Is AI take-off as accurate as a manual take-off?
What our own test found
In October 2026 we ran real tender PDFs through QuantRead and compared the results with manual take-offs. The full details are in our case study. In short:
- On one tagged elevation, window, door and louvre counts matched the tags exactly.
- On a third elevation, the derived window and door perimeter was 221.0 m against a manual take-off of 238.9 m, about 92.5% of the manual metres.
- On that same elevation the opening count differed, because lobby and low-level units were grouped differently from the manual count.
That is a small test on a handful of drawings, not a general accuracy figure. Be wary of any tool that quotes one headline percentage for every drawing.
Manual vs AI, side by side
| Manual take-off | AI take-off | |
|---|---|---|
| Speed | Slow: every line clicked or scaled | Fast first pass |
| Judgement | Strong: knows what is implied | Weaker: measures what is drawn |
| Arithmetic errors | Possible in spreadsheets | Totals worked out in code |
| Consistency on repeat sheets | Tiring, error-prone | Consistent |
| Audit trail | Depends on the person | Every line labelled with its method |
Using AI without losing accuracy
- Let the AI do the first pass.
- Check every assumed and low-confidence line.
- Spot-check counts on each sheet.
- Add the implied items a person would know about.
- Check on site before ordering.
Tags: AI, accuracy, case study