Early-stage building design decides most of a project's yield, energy use and carbon, yet teams usually test only two or three schemes by hand in CAD or spreadsheets. Generative tools take site data, zoning and programme, produce hundreds of code-aware layouts and masses, score them with fast surrogate models for daylight, wind, noise and area efficiency, and export the chosen option to Revit or IFC. Typical benefit: feasibility questions answered in hours rather than days, and far more options explored before design is locked in.
Problem & value
- Today: site feasibility is a manual loop. An architect or developer reads the zoning rules, sketches a massing, hand-calculates gross/net area and parking, orders daylight or wind studies from a consultant (days to weeks) and redoes it when the brief changes. Few options are compared.
- AI value: rule-aware generation plus instant feedback. Autodesk reports that Forma Site Design provides sun hours, daylight potential, wind, noise, solar energy, microclimate and carbon analyses, with "rapid AI" versions of wind and noise for interactive use [1][2]. Vendors such as TestFit position real-time unit/parking counts for feasibility [3]. Benefit figures differ by vendor and project; rely on a pilot with your own sites rather than marketing numbers.
- Who: developers (go/no-go on land), architects (competition and concept phase), municipalities (testing zoning or masterplan scenarios, e.g. density, shadow and wind impact).
Variants
- Site planning and massing: place buildings, roads and open space; optimise GFA, daylight, wind comfort, noise.
- Multifamily / unit-mix layout: parametric unit, core and parking placement against local code (e.g. TestFit-type tools).
- Floor-plan generation: room layouts inside a given footprint (research models, graph- and diffusion-based; Finch-style parametric tools for plans and facades).
- Feasibility-as-a-service for planners and lenders: automated property-level yield and rule checks at city scale (Archistar-type platforms) [4].
- Municipal scenario testing: compare zoning variants for shadow, wind, noise; links to PRED-036.
- Private use: homeowners testing extension or plot layouts with consumer floor-plan tools (low reliability).
How it works
- site data (cadastre, terrain, context buildings, climate) + zoning/code rules + programme + objectives
- rule engine (constraints as parametric code)
- generator (parametric/evolutionary search, optimiser, or learned generative model)
- surrogate analyses (daylight, sun, wind, noise, area/yield, carbon, cost)
- multi-objective ranking (Pareto front) and option comparison
- human selection and edit (architect/developer)
- export (IFC / Revit / Rhino / OBJ)
- detailed design, permit documents
- Generation: mostly constraint-based parametric engines, search/optimisation (genetic algorithms, MIP) and, increasingly, learned models (GANs, diffusion, graph networks for floor plans). An LLM turns a text brief into parameters and reads zoning text into rules; it does not draw the geometry.
- Surrogate models: neural networks trained on CFD, radiance and acoustic simulations return wind, noise or daylight maps in seconds instead of hours. They are approximations; Autodesk describes separate "rapid AI" and standard (simulation-based) wind and noise versions [1][2].
- Human-in-the-loop: professionals define the rules and objectives, choose and refine options, and must verify with certified simulation, structural, fire and code review before permit submission.
Data preparation
The data is spatial and rule-based. Site data (cadastre, terrain, neighbouring buildings, CityGML LoD2) comes from state survey offices, GIS or OpenStreetMap; zoning rules sit in Bebauungspläne, PDFs and building-code texts; programme and cost figures live in Excel files, the ERP or earlier project folders; completed plans are in the DMS, CAD or BIM archives. Typical gaps are scanned or non-georeferenced plans, inconsistent coordinate systems, zoning rules that exist only as prose, missing programme parameters, and project archives without metadata.
Preparation steps: - Convert zoning and code rules into a versioned, machine-readable rule sheet (plot ratio, height, setbacks, parking, fire and accessibility rules) with a date and source for each rule; an LLM can draft it, a planner must verify it. - Collect GIS layers into one coordinate system and a project template; check licence terms of survey data. - Tidy programme and benchmark spreadsheets (PREP-011) and define KPI definitions such as net area and yield (PREP-012). - Digitise and tag past plans if you plan Path D (PREP-006, PREP-010); define owners for the data (PREP-004).
