How AI Automates Doors &
Windows Takeoff

What a doors and windows takeoff involves and the manual pain

A doors and windows takeoff is more than counting rectangles on a floor plan. The deliverable is a complete opening schedule: counts by type, door leaves, frames, glass area for windows and glazed doors, and hardware sets keyed to the hardware schedule. Each of those line items carries different material and labor cost, so getting the breakdown right matters as much as getting the count right.

The challenge is that every opening on the plan carries a mark — a tag like D-01 or W-14 — and that mark is a pointer into a separate door or window schedule that defines size, material, fire rating, and hardware group. Manually, an estimator must locate each mark on the plan, look it up in the schedule, find the hardware group, then expand that hardware group into individual finish hardware items from a third document. On a medium commercial project with 150 to 300 openings, this cross-referencing alone can take 6 to 16 hours, and a single misread mark or skipped schedule row silently distorts the whole BOQ.

• Output: opening counts by type, door leaves, frames, glass area, and hardware sets per opening
• Each opening must be tied to a mark that defines size, material, rating, and hardware group
• Manual packages typically take 6–16 hours due to plan-to-schedule cross-referencing

Step 1 — Plan ingest and sheet classification

The first thing AI does with a set of architectural PDFs is classify every sheet before measuring anything. Floor plans, reflected ceiling plans, the door schedule, the window schedule, frame type sheets, and the hardware schedule are each distinct document types with different roles in the takeoff. Mixing them up produces garbage output, so this classification step is foundational.

AI isolates the door and window schedules and reads them in full — mapping each mark to its size, material (hollow metal, wood, aluminum storefront), fire rating, and hardware group assignment. Frame elevation or frame type sheets are tagged separately so frame quantities can be pulled from the right source. Only once this document map is in place does the AI begin working the floor plans.

• Sheets classified: floor plans, door schedule, window schedule, frame types, hardware schedule/sets
• Door and window schedules parsed: mark → size, material, rating, hardware group
• Frame elevation sheets tagged for frame quantity extraction

Step 2 — Scale detection and calibration

Architectural floor plans are drawn at a stated scale — typically 1/8" = 1’-0" or 1/4" = 1’-0" for commercial work — and AI reads that annotation directly. It then validates the stated scale against a dimensioned reference grid or a known dimension line on the same sheet, catching cases where a sheet was printed at a non-standard size or where the scale block is stale from a reused border.

For doors and windows, this matters less than for, say, concrete or flooring: opening sizes come from the schedule rather than from graphically measuring the plan. But per-sheet calibration still confirms that opening locations are read consistently across floors, and it provides a cross-check when a plan callout next to an opening contradicts the schedule entry — a discrepancy the AI flags for estimator review.

• Scale read from plan annotation, validated against dimensioned grid
• Opening sizes drawn from schedule, not graphic measurement
• Per-sheet calibration confirms locations; plan-vs-schedule conflicts are flagged

Step 3 — Symbol recognition and reading schedules

Door and window symbols on architectural floor plans follow recognizable conventions — a door is typically shown as a quarter-circle swing arc with a rectangle for the leaf; a window is a pair of parallel lines in a wall break — but in practice symbols vary by architect and CAD standard. AI is trained to recognize the common symbol families and, critically, to find the mark or tag number associated with each symbol. That tag is the key that unlocks everything downstream.

Once the AI has located a door or window symbol and read its mark, it looks that mark up in the schedule it already parsed in Step 1. The result is a rich record for each opening: width, height, leaf configuration (single, pair, Dutch), material, frame type, and fire rating from the door schedule; hardware group from the hardware schedule. Sidelites, transoms, and borrowed lites are recognized as associated elements of the parent opening and added to that opening's record rather than counted as independent openings, since they share the frame assembly.

• Door and window symbols detected; mark/tag read from plan callout
• Each mark looked up in the pre-parsed schedule: size, type, material, rating, frame, hardware group
• Sidelites, transoms, and borrowed lites tied to their parent opening

Step 4 — Measurement and quantity computation

With every opening identified and linked to its schedule row, AI computes quantities. Door openings are counted by mark and type. For each opening, the leaf count is tallied: a single door is one leaf; a pair counts as two leaves with one frame. This distinction matters for labor — hanging two leaves is different from hanging one — and for hardware, since each leaf typically gets its own set of hinges while the closer and lockset may be shared.

Glass area for windows and glazed doors is computed as opening width × height × quantity, then grouped by glass type (clear, tinted, tempered, fire-rated) and framing system. Hardware is expanded by hardware group: the AI uses the hardware schedule to look up what each group contains — number and size of hinges, lockset type, door closer, wall stops, door silencers, threshold, and perimeter seal — and multiplies by the count of doors assigned to that group. The result is a hardware takeoff driven by the schedule rather than estimated from rules of thumb.

• Openings counted by mark and type; leaves per opening tallied (pairs = 2 leaves, 1 frame)
• Glass area = opening size × quantity, grouped by glass type
• Hardware expanded from hardware group: hinges, lockset, closer, stops, seals per opening

Step 5 — Assembly mapping, waste, and BOQ output

Each opening is mapped to a CSI Division 08 assembly: door, frame, and finish hardware installation. Fire-rated openings — those carrying a 20-minute, 45-minute, 60-minute, or 90-minute label — are routed to rated assemblies that carry the appropriate rated frame and fire-rated hardware items (rated closers, intumescent seals, and coordinator hardware for pairs). This separation is important for both pricing and compliance documentation.

Hardware sets are exploded into individual finish hardware line items organized by group, so the estimator can review exactly what Pilars counted for Group HW-3 or Group HW-7 against the specification. The final output is a CSI Division 08 BOQ listing openings by type, frame type and count, glass area by type, and hardware items by group — exportable to Excel for transfer into the estimate. No glass waste factor is applied without the estimator setting it, since glazing waste is project-specific.

• Openings mapped to door + frame + hardware-set assemblies; fire-rated openings routed to rated items
• Hardware sets exploded into individual finish hardware line items by group
• Output: CSI Division 08 BOQ with openings, frames, glass, and hardware sets, exportable to Excel

Step 6 — Estimator review and accuracy

AI performs well at the two hardest parts of a manual opening takeoff: matching marks to schedule rows without transcription errors, and maintaining that linkage consistently across hundreds of openings. Opening-count accuracy is typically 95 to 98% on clean plans with complete, legible schedules — meaning an estimator reviewing a 200-opening package will generally find two to ten discrepancies rather than the dozens that can accumulate in a rushed manual count.

The AI is weaker in two predictable situations. First, when the hardware schedule is incomplete — groups are listed but items within a group are not specified — it cannot expand hardware and will flag those openings for manual resolution. Second, when openings on the plan lack a mark entirely (a drafting omission), the AI reports the count of unmarked openings rather than guessing their type. Both failure modes surface to the estimator as specific flags rather than silent errors, which is the right behavior: the estimator knows exactly where to focus the review. That review typically takes 1 to 2 hours, versus 1 to 2 days for a fully manual takeoff of comparable scope.

• Opening-count accuracy: 95–98% on clean plans with complete schedules
• Flags: incomplete hardware schedules and unmarked openings surfaced explicitly
• Estimator review: 1–2 hours vs. 1–2 days fully manual

Questions estimators actually ask