AI for architects and planners: BIM, building applications, tenders – where language models actually save time

What the industry does – and where AI fits

The Swiss Society of Engineers and Architects SIA has roughly 11,000 members, of which about 4,500 architecture and 6,500 engineering/planning offices (SIA annual report 2024). The industry is fragmented: solo practitioners, SME offices with 5 to 30 staff, general planners and large networks such as Burckhardt, Itten Brechbühl, BIG Switzerland. Daily work mixes three domains: design and visualisation (architecture in the narrow sense), technical planning (BIM, structure, HVAC, electrical) and administration (building applications, tenders, contracts, correspondence).

In the design and visualisation layer, AI is already shipped by the software vendors: Adobe Firefly for moodboards, Vectorworks Cloud-AI, Autodesk Forma for early BIM phases, Enscape and Twinmotion with AI textures. These tools are part of their platform licence.

The administrative layer, on the other hand, is Word-Excel-PDF every day – and exactly the area where a language model produces practical results today. Building-application texts (descriptions, justification of deviations), tender analyses (reading specifications, writing your own suitability text), contracts (SIA-based works contracts, fee agreements), client correspondence (updating the building owner, dealing with authorities) – all text work that a well-tuned German-language model competently prepares.

Important: structural analysis, seismic verification, fire safety concept are not the field of AI language models. Structure is engineer territory, signed and countersigned, with liability. A language model must neither compute nor review a structural report – attempting that risks liability damage and SIA-code issues.

Why it matters now

Three market trends are forcing Swiss architecture offices into efficiency investment in 2025/2026.

First: Swiss construction grew only 1.8 percent nominal in 2024 (BFS construction statistics 2024), residential reservations are dropping, public tenders are getting harder. Offices that answer tenders faster, more completely and with more variants win. Tender analysis today is a four to eight hour job per call for a 5-person office; AI preparation cuts it in half measurably.

Second: fees under pressure. SIA regulation 102 (architectural services) provides reference rates, but building owners negotiate aggressively. The mix of flat fees and rising labour cost forces offices to remove administrative work from the creative hourly rate.

Third: BIM mandate in public construction. Federal, cantonal and large city authorities require BIM models at LOD 300 to 400 (federal BIM strategy 2025-2030). BIM models produce thousands of data records per project – quantities, component attributes, material documentation. AI-assisted quantity sanity checks and automatic attribute validation against norms (SIA 416, eBKP-H) are production-ready in 2025 and relieve the BIM manager.

Legally the situation is more comfortable than in medicine or fiduciary: personal data is rare; the sensitivity sits with business secrets and competition protection. Dumping a competitor entry from a cancelled tender into a US cloud tool is a problem of a different kind – the awarding authority will not appreciate it. EU/CH hosting therefore remains the default.

Where AI plugs into the office

Four use-cases cover most of the leverage. All four are administrative and text-based – not structural engineering.

Tender analysis. A 80-page public tender is loaded via a RAG pipeline (rag-eigenes-wissen). The model extracts specification requirements, evaluation criteria, suitability evidence, deadlines and warranty clauses into a structured list. The office decides in 30 minutes whether to bid – instead of after 4 hours of reading. If yes, the model drafts the suitability text from earlier successful submissions.

Building application texts. The descriptive document, the justification of deviations from regulation, the statement on a third-party objection – formulaic texts with a clear structure. A language model with access to the office archive (past applications, regulation excerpts, authority comments) delivers a draft that needs only proofreading in 80 percent of cases. See prompt-engineering-grundlagen for template structure.

BIM attribute checks. From an Allplan, Archicad or Revit model, IFC export produces component data records. A pipeline (n8n-workflow-automation) checks per component class whether mandatory attributes (eBKP-H code, U-value, fire class, manufacturer) are present and consistent. Without AI: every record is checked manually. With AI: the model flags only the 3 to 8 percent of records that are unclear or contradictory.

