A model built for one job: reading the drawing
Every AI takeoff product on the market rests on the same hidden foundation — the ability to look at a construction drawing and pull out what is actually on it. The counts, the dimensions, the schedules, the callouts, the symbols, the notes buried in the corner of sheet E-401. We call this layer information extraction, and it is the single thing that determines whether everything built on top is trustworthy or fiction. If the model miscounts the receptacles or misreads the luminaire schedule, no amount of clever pricing logic downstream can save the bid.
So we started there. Pilars AI is not a general-purpose model handed a blueprint and told to figure it out. It was built from the ground up for this one job: reading a plan set — scaled drawings, cross-sheet references, panel schedules, keynotes, revision clouds, and the unwritten trade conventions that a human estimator absorbs over fifteen years — and extracting the ground truth of what the project contains.
How Pilars AI performs
We evaluate Pilars AI on a held-out set of real commercial construction drawings, scored against estimator-validated ground truth. We report two numbers, because reading a drawing is really two problems.
Accuracy pulling structured data — counts, quantities, dimensions, schedule and callout values — from the plan set.
Accuracy reading purely visual information — symbols, linework, hatching and geometry that carry no text label.
Market and general-purpose ranges reflect the spread we observe when running the same held-out construction drawings through competing tools and off-the-shelf vision models. On this evaluation Pilars AI sits among the highest information-extraction accuracy on construction drawings we have measured.
The gap between the two numbers is the point. Information Extraction — 96.43% — covers everything anchored to text or structure: a fixture schedule, a door count, a dimension string, a keynote legend. Visual Information Extraction — 88.12% — is the harder problem: a symbol with no label, a run of linework that only means something because of where it sits, hatching that implies a material. Humans do this effortlessly and cannot explain how. Getting a model to 88.12% here is what separates a tool that reads drawings from a tool that reads text that happens to be on a drawing.
Why extraction is the floor, not the ceiling
Here is the part that matters most, and the part most easily misunderstood. Extraction is the floor. It is not the product.
Knowing that there are 214 duplex receptacles on the second floor is not a takeoff. It is not an estimate. It is not a bid. It is a fact. The value an estimator creates lives entirely in what happens after the fact is known: which of those receptacles are on a dedicated circuit, what the homerun length really is once you account for the vertical rise the plan flattened, which GC always under-scopes temp power, when a "typical" note quietly changes the count on three other sheets, and whether the spec book contradicts the drawing.
That layer is not extraction. It is reasoning — and reasoning is what Pilars AI does on top of a rock-solid extraction floor. Because the two flagship numbers are high, everything above them has clean inputs to reason from. You cannot reason your way out of a bad read.
"The extraction has to be near-perfect before the smart stuff even matters. If the model can't be trusted to count what's on the sheet, I can't trust anything it concludes. That's the part Pilars got right first."
Senior Estimator · Commercial Electrical · design partner
The reasoning was learned with estimators
We did not invent the reasoning layer in a lab. Every judgment Pilars AI applies on top of extraction — the waste factors, the cross-sheet reconciliation, the scope calls, the labor adjustments, the "this note means these three other sheets changed" instinct — was learned with practicing estimators, working over real ERP books and real bid histories, correcting the model and telling us why.
That is the difference between a model that scores well on a benchmark and a model that produces a takeoff an estimator will actually bid from. The benchmark tells you the floor is solid. The estimators are what taught the model how to stand on it.
Not "H1." Pilars AI.
Other teams ship a model and give it a version number. Ours is not H1, and it is not a generation label. It is Pilars AI — Building Pillars for Construction Technology. The name is the thesis: the pillars a modern construction business stands on — the takeoff, the estimate, the bid, the scope — should be built on a model that reads the drawing correctly, every time, and then reasons about it the way your best estimator would.
What this means for your shop
- The floor is trustworthy. 96.43% information extraction means the counts, quantities and schedule values the model hands you are accurate enough to review, not re-do.
- The hard visual reads are covered. 88.12% on visual extraction means symbols and unlabeled geometry — the part that trips up text-only tools — are handled, not skipped.
- The intelligence on top is real reasoning. Waste factors, cross-sheet checks and scope judgment are applied automatically, learned from estimators, on top of clean inputs.
- Your estimator moves up a level. Instead of counting on a PDF, they review a takeoff and spend their judgment where judgment actually wins bids.
Bottom line
A construction takeoff is only as good as the read underneath it. Pilars AI puts that read among the most accurate in the market — 96.43% on Information Extraction, 88.12% on Visual Information Extraction — and then layers on reasoning learned directly from estimators. That is the foundation the whole platform stands on, and it is why we named the model after the thing it is built to be: the pillars construction technology stands on.