Technical building services (TGA: heating, ventilation, cooling, sanitary, electrical) is the largest time sink in many engineering offices: room data are typed into calculation tools, products are picked from data sheets, schematics are drawn and every architect change ripples through the plans. A copilot can pre-fill the calculation tools, check results for plausibility, compare products and find clashes today. Automatic routing and drawing are still early. Certified calculations stay in approved software, and the licensed engineer signs.
Problem & value
- Today: the planner reads architect plans or an IFC/Revit model, re-enters rooms, areas and constructions into heat-load, pipe-network and energy-proof software, selects products from PDF data sheets, draws schematics and layouts, coordinates in the federated BIM model and updates plans after each architect change. In the underlying gap analysis for engineering offices the biggest blocks are schematics and layouts, sizing of ducts/pipes/cables, heating and cooling load, plan updates after changes and BIM coordination; SHK firms mostly need the heat-load part.
- AI value: less re-keying, earlier error detection, faster product comparison, less rework after changes. Realistic range for a split "calculation pre-fill and checking" scope is 10-30 percent of the affected time; routing and drawing automation is not yet reliable enough to count on.
- Evidence level: vendor material is mostly promotional; independent field studies are scarce. Measure on your own projects.
- Not the same as: quantity takeoff from plans (DOC-030), rule checking of permit documents (DOC-016), reading existing drawings (DOC-013), operating-phase HVAC control (OPT-007).
Variants
- Calculation pre-fill: room book and building data from IFC/plans into heat-load and energy tools. Feasible today in parts.
- Result plausibility check: LLM/rule checks on loads per m2, flow velocities, pressure drops, missing rooms. Feasible today.
- Product selection: compare data sheets (PDF, VDI 3805 / ETIM) against requirements.
- Clash detection and change impact: rule-based clash tools exist; AI summarises and prioritises clashes and carries architect changes into the MEP model.
- Routing and schematics: generative routing of ducts, pipes and cable trays, auto-generated schematics. Early; beta tools exist, but no verified production evidence.
- SHK craft variant: heat-load and design for retrofit boilers/heat pumps, including hydraulic balancing documentation.
- Private use: none relevant; homeowners use simplified online tools.
How it works
- architect plans / IFC model / room book
- intake (IFC parse, PDF/DWG plan reading by VLM, requirement extraction)
- data mapping (rooms, zones, constructions, use profiles
- calculation tool fields)
- pre-fill of heat/cooling load, pipe/duct sizing and energy-proof tools via import/API
- deterministic calculation in approved software (DIN EN 12831, DIN V 18599)
- AI plausibility check (outliers, missing rooms, norm limits) + product shortlist from data sheets
- draft variants / schematics / routing proposal (human selects and corrects)
- clash detection in federated BIM model + change-impact notes after architect updates
- engineer review and sign-off, release via BIM coordination (BCF issues)
- Key principle: the AI prepares and checks; every number that ends up in a proof comes from validated calculation software, not from the LLM.
- Extraction: IFC is structured (parsed with open libraries); PDF/DWG plans need a VLM or document parser and are the error-prone input.
- Routing/drawing: research and beta tools combine search/optimisation with clash checks; output is a proposal in the native model, never a released drawing.
- Human-in-the-loop: pre-filled data are reviewed per room; calculation results, product choice, routing and schematics are approved by the responsible planner.
Data preparation
Which data: architect models (IFC, Revit) and plans, room books, building physics data, product data sheets and manufacturer libraries (VDI 3805, ETIM/ECLASS), past projects with calculations, company standards and symbol libraries.
Where it lives: project folders on file servers, BIM collaboration platforms, the TGA calculation tools (project files in proprietary formats), the CAD/BIM system and in the heads of senior planners.
Typical gaps: architects deliver 2D PDFs or poorly structured IFC (missing room volumes, wrong space types, no U-values), manufacturer data without usable properties, calculation projects that cannot be read in bulk, inconsistent layer and naming conventions, and no BIM execution plan (BIM-Projektabwicklungsplan) defining what is exchanged when.
