the door count, which is why every result carries a check-on-site list.)
- On a third elevation, the 1:150 scale was read from the title block and checked against level datums. After the measuring method was tightened, the derived full window and door perimeter was 221.0 m against a manual take-off of 238.9 m (92.5% of the manual metres).
- Movement joints were taken from the written dimensions on the sheet, with each run listed.
What did not match, and what we changed
- On that third elevation, the opening count still differed: 38 from the reader against 24 in the manual count, because lobby and low-level units were grouped differently.
- An early run folded curtain-wall bays into the window count and perimeters. Curtain-wall and glazed-screen bays are now always a separate line, never mixed with windows.
- Door sizes too small to scale are marked “ASSUMED sizes: check” with low confidence.
- Fire-stopping counts were low confidence where the sheet had no fire-rating text. Flooring areas on one plan were scaled with no written areas to check against.
Time
Successful test runs typically took about 3 to 16 minutes per drawing, mostly the AI reading large sheets at high resolution.
What this means
The 92.5% figure is one perimeter on one sheet, not a product-wide accuracy claim. What the tests show is that the reader gets counts and written dimensions right more often than not, shows its working, and flags the lines that need a person. That is what QuantRead is for: a fast, checkable first pass, not a replacement for your own measure.
Every result is an estimate: it states the scale used, the assumptions, a confidence per line and a “check on site” list. It is not a survey, a bill of quantities or a price.