How it works

Three stages. One narrow AI step.

Most BoQ tools are either pure AI — unreliable — or pure manual — slow. Finorly runs three deterministic stages with one narrow AI step in the middle, so you get speed without giving up auditability.

~30 seconds per drawingEvery item traceableNo hallucinated quantities
pipeline preview~30s
01ExtractRead text, tables, blocks, and geometry from PDF, DXF, or DWG.code
02ClassifyUse a narrow AI step to map each item to the right standard section.AI
03ValidateRun deterministic rules for arithmetic, flags, and source traceability.code
AI does classificationnever arithmetic
The pipeline

Three stages. The middle one is the only place AI runs.

Extract is code. Classify is a narrow AI model that maps to your standard. Validate is a rules engine. If any stage disagrees, you see a flag before a number lands in your BoQ.

STAGE 01

Extract

deterministic
TEXT · positionsTABLES · rowsGEOMETRY · polys

We pull text, tables, and geometry from your file with code — not AI. PDFs get positional text extraction; DXF files are parsed natively for schedules and attributed blocks.

PDFDWGDXF
→ no AI yet
STAGE 02

Classify

narrow AI
"Cast concrete C25""Excav. to RL"CLASSIFYai modelNRM2 § 11.3.1NRM2 § 5.2.4? flagged for review

An AI model maps each line item to your chosen standard. It classifies — it doesn't invent. Every suggestion is a draft until the rules engine validates it.

NRM2POMICESMM4SMM7
→ AI draft, never authoritative
STAGE 03

Validate

deterministic
qty × rate = amount ✓NRM2 deduction rule ✓wastage % mismatch ⚠ flagunits normalised ✓

A rules engine runs the measurement logic for your standard — deductions, wastage, two-of-three math (qty × rate = amount). Any conflict becomes a flag you can review.

rulesarithmeticflags
→ the math is always code
Inputs

File types we accept — and the ones we refuse.

We’d rather reject a file than guess at it. The extraction stage is deterministic; if there’s nothing to parse, we don’t rasterise and we don’t hallucinate.

PDF

Text-based PDFs

Exports from CAD or Excel. We extract text positionally — layout preserved, tables intact, no OCR guesswork.

Fully supported · 3–6 credits
DXF

AutoCAD exchange

Parsed natively. Schedules, blocks, and attributes are read directly from the file — no rasterisation, no conversion loss.

Fully supported · native parsing
DWG

AutoCAD native

Converted to DXF internally using the same engine as our free conversion tools, then parsed the same way.

Fully supported · via DXF
What we refuse: scanned PDFs, image-only files, pure-geometry drawings with no schedule data, and anything locked or encrypted. We’d rather refuse than guess.
Traceability

Every number has a source.

Click any line in the editor to see the exact region of your drawing it was extracted from. Nothing floats. If a quantity can’t be traced, it doesn’t ship.

BoQ editoritem 3.04 selected
#descriptionqtyunit
3.01Excavation to reduced level412.0
3.02Disposal of excavated material412.0
3.03Imported fill, well-graded148.5
3.04In-situ concrete, foundation slab62.4
3.05Reinforcement, high-yield bars5,820kg
3.06Formwork to foundations184.8
Every row → one click → the source polygon on your drawing.
click · trace
Source · drawing A-104foundation plan
A-104 · FOUNDATION PLAN
1:100
62.4 m³
↑ item 3.04
EXTRACTED REGION · VERIFIED
The exact polygon the 62.4 m³ was computed from.
The boundary

Where AI helps — and where it doesn’t.

The failure mode of every “AI BoQ tool” is asking the model to do arithmetic. We don’t. Classification is probabilistic; math is not.

AI · YES

What the model actually does

Narrow, bounded tasks where language ambiguity is real and rules would overfit.

  • Mapping descriptions to standard taxonomies (NRM2 / POMI / CESMM4 / SMM7)
  • Disambiguating unit abbreviations (m² vs sq m vs SQ.M.)
  • Merging duplicate items across cover sheets and detail pages
  • Suggesting sections when descriptions are incomplete
  • Confidence scoring so the editor surfaces shaky items first
AI · NO

What the model never touches

Anything that needs to be right, not plausible. Handled by a rules engine you could audit.

  • Arithmetic — qty × rate, totals, subtotals, deductions
  • Final item approval — always user-driven, never auto-confirmed
  • Standard rule application — that's code, not a model
  • Unit conversions and wastage calculations
  • Determining which items were extracted from which region
Standards

Four standards out of the box. Pluggable for regional variants.

Each standard is a module — taxonomy, column schema, and measurement rules defined separately. Pick one per project; your firm’s conventions apply at export.

+ regional variants on request
RICS · UK building works
NRM2 — New Rules of Measurement

The default for UK building projects since 2013. Work sections from groundworks through finishes, with standardised descriptions and quantity rules.

Example work section
§ 11 · In-situ concrete works · m³
§ 11.3.1 · Concrete foundations · grade C25 · m³
§ 15 · Structural metalwork · t
Column schema
Item · Description · Quantity · Unit · Rate · Amount
What you get

Fast enough to matter. Honest enough to sign off.

~30s
From upload to draft BoQ

Extract, classify, validate, flag. The draft lands in the editor before you've finished your coffee.

100%
Every line traceable

If the pipeline can't attach a source region and a standard rule to a number, it's flagged — not shipped.

0
Hallucinated quantities

The model never does arithmetic. Quantities come from geometry or schedules — or they don't appear at all.

Before you upload

A few sensible questions.

Full answers on the FAQ page →

What if the AI classifies something wrong?+
Every item is editable. The editor shows confidence scores and surfaces low-confidence items first, so you review the shaky ones — not all of them. Your corrections don't leave your workspace.
Can it handle our firm's custom template?+
Yes. The standards layer handles classification; export templates are separate. Upload yours once and every export uses it — XLSX, CSV, or PDF.
What happens to my files?+
Deleted after processing unless you save to a project. We don't train on your data. EU-hosted. If you need a signed DPA before uploading, email us — we'll send one the same day.
How does it handle scanned or image PDFs?+
It doesn't. Scanned PDFs are rejected at upload with an explicit message. Our extraction is deterministic — if there's no text layer, there's nothing honest to extract.
Can I add a custom regional standard?+
On Business plans, yes. Each standard is defined by taxonomy, column schema, and measurement rules — we can add regional variants without rewriting the pipeline. Usually a two-week turnaround.
01 · EXTRACT02 · CLASSIFY AI03 · VALIDATE~30 SECONDS · FULLY AUDITABLE
See it run

Try the pipeline with your own drawing.

Upload a PDF, DXF, or DWG and see extraction, classification, review flags, and export path in one run.