Estimators spend much of bid time counting doors, measuring wall areas and typing quantities into a bill of quantities (Leistungsverzeichnis, LV). AI takeoff tools detect rooms, walls, openings and symbols on plan sets or read quantities from BIM models, propose a priced LV or a DIN 276 cost estimate, and leave checking and judgement to the estimator. Result: a first-pass takeoff in minutes instead of hours, so more bids can be answered with the same team.
Problem & value
- Today: contractors and trades (Handwerk) receive plan PDFs, sometimes with a GAEB tender LV, and measure on screen (Bluebeam, PlanSwift-style tools) or on paper, then price in AVA/estimating software (e.g. California.pro, Nevaris, ORCA AVA, RIB iTWO). Planners do the same by cost group (Kostengruppe) for early-phase cost estimates. The work is repetitive, error-prone and a bid bottleneck: many tenders are never answered for lack of estimator capacity.
- AI value: automatic recognition of rooms, wall types, openings and symbols; scale detection; mapping of measured elements to LV positions or cost groups; consistency checks against the tender LV.
- Evidence is mostly vendor-reported. Vendors claim accuracy of about 98 % for area takeoffs (e.g. Beam AI in software-directory listings; Beam itself cites results within ±1 % of in-house takeoffs for QA-reviewed work); an independent 2026 case study of Togal.AI against contractor takeoffs found small but statistically significant deviations, with consistent under-measurement of area quantities [3]. Treat figures as a first-pass speed-up, not as a replacement for review.
Variants
- Trade-specific takeoff (SMB): flooring, drywall, painting, roofing, electrical (counting sockets and luminaires), HVAC/sanitary symbols; typical for Handwerk bids.
- General contractor / main-trade bid: full building takeoff across trades, reconciled with the tender LV (GAEB X83) and priced via X84 offer.
- Planner early-phase estimate: quantities by DIN 276 cost group (KG 300/400) from area/volume indicators (BGF/BRI) and reference-project cost data, as a precursor to detailed LV.
- BIM-based 5D: quantities extracted directly from IFC/Revit models; AI mainly maps elements to positions and flags missing model information (LOD gaps).
- Automated vs. assisted takeoff: fully automatic first pass (duplicate variant) versus interactive assistant where the estimator clicks one example and the tool finds similar elements.
- Private use: homeowners estimating renovation costs from a floor plan photo; low accuracy, indicative only.
How it works
- plans (PDF/DWG/IFC) + tender LV (GAEB X83) + specification
- ingest: vector/raster separation, sheet classification, scale and title-block reading
- detection: rooms/zones, walls, openings, symbols (segmentation + object detection + VLM)
- quantification: areas, lengths, counts, volumes per element type and level
- mapping: elements
- LV positions / STLB-Bau text / DIN 276 cost group (LLM + retrieval)
- pricing: unit rates from own history, price database, supplier quotes; surcharges
- estimator review (overlay check, assumptions, risk items)
- GAEB X84 offer / estimate / Excel
- Detection: instance segmentation and detectors trained on floor plans; VLMs help with legends, notes, hatches and plan text. Vector PDFs and DWG give exact geometry; scans need more correction.
- Mapping and pricing: an LLM with retrieval over the tender specification and the firm's previous LVs proposes the matching position; price comes from the database, not from the LLM.
- Human-in-the-loop: the estimator confirms scale, checks overlays per sheet, resolves conflicts between plan and specification, and owns the final price and risk surcharge.
Data preparation
Needed are plans (PDF, DWG, IFC/BIM), tender documents, specifications, price and cost data (own unit prices, supplier quotes, BKI or similar catalogues) and finished past bills of quantities. In an SMB these sit on file shares per project, in e-mail attachments, in the AVA/ERP system and partly on paper. Typical gaps are scanned plans without scale, mixed plan versions, inconsistent layer names and legends, unit prices that exist only in old Excel sheets, and missing links between LV item and own cost position.
