From drawings to BOQ in under a minute — inside the dual-path AI pipeline

A Bill of Quantities is only useful if you can trust the numbers. That is why CAIS does not treat quantity take-off as a single black-box prompt.

Two paths, one answer. Every drawing goes through two independent extraction paths at the same time. The image path rasterizes the sheet and visually counts elements — columns, footings, doors — the way a human estimator scans a plan. The text path reads the PDF's embedded text layer to pull schedules, dimension chains, and levels with high precision. We then merge the two: counts from the image path, dimensions from the text path. Where they disagree, the lower-confidence value is flagged for review.

Calculation is a separate step. Once elements and dimensions are extracted, a second AI pass does only the maths — applying the right formula per element type across nine BOQ sections (foundations, columns, beams, slabs, walls, finishes, doors and windows, waterproofing, and miscellaneous). Separating observation from calculation keeps each step auditable.

42 sanity checks. Before an Excel is produced, the result runs through 42 deterministic checks: do beam volumes reconcile with concrete totals? Does rebar tonnage fall in a sane ratio to concrete? Is any dimension missing rather than assumed? When a value cannot be found from any source, CAIS sets it to null and labels it DIMENSION_MISSING — it never invents a number.

BIM is even tighter. If you upload IFC or Revit instead of PDF, CAIS parses geometry and schedules natively and reaches 95–99% accuracy. The output is the same 13-sheet, FIDIC-compliant Excel either way.

The result: a draft BOQ in well under a minute for most projects, structured so a quantity surveyor can review and sign off rather than start from scratch.