Architectural Drawing Data Extraction with AI
Dimensions, door and window schedules, room areas, symbols and annotations — pulled off the sheet as structured data, including from scans and drawings nobody drew in CAD.
- Reads dense, marked-up and scanned sheets, not just clean CAD exports
- Trained on the conventions your own office draws to
- Every figure checked back against the source sheet
- Projects delivered
- 50+Projects delivered
- Formats accepted
- PDF, CAD, scanFormats accepted
- Figures traced to source
- Per sheetFigures traced to source
- Reply from the engineer
- 24hReply from the engineer
What is architectural drawing data extraction?
Architectural drawing data extraction is software that reads a drawing the way a person reads it — finding the dimension strings, the schedules, the symbols and the annotations — and writes what it finds into a structured form: a spreadsheet, a database, or straight into whatever system needs the numbers.
It is not the same problem as reading a document. A drawing has no reading order. Information sits in tables, in leader lines pointing at things, in symbols that mean something only by reference to a legend, and in handwriting somebody added on site. Text extraction alone gives you a pile of disconnected strings; the work is in knowing which dimension belongs to which wall.
The reason it is worth doing is that the numbers get retyped anyway. Someone takes off quantities, someone builds a door schedule, someone transfers room areas into a valuation model. That retyping is where errors enter, and it is slow enough that it is often the reason a bid goes out late.
What we do not claim is that it removes the checker. A figure taken off a drawing wrongly is worse than one not taken off at all, so every extracted value carries a reference back to the sheet and the position it came from, and low-confidence readings are flagged rather than filled in.
How it works
- 1
Send the sheets
PDF, CAD export, or a scan of a paper drawing. Whole sets are fine; we work out the sheet types.
- 2
Read the drawing
Dimension strings, schedules, symbols and annotations are located and linked to what they refer to.
- 3
Structure it
Values are normalised into consistent units and field names, so two sheets drawn differently come out the same.
- 4
Check and hand over
Each figure carries a link back to its position on the sheet, and anything uncertain is flagged for review.
What gets extracted
| Field | Example | Confidence |
|---|---|---|
| Room name and area | Kitchen / Dining — 24.6 m² | 0.95 |
| Dimension strings | 3600 / 1200 / 3600 to gridline C | 0.93 |
| Door and window schedule | D-04, FD30S, 926 × 2040 | 0.92 |
| Symbol counts | 18 double sockets, 6 downlights | 0.89 |
| Annotations | "Verify on site" against the stair core | 0.87 |
| Revision and sheet metadata | Rev C, issued for construction | 0.96 |
We have built this
We built a floor plan analyser that reads scanned plans and returns rooms, dimensions and areas as structured data, and a separate pipeline for technical engineering drawings where combining vision preprocessing with few-shot prompting lifted exact-match accuracy by 36 percentage points. Both case studies cover the failures as well as the results — in particular what happens when a drawing's legend does not match the symbols actually used on the sheet.
Read the full case study →Manual take-off vs generic OCR vs purpose-built
| Manual take-off | Generic OCR | Purpose-built | |
|---|---|---|---|
| Time per sheet set | 8-16 hours | Minutes, then hours fixing | Minutes |
| Understands the drawing | Yes | No — returns loose text | Links values to what they describe |
| Scanned and marked-up sheets | Yes | Poorly | Yes, that is the normal case |
| Traceable to source | In the estimator's head | No | Every figure, to sheet and position |
| Consistent across offices | Varies | No | Same conventions applied every time |
Delivers into: Revit · AutoCAD · Bluebeam · Excel · Power BI · REST API
Frequently asked questions
Does it work on scanned paper drawings?
Yes, and that is usually the reason people call. A clean CAD export is the easy case and plenty of tools handle it. The hard case — a scan of a drawing from 1994 with site annotations in biro — is the one that actually blocks work, and it is the case we built for. Accuracy is lower on those sheets than on native PDFs, and we will show you the difference on your own drawings rather than quoting you one number.
What about drawings that do not follow a standard?
Every office has house conventions, and most sets have at least one sheet that breaks them. We train on your drawings rather than on a generic corpus, which is why the assessment starts by asking for a representative set — including the awkward ones, not the tidy ones you would normally show a supplier.
Can it do quantity take-off?
It can produce the measured quantities that a take-off is built from — areas, lengths, counts by symbol. Whether that amounts to a take-off depends on your pricing and build-up, which is yours and not something we would want to guess at. Most clients take the structured output into their existing estimating tool rather than replacing it.
How do you handle a wrong reading?
By making it visible rather than plausible. Every value links back to the exact position on the sheet it came from, so a checker confirms it in a glance instead of re-measuring. Anything the model is unsure about is flagged rather than filled in, because a confident wrong dimension is the one failure mode that would make the whole thing a liability.
Do you need the CAD files?
No. If you have them, the results are better and faster, because the geometry is exact rather than inferred. But the whole point of this is the sets where CAD no longer exists, or was never issued to you, and those work.
Where does the output go?
Wherever you need it. Spreadsheet is the common answer, Revit and Bluebeam for teams already working there, or a REST endpoint if you have your own system. We do not ask you to adopt a new tool to read the output.
What does a project look like?
It starts with a free assessment: send a representative set of drawings and we will extract from them and show you exactly what came out, what did not, and the accuracy against your own figures. You get that within 24 hours, from the engineer who would build it. If the sheets turn out to be too poor to read reliably, we will tell you that instead of selling you a pilot.
What manual take-off is costing you
See what manual data entry is costing you
Rough estimate at ~90% automation over 220 working days a year. Your exact numbers come from the free assessment.
Get a Free Assessment
Send us a few of your documents and we'll tell you exactly what can be read from them, how accurately, and how much time it would save. No commitment.
See what manual data entry is costing you
Rough estimate at ~90% automation over 220 working days a year. Your exact numbers come from the free assessment.
Prefer to talk live? Book a call directly.
No preparation needed · Get actionable insights · Zero obligation
