APPLIED AI · CONSTRUCTION

Construction Take-off Suite

A drawing-to-BOQ/BOM platform for tender take-off. Deterministic measurement with full provenance, deployed across 4 live GCC projects.

YEAR

2026

INDUSTRY

Construction & tender pricing

SERVICE USED

Applied AI, deterministic pipelines, systems integration

CHALLENGE

Priceable bills with anti-fabrication guarantees

(QTF® — THE PROBLEM)

Tender take-off is the most expensive repeated manual exercise in construction. Two failure modes dominate: silent quantity error and silent scope error — and existing AI take-off tooling addresses neither honestly.

A quantity surveyor reads several thousand drawings, spreadsheets and specification pages, measures every element by hand, and re-does a large fraction of that work every time an addendum lands. Existing tooling produces a number without saying where the number came from — and stays quiet about what it could not measure. We built the opposite.

The Automated Construction Take-off Suite is a client-agnostic pipeline that ingests a raw tender dump in whatever structure the consultant sent it, and produces a priced-format Bill of Quantities, a procurement Bill of Materials, and a written account of everything wrong with the issued documents. Ten stages, each with declared inputs, outputs and failure behaviour.

What the pipeline actually does

Intake and inventory. CAD-to-vector conversion and multi-sheet PDF splitting. Sheet identity resolved from title blocks, never filenames, with an explicit revision supersede graph. Semantic extraction of grids, level datums, layer roles, legends and room polygons into a queryable project graph. A completeness engine that asks whether the issued set is actually complete enough to price. Twenty-five named consistency checks — dimension chains, tag-vs-schedule tallies, cross-discipline geometry, BOQ↔spec↔drawing triangulation. A rule-pack compiler that freezes the measurement standard and client amendments as an immutable versioned bundle. Deterministic per-trade measurement, where every quantity carries a formula and a source. Assembly into the client’s own workbook format, then a branched BOM layer with two-way reconciliation. Finally an independent audit loop and a revision watch that turns each addendum into an hour-long delta bill, not a re-run.

Two design rules make it client-agnostic rather than client-specific. Consultant conventions live in a per-project profile — title-block geometry, sheet-number grammar, revision tokens, layer dictionary. What a project type requires in order to be priceable lives in a versioned required-set ontology. Onboarding a new consultant is a configuration pass, not a code change.

Security and confidentiality

Tender drawings are among the most commercially sensitive documents a contractor handles. Confidentiality and evidential integrity were designed in, not bolted on.

• Local-first by default. CAD conversion, OCR and measurement run on machines under our control with locally installed engines. No drawing is uploaded to a third-party conversion or OCR service.
• Consent-gated cloud fallback. Where a cloud service is the only route for a stubborn file, it is reachable only behind an explicit per-project consent flag recorded in the project profile.
• Isolated per-project workspaces. Each engagement has its own corpus directory, artefact store, and dependency environment. No shared scratch space between clients.
• SHA-256 hashing at intake. Every ingested file’s hash is its identity for the rest of the pipeline. Any substitution or truncation is detected, not assumed.
• Client IP stays client-scoped. Consumption norms, unit rates and purchase-department sheets are per-client artefacts with their own sign-off gate.

Anti-fabrication controls

The commercial risk in automated take-off is not a hacked file — it is a confident wrong number. These controls are what the suite exists to enforce.

• Deterministic-only computation. No language model computes, adjusts or rounds any quantity. Machine learning is confined to classification, table normalisation, description phrasing and narrative drafting — and is fenced by a verbatim-numeric guard: every number in an output must exist in the input, or the step fails hard rather than inventing one.
• Full provenance on every line. Each quantity carries file, sheet, revision, page bounding box or CAD entity handles, extraction method, formula used, and a confidence class — MEASURED, SCHEDULE, DERIVED, INFERRED or MANUAL.
• Independent mechanical audit. A separate process with a fresh context reads only persisted artefacts and raw sources. It re-executes the entire check suite (results must match bitwise), dereferences 100% of evidence pointers, and independently re-derives every material or contested line.
• Canary runs against rubber-stamping. Deliberately perturbed shadow bills are injected. An auditor that fails to catch a canary invalidates its own run.
• The false-absence rule. Claiming a document was not issued when it was is our highest-severity internal failure class. No such claim may be made without a corpus-wide corroboration pass across filenames, CAD text, PDF text and OCR — and the claim must be scoped to exactly what was searched.
• Never measure pixels. Raster-only sheets are OCR’d for text and flagged. Geometry is never measured from a scan — the system raises a document request instead.

Results

Client A — industrial and laboratory facility, GCC free-zone. Pilot proven end-to-end: schedule harvest, CAD geometry measurement, deterministic rules engine, client-format bill renderer, true-BOM generation. Acceptance tolerances met on the benchmark trades (±3% concrete and formwork, ±5% rebar). All four deliberately seeded cross-discipline discrepancies detected with dereferencing evidence, including a plant-room area conflict across three disciplines and a tank-capacity conflict between schedule and specification.

Client B — G+2P+8 residential tower tender, roughly 1,557 documents across five volumes and three addenda. Full package ingested and resolved: 553 sheet instances across ~500 sheet numbers. A 6.3 GB duplicate archive identified as an exact re-zip of the package itself.

• 1,466 bill lines / 5,934 measured items delivered as a 22-sheet consolidated workbook in the client’s own layout — 979 machine-quantified, 487 explicitly flagged MANUAL with a stated reason.
• Nine sampled section quantities were independently re-derived from primary sources; all nine matched the persisted calculations.
• 216 documented findings supported a 165-item RFI schedule. A 32-item technical query register (3 critical, 24 major) and a 21-item drawing-defect register with 57 drawn-versus-scheduled conflicts carry rendered CAD and PDF evidence figures.
• Priced-consequence errors caught that a manual read had missed: four CAD files declaring the wrong drawing units, a finishing schedule issued under another project’s plot header, pile-cap and ground-beam schedules cited by sheets that were never issued.
• Regression suite of 440 automated tests, with every corpus finding preserved as a fixture — including the false claims we caught ourselves making.

The bill is permitted to be incomplete. It is not permitted to be wrong.

Two structurally dissimilar clients — different consultants, different naming conventions, different disciplines, different output formats — ran through the same engine with no change to engine code. Only a new profile and ontology entries.

What a client receives

• Bill of Quantities in their own workbook format, every line traceable to a source.
• Bill of Materials, reconciled two-way against the BOQ.
• Findings register and a consultant-ready clarification schedule with drafted question text and evidence figures.
• Basis of estimate — measurement rule-pack version, governing document set, and the full assumptions ledger.
• Audit certificate and residual-risk list: what was verified, how, and what remains open.
• Revision watch: the next addendum arrives as a delta bill, not a repeat exercise.

The deliverable is a draft with provenance, reviewed and signed by a human quantity surveyor. That boundary is deliberate and permanent — the machine measures and evidences, a qualified QS signs.

(QTF® — CLIENT REVIEW)

The bill of quantities used to be a black box we hoped was correct. Now every line has a source, and the queries come in before the price does. That is a different conversation with the consultant.