Quantity takeoff from drawings

The job

At a $5–10M GC, takeoff is usually one estimator with a PDF set and a deadline: measure everything, twice if the addendum lands late. It is the slowest, most mechanical part of getting a bid out, and it caps how many bids the company can chase in a month.

“Done” for this job: a quantity sheet by trade, tied to sheet references, that the estimator has reviewed and would defend in a bid interview. The agent’s output is never “done” — it’s the draft the estimator marks up.

Use

Several products run computer vision over your drawing set and return editable first-pass quantities [1] [2] [3]. The workflow is the same regardless of which you pick:

1. Upload the set. Full architectural set, best available scale. Auto-detection quality tracks drawing quality — a fuzzy scan gives a fuzzy takeoff.
2. Run the automatic pass. The tool detects rooms, walls, and counts [1]. Treat the output as a junior’s first attempt, not a result.
3. Review by exception. Walk the set and correct: split, merge, and reclassify what the model got wrong. This is where the estimator’s hours now go.
4. Export to your pricing. Quantities out to Excel or your estimating system [1] — these tools do takeoff, not costing. Keep pricing where your cost history lives.

Vendor accuracy and speed claims are marketing until they happen on your drawings; independent research is thinner and more cautious than the ads [4]. The pilot below is how you find your own number. Published research recycled from the prior listing found AI plus manual cleanup about 70–76% faster and most adjusted quantities within 5% of On-Screen Takeoff on a small architectural sample; it also says not to bid the raw AI.

Build

A general-purpose model with vision can read plan sheets and produce door, fixture, and room schedules and rough area estimates from a PDF. It can help with counts and schedules, or provide an independent sanity-check against purchased output. Precise measured quantities need purpose-built tools; don’t bid off an LLM’s scale reading.

A workable v1: a prompt template that returns a CSV door/room/fixture schedule with sheet references, run per addendum. Keep the transcript with the bid file so counts are auditable.

Deploy

1. Pick a recently completed bid where you trust the numbers—your answer key.
2. Run the agent on that set and time upload, automatic pass, estimator review, and export.
3. Diff against the answer key by trade; record failures and whether review caught them.
4. Decide using a threshold set in advance. If it misses, stop and retry later.

Use your drawings, not a vendor demo set. Buy only if the full cycle returns time and the estimator will still check every sheet.

Manage

• Human review is a rule. No quantity leaves takeoff without an estimator’s pass.
• Confirm scale on every sheet. Auto-detection is a hint, not acceptance.
• Know the failure modes. Clean architectural floor plans are the best fit. Civil/site, structural, elevations, RCPs, poor scans, and busy annotations require more manual work [1].
• Do the seat math yourself. The published Growth price is $299/user/mo billed yearly. It must beat hours saved on your actual bid volume.
• Keep exports. Quantities can leave in a spreadsheet; classifications and chat history may not.
• Check data handling. Drawings can be under NDA. Review security claims and contract terms before uploading.