Preparation steps: 1. Organise project folders with one current plan set and version status (PREP-009, PREP-010). 2. Vectorise or OCR scanned plans and check the scale (PREP-006, DOC-014). 3. Turn price lists and Excel calculations into tidy tables with a unique item key (PREP-011, PREP-012). 4. Export past LVs as GAEB files to harvest reference mappings (PREP-020, PREP-005). 5. Keep 10-20 completed projects with actual quantities as an evaluation set (PREP-018).
Rough share of total effort: about 30-40 percent, mostly the cost-data clean-up. Data is mostly not personal.
Implementation paths
Which path for whom
- SMB: A — takeoff and estimating products with GAEB support are mature, available at SMB prices, and the process (LV from plans) is standard.
- Enterprise: B — estimating is core to margin and bidding, so integration with own cost databases, BIM models and ERP/AVA systems justifies building on APIs.
- Developer: B — cloud vision APIs plus own symbol and cost-data pipelines fit best; self-hosting needs HW3 and rarely pays.
Path A — Buy (recommended)
- AI takeoff tools: Togal.AI, Kreo, Beam AI, Bluebeam (Revu) with AI features, Stack, Procore estimating products, InEight Estimate; mostly US-built, check EU hosting and GDPR terms.
- DACH/EU AVA and estimating suites with GAEB: California.pro, Nevaris, ORCA AVA, RIB iTWO (Schneider Electric heritage), SIDOUN Globe (Freiburg), plus STLB-Bau / Dynamische BauDaten for standardised position texts [1][2]. BIM quantity tools such as Solibri, BIMcollab or dRofus-class tools for model-based takeoff.
- Choose A when: you bid regularly on typical building types and need GAEB import/export out of the box.
Path B — Build on cloud APIs
- Parsing: Azure AI Document Intelligence, Google Document AI or Mistral OCR for title blocks, schedules (door/window lists) and specification text; frontier VLMs (Claude, Gemini, GPT) for legends and symbol interpretation with structured outputs.
- Geometry: use vector data from PDF/DXF (e.g. ezdxf) and IFC (IfcOpenShell) where available, and let the model only classify and map.
- GAEB I/O: pyGAEB parses and writes GAEB DA XML 2.0–3.3 including X31 quantity data and has an optional LLM classification layer [4].
- Choose B when: you want own pricing logic or integration with an in-house ERP/AVA.
Path C — Self-host (open source)
- Detection/segmentation: YOLO-family or RT-DETR and SAM-class models fine-tuned on floor plans; Docling or PaddleOCR for text and tables; open VLMs (Qwen-VL family) via vLLM for legends and notes.
- BIM/geometry: IfcOpenShell, ezdxf, BlenderBIM for model quantities.
- Hardware: HW2 for detection and a mid-size VLM on a few hundred sheets a day; HW3 when running larger VLMs for specification reasoning. Ops effort is moderate; mainly relevant for large firms with confidential bids.
Path D — Custom training
Worth it when plans follow a firm-specific convention (e.g. a prefab builder, a utility, a retail chain with standard layouts). Needs a few hundred annotated sheets and corrected takeoffs from past projects; estimator corrections in the review step are the natural training data.
State of the art (2026)
- 2D takeoff: segmentation-based detection of rooms and walls on clean vector PDFs works well; vendors advertise minutes per plan set. Independent testing remains thin and shows systematic bias in some area quantities [3].
- Weak points: hand-drawn or low-resolution scans, inconsistent symbol libraries, scale errors, multiple details per sheet, hidden quantities (what is not drawn: fixings, waste, site setup), vertical elements across sheets, and interpreting specification text against drawing conflicts.
- BIM: quantity extraction from IFC is deterministic when the model is complete; AI helps where classification is missing. Quality depends on model discipline, not on AI.
- LLM role: position mapping and plausibility checks ("price per m2 outside range") are useful; numeric measurement by general VLMs alone is unreliable and should be backed by geometry.
