Can ChatGPT do a drywall takeoff? We tested it on a real plan sheet

Short answer: not yet, and not in the way that matters most. On October 9, 2026 we gave free ChatGPT (signed out, the version most people try first) a real first-floor plan from an apartment renovation and asked for a framing and drywall takeoff. It first declined to give a number, then, when pushed, estimated 220 LF of new walls, 195 studs and 91 sheets. The sheet draws no new walls at all: every wall on it is marked "existing to remain."

We also ran our own tool on the same sheet. It got the same thing wrong, in a different way. Both results are below, along with what we changed.

The test

The plan set is the same 21-page apartment renovation we used in our first public test. We used sheet A201 (first floor plan, 1/8" = 1'-0") and the wall-types table from sheet A901. We asked the question the way a sub would:

We gave it every advantage we could: the scale, the building length (about 138 ft) and the default wall type from the legend (A4: 2x4 studs at 16" o.c., 5/8" gypsum both sides).

What happened, step by step

• The PDF was rejected. Signed-out ChatGPT only accepted images, so we exported the sheets as PNGs, which is what most people would do next.
• Full sheets: "too low-resolution." It asked for sharper images and gave the formulas instead (LF × 0.75 = studs, LF × 0.375 = sheets).
• A sharp 250-dpi crop got shrunk anyway. The upload came back as a 1,024-pixel-wide image. At this scale that is about 6 pixels per foot, so a 4¾" wall is about 2 pixels wide.
• It said it couldn't tell new walls from existing. That was fair: we had cropped off the legend. So we typed the legend in: existing walls are gray with gray fill, new walls are solid black with black hatch, demolition is dashed.
• With the legend in hand, it gave numbers. 220 LF of new partitions (range 200–240), 195 studs including an 18% allowance, and 91 sheets of 4x12 including 10% waste. It labelled them "budgetary" and said to verify against the PDF.

How we know it's zero

We didn't eyeball it. A CAD-exported PDF stores every line with its color, so we read the sheet's own vector data. Every wall on A201 is drawn in gray: about 1,400 LF of it. The only black linework on the plan is door swings, fixtures, tags and symbols. The renovation work on this floor is finishes, the elevator and MEP, and the community-room changes are drawn separately on the enlarged plans (sheet A207, which the dashed box on A201 points to).

So ChatGPT's 220 LF wasn't a measurement that came out a bit high. It was a plausible-sounding number for walls that aren't there. Priced as framing and board, that is real money in a bid, and the answer looked careful: formulas, a waste factor, a range.

Our own tool got it wrong too

Before writing this, we ran the same sheet through ReadMyBlueprint. It counted 1,280 LF of walls on A201: all the existing ones.

We already had a rule for this. When a plan set's legend labels a swatch "EXISTING PARTITION" or "EXISTING WALL", walls drawn in that fill aren't counted. On this sheet the rule failed for three reasons:

• The label was too long. "EXISTING WALLS / ELEMENTS TO REMAIN (GRAY LINES, GRAY FILL, DOORS 45°)" is 72 characters. We ignored any label over 60, to stay clear of general notes.
• The demolition row looked like an "existing" label. In "EXISTING WALLS / ELEMENTS TO BE REMOVED", the words in the middle hid "removed" from our check.
• The swatch gray doesn't match the wall gray. The legend swatch is drawn at 75% gray, but the walls on the plan are at 88% gray. We required an exact match.

The 24 LF left on A201 are short wall-shaped pieces we haven't traced yet, so a person should look at them. Our tool also skips enlarged-plan sheets like A207, so it doesn't count the same wall twice. That means the community-room walls on this job are still a manual count today. And when the tool sets walls aside, it says so on the sheet ("Not counted: 1,403 LF of walls drawn in the fill this plan set's legend marks as EXISTING"), so you can check it rather than trust it.

What this means if you're tempted to use ChatGPT for takeoffs

• Use it for the math, not the measuring. Its formulas were right (16" o.c. = 0.75 studs per foot, 18 sq ft of board per foot of 9' wall). It can't see the drawing well enough to supply the footage.
• Watch what it does under pressure. Its first answer, "I can't tell," was the honest one. Asked again, it produced a number. If you push a chatbot for a figure, you'll get one.
• On renovations, new vs. existing matters more than anything else. Every tool, ours included, has to read the legend to get this right. Check it on any takeoff before you price it.
• Check images against the PDF. A CAD-exported PDF holds the exact linework. A screenshot throws most of it away before anyone reads it.