How AI Automates
HVAC Takeoff

What an HVAC takeoff involves and the manual pain

A complete HVAC takeoff has several distinct output streams. On the equipment side you need counts of RTUs, AHUs, VAV boxes, fan coil units, exhaust fans, and split systems. On the air-distribution side you need diffuser, register, and grille counts per type and CFM. Then there is duct — measured in linear feet by size and converted to sheet-metal pounds — plus piping for hydronic and refrigerant circuits on separate plans.

The sheet-metal conversion is where manual HVAC takeoff becomes time-consuming. Rectangular duct is priced by weight, not length, because the cost driver is material and fabrication. That means every duct segment must be looked up in SMACNA gauge tables based on its width, height, and pressure class to compute pounds per linear foot before any cost can be applied. On a mid-size commercial project with dozens of duct sizes and hundreds of fittings, this alone can consume 15–35 hours of careful counting and arithmetic.

Fittings compound the problem. An elbow, transition, or tap-in is not simply a change in direction — it contributes additional sheet-metal area beyond the straight run on each side. Estimators must either apply equivalent-length rules or compute the fitting surface area directly. Either way, it is painstaking work that must be done correctly before any number in the BOQ can be trusted.

Step 1 — Plan ingest and sheet classification

The first thing AI does when it receives a mechanical set is sort the sheets by type. M-series drawings follow a consistent convention across most commercial projects: M0 or M0.x pages carry the mechanical legend, symbols list, and master equipment schedule; M1 pages contain the HVAC floor plans showing duct layout and air devices; M2 pages show hydronic and refrigerant piping; M3 has construction details; and M4 covers controls wiring and equipment schedules in detail.

AI reads sheet titles, title-block data, and drawing content to assign each page to one of these categories. Ductwork plans are separated from piping plans because they require different measurement routines. Equipment schedules — RTU, AHU, VAV, diffuser, pump — are tagged specifically for table parsing, since the structured data in those tables (CFM ratings, tonnage, electrical data, weights) will be needed later to validate measured quantities and build the equipment BOQ.

This classification step matters because applying a duct-measurement routine to a piping plan, or a table-parser to a floor plan, produces garbage. Getting the sheet type right before any measurement begins is prerequisite work, not a shortcut.

Step 2 — Scale detection and calibration

Before any duct length can be measured, AI must establish an accurate scale for each sheet. It reads the plotted scale notation in the title block or drawing border, then validates that against a known dimensioned element — typically a grid reference, a room dimension string, or a labeled duct segment — to confirm the PDF was not reproduced at a non-standard size.

This validation step exists because reduced-format prints are common. A set that was designed at 1:50 but printed on a smaller sheet will have a different pixels-per-foot ratio than the title block claims. AI detects this discrepancy by comparing the calibration measurement against the stated scale and adjusting accordingly, per sheet, rather than assuming a single scale applies across the entire document.

HVAC plans have a built-in cross-check that most other trades lack: duct width is almost always labeled directly on the plan as a callout (for example, 24×12 or 18×10). AI uses these callouts to verify its calibration — if the measured width of a labeled 24-inch duct does not match the scale expectation, the calibration is flagged for review before measurement proceeds.

Step 3 — Symbol recognition and reading schedules

With calibration confirmed, AI applies vision models to the floor plans to detect air devices and equipment outlines. Standard HVAC symbols — rectangular diffusers, round ceiling diffusers, return and exhaust grilles, linear slot diffusers, VAV boxes, and fan coil unit outlines — are recognized by shape, size relative to the plan scale, and the annotation patterns that surround them. Each detected symbol is assigned a type and linked to its schedule tag where one is called out on the plan.

In parallel, AI parses the equipment schedules using table-recognition techniques. RTU and AHU schedules typically list each unit by tag number with design CFM, cooling tonnage, heating capacity, electrical characteristics, and sometimes operating weight. VAV schedules list each box by tag with minimum and maximum CFM. Diffuser schedules list neck size and design CFM per device. AI extracts all of this into a structured lookup table that connects plan symbols to their scheduled properties.