Preparation steps: 1. Agree an information requirement with architects: IFC with spaces, constructions and use profiles (PREP-015). 2. Standardise symbol, naming and layer conventions in the office (PREP-019). 3. Collect past projects with verified results as test cases, anonymised for client data (PREP-018). 4. Write down sizing rules and company checks in retrievable form (PREP-013, PREP-014). 5. Build a product database from manufacturer data (VDI 3805, ETIM) (DATA-010). 6. Restrict access to confidential client projects before enabling cloud features (PREP-023, PREP-024).
Rough share of effort: about 40 percent of the project; interfaces and data quality dominate. See GUIDE-003.
Implementation paths
Which path for whom
- SMB — Path A: small offices and SHK firms get the best value from the AI and import features of the TGA software they already use (heat-load, pipe-network, BIM add-ins); building custom pipelines is out of reach.
- Enterprise — Path B: larger offices have project standards, product databases and BIM workflows that must be wired together through APIs (IFC, Revit/Dynamo, BCF) with cloud models.
- Developer — Path C: open IFC tooling (IfcOpenShell) with open-weight LLM/VLM on HW3 gives control over client geometry and is the base for product builders.
Path A — Buy (recommended for most SMBs)
- TGA calculation and BIM suites (DE): Hottgenroth, Solar-Computer, Linear, Dendrit, ZVPlan, Autarc, MagiCAD (add-in for Revit/AutoCAD), Autodesk Revit with its MEP and generative-design tools. Solar-Computer advertises an AI assistant and "BIM in the TGA in the AI era" for 2026; verify concrete functions in a demo.
- Choose A when: you already use one such suite and its import interfaces (IFC, Revit) cover your clients' data. Check EU hosting, whether project data train vendor models, and certification of energy-proof modules.
Path B — Build on cloud APIs
- Pre-fill and check: VLM/LLM (Claude, Gemini, GPT, Mistral for EU hosting) with structured outputs for plan and data-sheet reading; results written into calculation tools through import files or APIs.
- BIM integration: IFC via open libraries, Revit API/Dynamo, BCF issues for clash handover.
- Orchestration: n8n or custom agents with tool calls to calculation engines; deterministic checks stay in code.
- Choose B when: you have a stable project standard and a developer team to maintain the integration.
Path C — Self-host (open source)
- Stack: IfcOpenShell, BlenderBIM/Bonsai, OpenStudio/EnergyPlus for simulation, open VLMs (Qwen-VL) and open-weight reasoning models, a vector store for norms and company rules.
- Hardware: HW3 for VLMs reading dense plans; HW2 suffices for IFC-only workflows with small LLMs.
- Effort: months; open tools do not replace DIN-certified calculation software.
Path D — Custom training
Rarely sensible. Fine-tuning a symbol-recognition or routing model needs thousands of consistently drawn schematics and layouts; few offices own such curated data. Reinforcement-learning routing remains research.
State of the art (2026)
- Mature and certified: deterministic calculation (DIN EN 12831 heat load, pipe-network sizing, DIN V 18599 energy proof) in specialised software; rule-based clash detection in BIM tools.
- Practical now: LLM/VLM assistance around these tools: reading plans and data sheets, mapping data, summarising clash and change lists, checking plausibility.
- Early: generative MEP routing. A much-cited system (GenMEP, generative routing with clash detection on IFC/Revit) was still in beta in its public description, and I found no verified 2026 product with independent results (a 2020s partner page describes a GenMEP Revit add-on with autorouting, but that is vendor material). Vendor claims such as a 43 percent time reduction come from vendor glossaries and are unverified.
- Limits and failure modes: wrong room geometry from poor plans, hallucinated product properties, routes that are clash-free but violate fire protection, accessibility, maintenance zones or acoustic rules, and silently changed inputs after architect updates. Geometric validity is not functional correctness.
- Practical rule: split the case. "Calculation pre-fill and checking" is feasible; "routing and drawing" is a pilot topic.
Shared resources & combinations
- Same stack: VLM plan reading as in DOC-013; code-generation sandbox as in SCI-024; IFC tooling as in VIS-040.
- Upstream: VIS-040 (as-built models for retrofit), SCI-013 (requirements to design parameters), GEN-024 (building massing and layouts).
- Downstream: DOC-030 (quantity takeoff from the same model), PRED-048 (energy performance and renovation roadmap), OPT-007 (building operation).
- Reuse: KNOW-020 (find similar past projects), SCI-005 (simulation surrogates for building simulation), SCI-025 (checking of design risks).
