AI for Quantity Take-Offs - Step-by-Step

Tim Fairley · 20:37 · Construction estimator revises his stance: with Claude Opus 4.8-era models + a packaged Claude skill, AI quantity take-offs are finally usable — but only inside a workflow designed around what AI measures reliably.

The revised position

• Old advice: AI helps decide what to count, set up assemblies, compile the estimate — but you measure and count yourself.
• What changed: modern models can write and run their own code — instead of visually counting footing symbols, they extract text tags from the PDF and count them programmatically (100% accurate).
• Standing caution: ~80% of direct costs scale directly with quantities; a bad take-off → under-quoted project. For a lump-sum bid on a 1M–10M job, still check quantities by hand (a couple of hours).
• Sweet spots: quick conceptual estimates for go/no-go decisions, and checking a take-off you already did.

Tool architecture

• Runs in Claude Cowork (Claude opened on a desktop folder — reads/writes all files in it). Drawing sets live there: structural, plumbing, architectural, mechanical.
• Whole workflow packaged as a skill (/construction-takeoff): Python scripts for splitting PDFs and extracting drawing sections, plus built-in checklists per trade (architectural / civil / mechanical / electrical) and the primary-vs-secondary quantity strategy.
• Prerequisite step: run his drawing analyzer skill first — splits the set into individual PDFs and builds markdown files + a structured database of every object on the drawings, so AI has a map of where information lives.
• Why Cowork and not chat: an 11MB drawing set simply won’t upload in Claude chat (“file is not supported or may be corrupted”) despite being under the stated limit. Desktop app handles big documents.

Three-step workflow

1. Analyze & plan — AI asks which trade you’re working (you don’t want structural steel extracted when doing concrete), then proposes assemblies: what to measure as primary quantities.
2. You review the plan — critical human-in-the-loop. Narrow scope to what AI does reliably (text-tag counting, extracting written dimensions); strike anything it would do badly (scaled measurements).
3. Measure & output — runs the measurements with the most reliable method available, outputs in a fixed template every time so results are familiar and auditable.

Primary vs secondary quantities

• Don’t hand AI the whole messy task. Pick few, well-defined primaries: slab area, footing count. Count footings via text tags, read sizes/types/rebar from schedules.
• Derive secondaries by formula from primaries: P1 slab measured at 11,640 m² → concrete volume from the depth noted on drawings. No scaled section-view measuring at all.
• Example scoping call: electrical — count power points (reliable text counting), derive cable with a flat ratio like 15 m per point, because cable is cheap and fast to install.

Reliability: vector vs raster

• Vector PDFs: text is selectable/highlightable → tag extraction works.
• Compressed-image PDFs: nothing extractable → the skill falls back to vision counting and flags it low confidence.
• Every measured quantity gets a confidence score, so you know exactly which lines to re-check — a cheap self-audit.
• Honest findings: text-tag counting “unbelievably accurate”; slab areas are mixed (perfect when length×width is written on the drawing, shaky otherwise). Marking up drawings with counted quantities is hit-and-miss — though one wrong slab markup correctly outed a bad section.

Sanity checks beat precision fixes

• Built into the skill: after measuring, verify orders of magnitude — cable tray runs the full width of Block C three times → total should be roughly 3× that width. 60m vs a 10m answer = hallucination caught.
• Won’t fix small precision errors, but kills the big blowups.

Building your own version

• Put past take-offs + their drawings in a Cowork folder, describe this architecture (primary/secondary split, vector-over-vision, confidence scores, sanity checks, fixed output template), and have Claude Code iterate with the goal feature: set an objective, let it test, diagnose its own misses, improve.
• Or grab his prebuilt skill via his Contractor OS membership (pitch).

Bottom line

Don’t ask “can AI do take-offs?” — ask “which measurements can AI do near-perfectly (tag/dimension extraction) and which need my eyes (scaled areas)?” Structure the workflow, output template, and confidence reporting around that line.