Duct size callouts and pressure-class notes along each run are read by OCR. The pressure class — low, medium, or high — determines which SMACNA gauge column applies and therefore how many pounds per square foot of duct surface the sheet metal weighs. Missing pressure-class callouts default to low-pressure and are flagged for estimator confirmation.

Step 4 — Measurement and quantity computation

Duct measurement follows the centerline of each run from the supply trunk out to the terminal device. AI traces each segment, records its length in calibrated units, and groups runs by the duct size callout associated with that segment. Rectangular duct is grouped by both dimensions (width × height) and pressure class, since both determine the SMACNA gauge and resulting weight per foot.

Once lengths are tallied by size, conversion to sheet-metal pounds proceeds using SMACNA gauge tables: the perimeter of the duct cross-section times the length gives surface area in square feet, and the gauge for that size and pressure class gives pounds per square foot. Round and oval duct follow the same logic by diameter. The output at this stage is total pounds of sheet metal broken out by size group — the primary cost driver for a mechanical sub's material and fabrication estimate.

Fittings are handled using equivalent-length tables. An elbow in a 24×12 duct at low pressure adds a defined number of equivalent linear feet of straight duct, accounting for the extra material in the curved section. Transitions, offsets, and tap-ins each have their own equivalent-length values. AI detects fitting symbols, classifies them by type, and adds the equivalent footage to the affected size group automatically rather than leaving it to the estimator to catch by memory.

Air devices are counted per type against the schedule. CFM totals per device type are summed and compared against the equipment schedule design airflow for that air-handling unit or RTU, providing a built-in sanity check on the count before the BOQ is assembled.

Step 5 — Assembly mapping, waste, and BOQ output

Raw duct weight and equipment counts are not yet a usable BOQ. Each quantity needs to expand into the assembly of materials and labor that the mechanical sub will actually price. AI maps duct to assemblies that include, in addition to the sheet metal itself: hanger rod and strapping (typically spaced per SMACNA support requirements), duct sealant, and insulation wrap where the plan calls it out. A sheet-metal waste factor of 10–15% is applied to account for scrap, offcuts, and fitting over-material — a standard allowance in mechanical estimating practice.

Equipment is mapped to set, rig, and connect labor units. An RTU gets crane or hoist time, roof curb installation, electrical connection, and start-up. An AHU gets rigging, piping connections, and controls integration. VAV boxes get installation and balancing labor. Air devices are mapped to their install labor unit per type — a ceiling diffuser in a standard grid ceiling is a different labor unit from a linear slot diffuser in a custom plenum.

The final output is a CSI Division 23 BOQ organized by work section: 23 31 (duct and accessories, in pounds by size group), 23 74 (packaged equipment, by tag), 23 36 (air terminal units), 23 09 (controls, itemized from the M4 sheets), and piping by size and material. The BOQ exports to Excel with the source sheet reference for each line item, so the estimator can trace any number back to the plan page where it was measured.

Step 6 — Estimator review and accuracy

AI handles the counting and measuring well; the estimator's job shifts to verification and judgment. Air-device and equipment count accuracy on clean, well-drafted PDFs typically falls in the 93–97% range — meaning on a 200-diffuser project, two to fourteen devices may need manual correction. That is a very different review burden from counting 200 diffusers by hand, but it is not a zero-review workflow.

Sheet-metal weight accuracy depends on two things the estimator must confirm: that the pressure-class assignment for each duct run is correct, and that the gauge table being applied matches the specification. A project spec requiring heavier gauge than SMACNA minimum will produce incorrect weight if AI defaults to standard tables without a spec override. AI flags runs where the pressure class was not explicitly called out on the plan so the estimator can apply project-specific requirements.

AI is explicitly weaker on complex fitting geometry, double-line duct that overlaps with structural or other system layers, and runs that are implied by the design but not drawn — for example, flexible duct from a VAV box to a ceiling diffuser that is shown only schematically. These cases are surfaced in the review output with confidence indicators rather than being silently estimated and buried in the quantity.

The practical result: estimator review of an AI-produced HVAC takeoff typically takes 2–4 hours, compared to 2–5 days for a fully manual takeoff of comparable scope. The time savings are concentrated in the counting and measuring steps; the judgment steps — scope review, spec reading, subcontractor coordination — remain with the estimator.

Questions estimators actually ask