How AI Automates
Earthwork Takeoff

What an earthwork takeoff involves and the manual pain

Earthwork takeoff is, at its core, a surface-differencing problem. The finished deliverable is a set of volumes — cut (material to remove), fill (material to import or reuse), net import or export, topsoil strip and respread, and separate trench excavation for utilities — all expressed in cubic yards. Those volumes drive hauling cost, equipment selection, compaction crew hours, and disposal fees, so errors compound quickly into bid-day surprises.

The volume itself comes from comparing two surfaces: the existing grade (as-surveyed) and the proposed finished grade shown on the civil plans. Every cell in a grid across the site contributes a small wedge of cut or fill based on the elevation difference at that point. Swell and shrink factors then adjust raw volumes to reflect how soil behaves when excavated (it loosens and expands by roughly 25% for common soil) or when compacted under fill (it densifies by 10–15%), producing the quantities a PM actually orders trucks for.

Manual earthwork takeoff is extremely slow. Generating a grid of elevations from hand-scaled contours, differencing them, applying soil factors, and balancing the mass diagram can take 15–40 hours on a mid-size commercial site. That time burden is why many sitework estimators only bid jobs where they have prior survey data or a strong relationship with the GC — not because the work is bad, but because the takeoff clock is brutal.

Step 1 — Plan ingest and sheet classification

Before any measurements happen, the AI needs to know which sheets are relevant to earthwork. A commercial bid set might contain 80–120 sheets covering architectural, structural, mechanical, electrical, and civil work. The AI classifies each sheet and isolates the three that drive earthwork: the existing conditions or topographic survey, the grading and drainage plan, and the utility plan (for trench excavation).

Within those sheets, the AI tags the major elements: the site boundary that defines earthwork limits, building pad areas and their finished-floor elevations, pavement areas and their subgrade depths, and any detention basin or retaining wall extents that bound a cut or fill zone. Scoping these limits correctly is critical — an earthwork takeoff that bleeds outside the contract boundary, or misses a basement, produces a number that's wrong from the start.

Step 2 — Scale detection and calibration

Civil grading plans commonly use engineering scales — 1″ = 20’, 1″ = 40’, and 1″ = 50’ are typical — rather than the architectural scales estimators encounter on floor plans. The AI reads the scale bar and title-block annotation to set horizontal calibration, then validates it against a dimensioned property line or easement callout elsewhere on the sheet. If the two don't agree within a tight tolerance, it flags the discrepancy rather than silently proceeding.

Vertical calibration is handled separately. The AI picks up the datum — typically NAVD 88 or a project benchmark elevation — from labeled contour lines and spot elevations. Horizontal scale tells the system how large the site is in plan; vertical datum tells it what the actual elevation numbers mean. Getting both right before any surface is built is what keeps computed volumes in the right order of magnitude.

Step 3 — Surface recognition and reading elevations

With scale and datum established, the AI builds two triangulated irregular networks (TINs): one for the existing grade and one for the proposed finished grade. Each TIN is built from two types of input. Contour lines are traced geometrically — the AI follows the polyline for each labeled contour interval and assigns the elevation from its annotation. Spot elevations are read by OCR directly from the plan, paired with their X/Y coordinate based on the leader or dot location.

For the proposed surface, the AI also picks up grading callouts — percent slopes, swale inverts, and curb and gutter flow-line elevations — and finished-floor elevations on the building pad. These constrain the proposed surface in areas where contour lines are sparse or where design intent (a flat parking field, for example) isn't fully expressed in contours alone.

Topsoil strip depth and subgrade preparation notes are read separately and flagged as a distinct volume layer, since stripping is priced differently from general excavation and respread happens at a different project phase. Similarly, utility trench widths, depths, and lengths are captured from the utility plan to produce trench excavation as an independent line item.

Step 4 — Measurement and quantity computation

With both surfaces built, volume computation follows the grid-differencing method. The site is overlaid with a regular grid — cell size is typically 5’ or 10’ depending on contour density and required precision. At each grid node, the AI interpolates existing and proposed elevations from the respective TINs and takes the difference. Positive differences (existing above proposed) are cut; negative differences (proposed above existing) are fill. Each cell's contribution is its elevation difference times cell area, divided by 27 to convert cubic feet to cubic yards.

Raw cut and fill volumes are then adjusted for soil behavior. A swell factor — commonly 25% for ordinary soil, higher for clay, lower for sand — increases the hauled volume of cut material, since it occupies more space in a truck than in the ground. A shrink or compaction factor — commonly 10–15% — increases the volume of fill material that must be imported to achieve the specified compacted depth. Net earthwork equals required fill minus usable cut after these adjustments, yielding the import or export quantity the PM will price.

• Grid cell area × elevation difference ÷ 27 = cubic yards per cell
• Cut (hauled) = raw cut × (1 + swell factor)
• Fill (required) = raw fill ÷ (1 − shrink factor)
• Net = fill required − usable cut → import (+) or export (−)

Step 5 — Assembly mapping, waste, and BOQ output

Raw volumes mean little to a bid until they're mapped to cost-bearing work items. The AI assigns each volume component to a CSI Division 31 assembly: general excavation, structural excavation for footings, backfill and compaction, hauling and disposal, topsoil stripping, topsoil respread, trench excavation, and trench backfill. Each assembly carries its own waste and over-excavation factor — trench walls need working room, building pads need a compaction bench — which the AI applies based on soil type notes and specification references read from the plans.

The output is a clean BOQ in cubic yards with separate line items for cut, fill, net import or export, topsoil strip, topsoil respread, and trench excavation by utility type. Over-excavation and backfill under slabs are broken out if the structural drawings reference a minimum bearing depth or soil replacement spec. The BOQ exports directly to Excel so the estimator can attach unit rates, equipment spread sheets, and trucking costs without reformatting.

Step 6 — Estimator review and accuracy

AI is genuinely strong on the core surface-differencing task when the civil drawings are reasonably legible: clean contour surfaces on well-scanned PDFs typically produce volume accuracy in the 90–97% range relative to a careful manual takeoff. The accuracy band widens when contour density is low (large contour intervals leave the TIN underdetermined), when survey data is illegible or printed at poor resolution, or when the grading plan uses non-standard symbology that the model hasn't seen at high frequency.

There are also scope items that AI cannot infer from grading sheets: subsurface rock, groundwater depth, buried obstructions, and soil bearing assumptions. These are not estimation failures — they are genuinely absent from the plan set and require geotechnical reports or site knowledge. The AI surfaces them as explicit flags in the output so the estimator knows exactly what judgment calls remain.

Estimator review of an AI-produced earthwork takeoff typically runs 2–4 hours, compared to 15–40 hours fully manual. The time is spent checking scale calibration on a few control dimensions, validating the surface extent against the boundary, and resolving the flagged risk items. The estimator doesn't rebuild the takeoff from scratch; they audit a completed draft.

Questions estimators actually ask