Effort, cost & KPIs
- Path A: weeks to a few months; add-on licences or AI subscriptions, mostly absorbed in the existing suite.
- Path B: 4-9 months for interfaces, test projects and validators; ongoing maintenance when CAD/BIM versions change.
- Path C: similar plus GPU server and operations.
- Cost drivers: data quality at the architect interface, interface maintenance, engineer review time.
- KPIs: time per room for data entry, share of pre-filled fields accepted unchanged, errors found by the plausibility check, clashes found before coordination meetings, hours per architect change, rework after review.
People & change
Involve early: management (scope, liability), the head of engineering as process owner, and the planners who use the tools every day; add IT and IT security for project data in the cloud and the suppliers whose product data are used. Legal/compliance checks professional liability and client confidentiality (GUIDE-010, GUIDE-013).
Works council: depends. Most offices are small, but where a Betriebsrat exists, usage logs make the tools objectively suited to monitor behaviour under §87(1) no. 6 BetrVG, and §90(1) no. 3 requires information on AI-supported work processes. A short works agreement excluding individual performance evaluation usually suffices (GUIDE-004). Personal data are limited to user logs and author metadata (GUIDE-005).
AI literacy: Art. 4 AI Act: train planners on reading plausibility reports, typical misreads of plans and why every number in a proof comes from the calculation software (GUIDE-006).
Pilot and concerns: start with one project type (e.g. residential heat load) and compare against a manual run. Typical concern: loss of craft and liability for AI-made errors. Position the copilot as a checker and pre-filler, keep sign-off with the licensed engineer (GUIDE-008, GUIDE-009).
Risks, regulation & compliance
- EU AI Act: minimal risk for design tooling; no Annex III use. Product law applies to the building and its components, not to the design tool; the planner's verification duty is unchanged. AI literacy (Art. 4) applies; the 2026 AI omnibus reportedly softened the wording to taking measures to support literacy, so check the current text.
- GEG and energy proof: the Gebäudeenergiegesetz (GEG) governs energy requirements; DIN V 18599 is the calculation series for overall building energy performance, and certified energy-proof software must be used for official proofs. The AI may prepare inputs but not replace the approved tool. The official portal lists the law in 2026 under the title "Gesetz zur Einsparung von Energie und zur Modernisierung der Wärmeversorgung in Gebäuden"; check the current text (a Gebäudemodernisierungsgesetz replacing the GEG was reported for 2026, with conflicting sources on its entry into force) and funding rules (BEG/KfW/BAFA) before relying on a calculation route.
- Heat load: DIN EN 12831 is the basis of professional heat-load calculations; funders require standard-compliant calculation.
- Fees and liability: HOAI service phases, professional indemnity insurance and the engineer's signature remain; AI output is a draft.
- GDPR: normally no personal data; client and user data in project files and logs need an AVV (Art. 28 GDPR) with cloud vendors.
- IP and confidentiality: client building data and security-relevant layouts (e.g. critical infrastructure) need EU hosting and contract protection.
- Mitigations: deterministic calculation, per-room review, versioned inputs, audit log of AI-made changes.
Getting started
- Pick 3 finished projects with known correct results; measure manual time per step.
- Ask your architects for IFC with spaces and constructions; check what your TGA software can import.
- Test the AI and import features of your existing suite on those projects; log accepted vs. corrected fields.
- Prototype a plausibility checker (rules plus LLM) on past calculations before attempting routing.
- Treat routing/schematic automation as a separate pilot with strict human review.
Sources
- baunetzwissen — BIM in der TGA-Planung — https://www.baunetzwissen.de/drucken/bim-in-der-tga-planung-5380443
- baunetzwissen — DIN V 18599 Normenreihe zur Bewertung von Gebäuden — https://www.baunetzwissen.de/nachhaltig-bauen/fachwissen/regelwerke/din-v-18599-normenreihe-zur-bewertung-von-gebaeuden-677364
- streit-software — Heizlastberechnung Software (DIN EN 12831, vendor overview) — https://www.streit-software.de/wissen/heizlastberechnung-software
- engineering.com — How to Use Generative Design in BIM (GenMEP, beta status, article from 2016) — https://www.engineering.com/how-to-use-generative-design-in-bim/
- Gesetze im Internet — Gebäudeenergiegesetz portal — https://www.gesetze-im-internet.de/geg/