Client communication. Weekly updates to the building owner, meeting minutes, replies to authorities. Here a light setup is enough: a mail plugin that proposes the draft, plus the office template. Daily total: 30 to 60 minutes less typing.

Further candidates worth piloting in 2026: AI-assisted search in the Swiss building product catalogue, automated norm lookups in SIA collections, clause-by-clause contract comparison against SIA 118. All text-driven, all traceable.

6 steps to AI-assisted tender analysis

1. 01Prepare the tender archive: collect past successful and unsuccessful submissions digitally, with metadata (client, volume, won yes/no, fee range).
2. 02RAG setup: office vector database local (Qdrant) or EU-hosted, embeddings via a multilingual model (Mistral Embed, Cohere multilingual). Separate client and competitor data by collection.
3. 03Build the spec extractor: prompt template that turns a tender into a structured list (requirements, evaluation criteria, suitability evidence, deadlines) – table output, not prose.
4. 04Go/no-go workflow: office lead receives the list plus a first suitability check after 15 minutes – decides whether to participate.
5. 05Suitability-text generator: model pulls similar past entries from the archive, drafts the text, project lead edits. Clearly mark "AI draft" until release.
6. 06Submission tracker: every entry and its outcome is stored in the same database. After 12 entries the model has enough material to estimate win probability from phase 1.

When an office should start

Entry runs in three stages, like the other industries: audit, pilot, managed.

The audit (ai-readiness-audit) captures the platforms in use (Allplan, Archicad, Revit, ArchiCloud, Vectorworks, Office 365, SharePoint), the office archive (past tenders, applications, contracts) and the data flows. It takes 2 to 3 days and outputs a prioritised pilot list. Offices with well-digitised archives can start RAG immediately – offices with paper archives need OCR and indexing first (see ai-belegerkennung-ocr for the principle).

The pilot focuses on one of the four use-cases, usually tender analysis, where time saving is most directly measurable. Pilot duration: 6 to 10 weeks, one project lead plus one student assistant. Success metric: hours per tender before vs. after, plus hit rate (won/submitted).

AI pays back when the office submits at least 6 to 8 tenders per year or handles 15 to 30 building applications per year. Below that, setup cost is hard to recover – an audit-only engagement with a 12-month follow-up makes more sense.

Where architect AI has limits

Three areas are not AI tasks, even though the temptation grows.

First: structural analysis, seismic, fire safety. These belong to a qualified engineer (SIA 260-269, BWA fire safety norm 2025). A language model produces no verifiable proof. It may ask a first sanity question ("is this support load realistic for a 6m timber beam ceiling?"), but that is not an engineering result. Submitting AI-written structural text for a building application risks permit loss and liability claims.

Second: legally binding contract text. SIA works contracts (SIA 118) and fee agreements remain a lawyer matter. AI may prepare a draft – the review against law, liability exclusion and insurance wording is done by the house lawyer. This applies especially to total- and general-planner contracts with Bauherrenschaft Schweiz or other institutional clients.

Third: competition entries in decision-critical form. A competition is judged on a creative contribution – AI must not produce it. What it may do: research the competition framework, prepare the site and statistics, suggest poster layouts. The actual design and explanatory text must come from the office, not the model.

There is a fourth practical limit: for very small offices (solo, 2 people) with fewer than 5 tenders/year ROI is tight. Then only the one-off audit recommendation is worth it – no setup.

Trade-offs

FAQ

Related topics

Sources

1. SIA Geschäftsbericht 2024 – Mitgliederzahlen und Branchenstruktur · 2024-12
2. BFS Bauwirtschaft Schweiz 2024 – Statistik des Bauwesens · 2025-03
3. Bundesamt für Bauten und Logistik – BIM-Strategie Bund 2025-2030 · 2025-06
4. Autodesk Forma & Revit AI – Produktdokumentation 2026 · 2026-02
5. SIA Norm 118 – Allgemeine Bedingungen für Bauarbeiten (Ausgabe 2013, in Revision 2026) · 2026-01