- Standards: GAEB DA XML 3.3 is the data exchange format for tender, offer and quantity phases (X83 tender, X84 offer, X31 quantity determination); STLB-Bau supplies standardised position texts [1][2].
Shared resources & combinations
- Same stack (ocr-docparse + vlm + segmentation): DOC-013 technical drawing understanding, DOC-014 parsing for RAG, DOC-016 permit/rule checking on plans; one VLM server can serve all.
- Upstream/downstream: DOC-015 tender requirements extraction feeds the takeoff; VIS-040 scan-to-BIM produces as-built models for renovation takeoff; OPT-027 schedule optimisation and ROBO-009 progress monitoring use the same quantities for planning and tracking; DOC-006 for delivery-note matching on site.
Effort, cost & KPIs
- Path A: days to weeks; subscription per seat (some tools list low double-digit monthly prices per user [3]); estimator onboarding is the main effort.
- Path B: 2–4 months for a pipeline with GAEB import/export and own price logic.
- Path C/D: 4–8 months incl. annotation; justified only at high bid volume.
- KPIs: time per takeoff, bids submitted per estimator, deviation vs. final measured quantities (per trade), win rate, share of positions auto-mapped, missed-item rate found in review.
People & change
Involve management for the bid strategy and budget, the head of estimating as process owner, the estimators as daily users, IT for plan storage and ERP/AVA links, and data owners of the price databases. Include suppliers if their quotes feed the cost base.
Works council: depends. A takeoff tool is not meant to monitor staff, but cloud products often log user activity and by the BAG's standard objective suitability to monitor is enough for §87(1) no. 6 BetrVG. Clarify the logging and agree on no performance evaluation of individual estimators. Under §90(1) no. 3 BetrVG inform the works council in good time. Where an expert is needed, §80(3) BetrVG covers engaging one.
Data protection: mostly project data without personal data; check plans with names of owners or residents and sign an AVV (Art. 28 GDPR). A DPIA is normally not required. AI literacy (AI Act Art. 4): support estimators with training on checking measured quantities.
Pilot with one trade and one estimator team on closed bids, compare to manual takeoff. Concern: "the AI threatens the estimator's job". Answer: it removes counting, and the estimator keeps judgement and bid sign-off. See GUIDE-003, GUIDE-008 and GUIDE-009.
Risks, regulation & compliance
- EU AI Act: minimal risk; AI literacy duty (Art. 4) applies. No Annex III category, as no natural-person decisions are made [5].
- GDPR: plans normally contain no personal data; architects' names, private homes and site photos can; check vendor DPAs and EU hosting for confidential tender documents.
- Contract/IP: tender documents and plans carry copyright and confidentiality (NDA); check that vendors do not train on uploaded plans.
- Liability and commercial risk: the contractor remains responsible for the quantities and price in its offer (VOB/B, lump-sum contracts); a missed item or scale error can make a bid loss-making. Mitigate with mandatory review, plausibility checks against reference ratios and logged assumptions.
- Public tenders: LV text must not be altered; keep tender LV and offer consistent (GAEB X83/X84).
Getting started
- Pick one trade and ten past projects with known final quantities; build a ground-truth set.
- Test 2–3 tools on those plans; compare deviations per quantity type, not total time only.
- Check GAEB DA XML import/export with your AVA software and your price database.
- Define the review checklist (scale, levels, excluded areas, hidden quantities) and an estimator sign-off.
- Roll out on live bids with parallel manual check for the first 10–20 offers; track deviations.
Sources
- GAEB (DVGW/Bundesverband Bauindustrie) standards: GAEB DA XML — https://www.gaeb.de/
- Dynamische BauDaten / STLB-Bau overview — https://www.bauprofessor.de/gaeb/
- Togal.AI case study vs. contractor takeoffs (EPiC Series, 2026) — https://easychair.org/publications/paper/pkrRp
- pyGAEB documentation — https://pygaeb.readthedocs.io/
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Togal.AI on AI takeoffs — https://www.togal.ai/blog/construction-takeoffs-with-ai-speed-and-accuracy-combined