How AI Automates
Electrical Takeoff

What an electrical takeoff involves and why it is slow by hand

An electrical takeoff produces three categories of output: device counts (receptacles, switches, data jacks, lighting fixtures), linear footage (conduit, wire, cable tray), and equipment counts (panels, transformers, disconnects). Each category requires careful attention to a different type of mark on the drawings, which is why manual takeoff is hard to parallelize and hard to rush without making errors.

The scale of the problem on commercial work is significant. A single 50,000 SF floor can contain 800 to 2,000 device symbols. Counting and tallying those by hand on paper takes a senior estimator 8 to 20 hours per floor — before any conduit or wire measurement begins. On a multi-story building, the hours add up quickly, and that compression is exactly what forces many electrical contractors to decline bid invitations they could otherwise win.

The most common error sources are predictable: missing a symbol in a dense mechanical or plenum corner, double-counting a device that appears on both a demolition overlay and the new-work sheet, misreading a fixture type tag (A1 vs. A1E vs. A1EM), and omitting the vertical conduit drops that never appear in plan view but add real footage to the installed conduit total. Each error type has a systematic fix in the AI approach described below.

Step 1 — Plan ingest and sheet classification

The process begins when the PDF set is uploaded. The AI splits the file at sheet boundaries and classifies each sheet by discipline and sub-discipline. For an electrical scope it isolates the E-series: E0 (cover sheet and legend), E1 (lighting plan), E2 (power plan), E3 (systems and low-voltage), E4 through E6 (panel schedules, load calculations, and one-line diagrams).

Separating lighting from power plans at this stage is important. A ceiling-mounted fixture and a floor-level receptacle may carry the same symbol geometry on a small-scale printout; processing them on separate passes with discipline-appropriate legends prevents cross-contamination in the device counts.

Title-block and sheet-index OCR runs on every page to extract the sheet number, revision letter, and issue date. Superseded sheets — identified by a later revision of the same sheet number — are excluded from counting automatically. This handles the common situation where an addendum replaces a few sheets mid-bid and the PDF set contains both versions.

Step 2 — Scale detection and calibration

Every linear measurement the AI produces is only as good as its scale calibration. The model reads the printed scale notation from the title block — for example, 1/8" = 1′-0" — and then validates it independently against either a known dimension annotated on the drawing or the graphic scale bar printed in the margin.

A common problem in real bid sets is that sheets are printed at a non-standard reduction. A set originally drawn at 1/8" scale printed at 11x17 on a printer configured for architectural D-size ends up at roughly 56% of nominal. If the AI used the title-block annotation alone, every conduit run would be off by nearly half. To guard against this, the model re-derives scale from a dimensioned grid line or a column-to-column span that appears on the sheet, and flags sheets where the two methods disagree by more than a small tolerance.

Calibration is done per sheet rather than per set because different disciplines within a set are often plotted at different scales, and even within a single discipline the mechanical room enlargement sheet will be at a different scale than the floor plan. An error at this step multiplies directly into every linear measurement on that sheet.

Step 3 — Symbol recognition and reading the legend

With sheets separated and scales confirmed, the computer-vision layer runs symbol detection across each plan view. Electrical symbols the model identifies include: duplex and quad receptacles, GFCI and WP receptacles, single-pole and three-way switches, dimmer and occupancy-sensor switches, recessed and pendant fixtures, surface-mounted and track fixtures, exit signs, emergency lighting heads, smoke detectors, junction boxes, and panelboards.

Symbol detection alone is not enough — the legend must be read to resolve what each symbol actually represents. Type tags printed next to fixtures (A1, F2, EM, EM-B) refer back to the fixture schedule, which lists the manufacturer model, wattage, lamp type, and mounting height. The AI parses that schedule and builds a lookup table that maps each tag to a real product description, enabling the BOQ to specify actual line items rather than generic placeholder descriptions.

Homerun arrows and circuit number annotations are recognized to associate each device or group of devices with its branch circuit and panel. This cross-reference is used later to validate counts against the panel schedule: if the schedule shows Panel LP-2 with 42 circuits and the plan-view homeruns only account for 38, the system flags the discrepancy for estimator review.

Step 4 — Measurement and quantity computation

Device counts are tallied per type per area or zone. Linear conduit measurement follows the routed run paths shown on the plan, converted to real-world feet at the calibrated scale. But plan-view measurement captures only horizontal runs; the AI adds estimated vertical drops based on the device's mounting height and the ceiling or slab height annotated on the architectural drawings. A device specified at 18" AFF on a floor with a 10′ finished ceiling adds approximately 8.5 feet of riser conduit that would otherwise be invisible in a 2D takeoff.

Wire footage is derived from conduit footage rather than measured independently. The calculation multiplies conduit run length by the number of conductors per circuit: a 3-wire 120V branch circuit (line, neutral, ground) multiplies by three. The model then adds 2 to 3 feet of slack per panel termination and 6 to 12 inches per device termination to reflect real-world pulling practice.

Quantities are broken out by AWG size and insulation type — THHN, XHHW, MC cable — because unit material cost and labor unit vary substantially between them. Grouping 12 AWG THHN with 4 AWG THWN-2 into a single wire line item would produce an unusable BOQ.

Step 5 — Assembly mapping, waste, and BOQ output

Raw quantities are passed through an assembly layer that expands each device into its installed components. A standard duplex receptacle assembly, for example, pulls in the device itself, the coverplate, the rough-in box, box connectors, and fasteners. Labor hours are assigned using NECA labor units (2025 update), which have been the industry standard for electrical installation time estimation since 1923 and are updated periodically to reflect changes in materials and installation methods.

Wire footage receives a 5 to 10 percent waste factor applied after the slack allowance, accounting for offcuts at junction boxes and the inevitable pulling waste on long straight runs. The resulting quantity is then rounded up to the next standard reel size — 250, 500, or 1000 feet — because wire is purchased by the reel and a partially-used reel still represents a cost. Conduit is similarly rounded to standard 10-foot stick counts.

The final output is a line-item bill of quantities organized by CSI Division 26, exportable to Excel or directly importable into estimating systems. Each line item carries the symbol source, the sheet reference, and the circuit association so an estimator can trace any number back to its origin in the drawing set.

Step 6 — Estimator review and accuracy

AI-generated electrical takeoff is strong at the tasks that are tedious and error-prone by hand: high-volume symbol detection across dozens of identical floor plates, consistent conduit measurement without fatigue, and systematic application of assembly rules. Its weaknesses are concentrated in areas where the drawing quality degrades or where human judgment is required to interpret ambiguous intent.

On clean CAD-quality PDFs, device-count accuracy typically runs 95 to 98 percent. Conduit and wire footage is usually within a few percent of a careful manual measurement after scale verification. The most reliable cross-check is the panel schedule: if the schedule circuit count matches the plan-view homerun count, the takeoff is almost certainly complete. Where they diverge, the AI surfaces the discrepancy rather than silently reconciling it.

Accuracy drops on poor-quality scans, hand-sketched revisions taped to plan sheets, and legends that use non-standard symbols without a clear key. In those cases the model flags the problem areas and passes them to the estimator rather than making a low-confidence guess. The practical result is that review of an AI-generated electrical takeoff usually takes 1 to 3 hours — the estimator's attention is focused on the flagged exceptions rather than spread across the entire drawing set — compared with 1 to 3 days for a fully manual takeoff of the same scope.

• NEC (NFPA 70, 2023 edition) referenced for conductor and conduit requirements
• NECA labor units (2025 update) applied for installation hours
• Code compliance checks can flag undersized conductors or missing equipment grounding conductors

Questions estimators actually ask