Speed is the headline. Under 60 seconds from PDF to takeoff is what people notice first. But speed without accuracy isn't useful — it's just fast noise.
We think about accuracy more carefully than almost anything else we do. Here's why, and how we measure it.
What Accuracy Means in a Takeoff
"Accurate" in the context of a painting takeoff means the measured square footage is close enough to the actual paintable surface that your estimate won't be significantly off. It doesn't mean perfect — every takeoff involves some degree of approximation.
Industry practice generally accepts a manual takeoff error of plus or minus 5–10% as normal variation between estimators working from the same drawings. Good estimators are consistent within that range. Bad ones aren't.
For an AI system, the relevant question is: how close to that human range can we get, and how consistent are we across different project types?
How We Measure Our Performance
We built a test set of projects that had already been estimated by experienced contractors. We run our pipeline on those same drawings and compare the outputs.
Our primary metric is Mean Absolute Percentage Error (MAPE) — the average absolute difference between our measurement and the human reference, expressed as a percentage. On our current test set, we achieve roughly 13% MAPE on wall square footage, which translates to about 87% accuracy on the first run.
That's not perfect. But it's within the ballpark of competent manual estimation, and it's improving.
Where We're Strong
On well-formatted vector PDFs — clean exports from modern CAD software — our accuracy is highest. These drawings have precise geometry, consistent symbols, and reliable dimension annotations. Our pipeline reads them cleanly.
Residential projects in the 1,500–4,000 square foot range are our sweet spot. The drawing conventions are familiar, room counts are manageable, and the geometry is typically straightforward.
Where We're Still Improving
Complex commercial projects with unusual layouts are harder. Drawings that mix multiple scales on a single page require more inference. Projects with non-standard door and window symbols occasionally produce undercounts.
Scanned drawings — PDF files that are essentially high-resolution images rather than vector data — are the hardest case. We support them, but accuracy drops meaningfully compared to vector PDFs. If you're regularly working from scanned documents, you'll want to be more careful about reviewing the output.
What This Means for Your Workflow
The right way to use Vector Takeoff right now is as a fast first draft. Use it to get a baseline measurement in under a minute, then spend your review time on the rooms or surfaces where your experience tells you to double-check.
This is actually how experienced estimators use any tool — they don't trust blindly, they verify strategically. The goal is to make that process faster, not to eliminate judgment.
Accuracy Will Keep Improving
Every project we process teaches us something. We track disagreements between our output and human-reviewed corrections. We add edge cases to our test suite. The model improves with each iteration.
We won't publish numbers we can't back up. When we say 87% accurate, that's based on real measurements against real projects — not a best-case demo we cherry-picked. And we'll update that number as it changes.
The goal is to make Vector Takeoff the most reliable automated takeoff tool in the industry. We're not there yet. We're working on it.