How AI Automates
Glazing Takeoff

What a glazing takeoff involves and the manual pain

A complete glazing takeoff is more layered than most trades because the work spans three structurally different systems. Punched windows and doors are counted by type mark and tied back to the window schedule for dimensions and glass spec. Storefront systems are measured by area — system width times height — and broken into framing footage for the vertical and horizontal members. Curtain wall adds another layer: the mullion grid, spandrel versus vision glass splits, anchor conditions, and sometimes project-specific pressure ratings that affect the framing section.

The deliverable is a Division 08 BOQ covering unit counts, glass area in square feet grouped by glass makeup, framing linear footage (mullions, sills, heads, rails), and hardware counts for locks, closers, and panic devices. On a commercial project of moderate complexity, assembling that package manually means walking every exterior elevation, cross-referencing each callout to the glazing schedule, and measuring area and framing separately for each storefront or curtain-wall bay. A full glazing package routinely takes 8 to 20 estimator-hours depending on building size and system complexity.

• Unit counts (windows, doors, storefront bays, curtain-wall panels)
• Glass area in SF, grouped by glass type (IGU, low-E, tempered, spandrel)
• Framing linear footage — mullions, sills, heads, rails
• Hardware counts per opening type

Step 1 — Plan ingest and sheet classification

The first job is figuring out which sheets in a full set actually contain glazing information. A typical commercial PDF has dozens of sheets — civil, structural, mechanical — and the glazing data is scattered across exterior elevations, the window and storefront schedule, glazing details, and floor plans that show opening locations by mark.

AI scans the sheet index and sheet titles to isolate the relevant pages, then classifies each: exterior elevation, glazing schedule, curtain-wall elevation, storefront detail, section, or plan with opening marks. The glazing schedule is the keystone document: it maps every type mark (W1, SF-A, CW-3, and so on) to width, height, glass makeup, frame system, and finish. Once that schedule is parsed, AI has a reference table it can use to validate every measured opening against a known dimension.

Sections and details are tagged separately so that mullion spacing, sill height, and head conditions can be pulled for framing calculations later. This classification step is what allows the rest of the pipeline to be systematic rather than ad-hoc.

Step 2 — Scale detection and calibration

Glazing accuracy depends on a reliable scale. Elevation sheets commonly carry a stated scale such as 1/8" = 1'-0" or 1/4" = 1'-0", and AI reads that annotation directly. It then cross-checks the derived scale against a known dimension on the same sheet — typically the floor-to-floor height, which appears as a dimension string on most elevations and can also be inferred from the structural grid.

For punched openings, this validation is a useful consistency check, because the window schedule often carries the nominal rough-opening size. If the measured pixel dimension, scaled up, agrees with the schedule callout to within a reasonable tolerance, AI flags that opening as confirmed. If there is a discrepancy — common when drawings have been printed at non-standard sizes or when a PDF was digitally combined from differently-scaled sheets — the sheet is flagged for estimator attention before quantities are committed.

Per-sheet calibration matters because a glazing set frequently mixes elevations at different scales, and a single global scale assumption would propagate area errors across the entire takeoff.

Step 3 — Object recognition and reading the schedule

With the sheets classified and scale confirmed, AI moves to detection. On elevations and floor plans it identifies window and door marks — the alphanumeric callouts in circles or rectangles that reference the glazing schedule — and records each instance with its sheet location and coordinates. Storefront systems appear as gridded rectangles on elevations; AI recognizes the characteristic mullion pattern and classifies the bounding region as a storefront or curtain-wall bay.

The parsed schedule then does the heavy lifting. Each detected type mark is matched to its schedule row, pulling nominal size, glass makeup (single-pane, insulated glass unit, low-E coating, tempered, laminated), and frame system (thermally broken aluminum, non-thermally broken, vinyl, fiberglass). This mapping is what separates a quantity takeoff from a count: the AI is not just counting marks, it is associating each one with a complete material specification that drives area grouping and hardware quantities.

For storefront and curtain-wall grids, AI reads the mullion spacing from the elevation or the detail — typically 24", 30", or 60" module — which feeds directly into the framing footage calculation in the next step.

Step 4 — Measurement and quantity computation

Quantity computation branches by system type. For punched openings, the scheduled dimensions are used directly rather than pixel-measured dimensions, since the schedule is the authoritative source. AI tallies units by type mark, and glass area for each type equals the scheduled glass size (not the rough opening) times the count. Hardware counts — locks, closers, panic devices — are derived from the opening classification in the schedule.

For storefront and curtain-wall, AI measures the system boundary on the elevation: overall width and height of each glazed bay. Glass area equals width times height. Framing footage is computed by counting the mullion grid: vertical mullions run the full height at the mullion module spacing, and horizontal rails run the full width at the vertical module spacing. The formula is straightforward once the module is known, but it scales quickly across a large curtain-wall elevation.

Glass area is grouped by glass type at this stage because IGU, tempered, spandrel, and low-E units carry materially different unit costs and lead times. A BOQ that lumps all glass area together is not useful for pricing, so the grouping is built into the computation rather than left to the estimator to sort out afterward.

Step 5 — Assembly mapping, waste, and BOQ output

Raw quantities become a priceable BOQ through assembly mapping. Each unit type expands into its components: a storefront system includes glass units, aluminum framing sections (head, sill, jamb, mullion), setting blocks, glazing gaskets, perimeter sealant, and anchor conditions. Curtain-wall adds back pans for spandrel zones, pressure plates, and snap covers. The detail sheets tagged in Step 1 provide the anchor and sill conditions that determine which assembly variant applies.

A modest waste allowance is applied to glass and framing material — typically 5 to 8 percent for glass and a rounding-up to standard stock lengths (12-foot or 24-foot extrusion lengths) for framing. These are conservative defaults that an estimator can override based on project conditions or supplier constraints.

The output is a structured Division 08 BOQ: punched window and door unit counts by type, glass area in SF by glass makeup, framing linear footage by member type, hardware counts by opening, and perimeter sealant in linear feet. The file exports to Excel with one row per line item, ready for pricing. Total processing time from PDF upload to draft BOQ is typically a few minutes for a standard commercial glazing package.

Step 6 — Estimator review and accuracy

AI is consistently strong at two things in glazing: counting scheduled openings on clean elevations, and measuring storefront and curtain-wall area from clearly drawn bays. Unit-count accuracy on clean plans with a legible schedule is typically 94 to 98 percent, meaning an estimator reviewing a 200-opening project might find three to twelve corrections rather than re-counting from scratch.

The weaker areas are well-defined and flagged rather than silently assumed. Complex curtain-wall geometry — angled facades, radius bays, multi-story unitized systems with non-rectangular panels — requires estimator attention because the pixel-based area calculation cannot reliably handle non-orthogonal boundaries. Vision-versus-spandrel splits within a curtain-wall elevation are flagged when the boundary between the two is not a drawn dimension line but an implied condition from the detail. Custom framing conditions — thermally broken versus non-thermally broken, pressure ratings, thermal isolators — are noted for review when the schedule is ambiguous.

The practical result is that estimator review on an AI-generated glazing takeoff runs 1.5 to 3 hours rather than the 8 to 20 hours of a fully manual effort. The estimator is reviewing and spot-checking rather than counting and measuring.

Questions estimators actually ask