Rough share of effort: 30-40 percent of the project, driven by rule encoding and GIS clean-up. Path A needs far less (a few days per pilot site).
Implementation paths
Which path for whom
- SMB: Path A, because Forma, TestFit or Finch cover standard housing and office typologies at seat prices without in-house developers.
- Enterprise: Path B, because developers and large practices need firm-specific rules, yield models and integration with internal BIM and cost tools that no product covers fully.
- Developer: Path C, because computational designers can build on Ladybug Tools, IfcOpenShell and open optimisers and keep rules and surrogate models as their own IP.
- Public: Path B, because municipalities need local rules (BauNVO, Bebauungsplan) coded into the workflow and EU-hosted or national cloud services; US-hosted SaaS is hard to procure.
Path A — Buy (recommended)
- Autodesk Forma (Site Design / Building Design), formerly Spacemaker (Norway-founded): site analysis, massing, IFC/OBJ and Revit/Rhino/Dynamo connections; Site Design is available to Revit subscribers [1][2].
- TestFit (US): real-time multifamily, parking and industrial feasibility with API/BIM export [3].
- Archistar (Australia): property feasibility and code-compliance at parcel scale [4].
- Finch (Sweden, EU): parametric floor-plan and facade optimisation for Rhino/Revit [5].
- Others: Digital Blueprint (Australia), Hypar (workflow functions to BIM), Giraffe, Snaptrude, ArchiLabs; Rhino with Grasshopper and Ladybug Tools for custom setups. Check: coverage of local EU/German rules (BauNVO, state building codes, DIN 5034 daylight), EU data hosting, IFC quality.
- Choose A when: fewer than ~50 projects/year, standard typologies (housing, offices), no in-house developers.
Path B — Build on cloud APIs
- Geometry and rules: Autodesk Platform Services (Forma APIs/extensions), Hypar, Speckle or IFC.js for BIM exchange; rules coded in Python/Grasshopper.
- LLM layer: a frontier model (Claude, GPT, Gemini, Mistral — EU provider) with structured outputs turns briefs and zoning ordinances into a rule/parameter schema; tool calls run the generator and simulators.
- Data: national cadastre/3D city models (e.g. CityGML LoD2 from German state surveys), OpenStreetMap, EU Copernicus data, weather files (EPW).
- Choose B when: you need firm-specific rules or typologies and integration with in-house tools.
Path C — Self-host (open source)
- Stack: Ladybug Tools (Honeybee/Radiance daylight, EnergyPlus), OpenFOAM or PALM-4U for wind, Speckle for BIM exchange, IfcOpenShell for IFC, Blender, Python optimisers (pymoo, OR-Tools, DEAP).
- Models: open diffusion and graph-based floor-plan research code; small open LLMs for brief parsing.
- Hardware: HW1 (CPU workstation) for parametric search; HW2 GPU to train/serve surrogates.
- Effort: needs computational-design staff; maintenance of rules per jurisdiction is the main cost.
Path D — Custom training
Train surrogate models on your own simulation runs (thousands of CFD/daylight runs on parametric massings) to get firm-specific fast feedback, or train floor-plan generators on a proprietary archive of plans. Needs clean, labelled plan data and licence rights to training drawings.
State of the art (2026)
- Interactive environmental feedback (surrogate wind, noise, daylight, sun) inside the massing tool is standard in leading products; vendor documentation and independent reviews agree on this, but reviewers disagree on how "generative" the form-making is: much is rule-based option generation and fast analysis rather than free creative synthesis [1][2].
- Feasibility automation (unit counts, parking, yield) for multifamily is mature; parcel-scale automated code checking is expanding by jurisdiction [3][4].
- Floor-plan generation: learned models (graph and diffusion) produce plausible layouts, but fidelity to fire escape, accessibility, structure and building services is weak; parametric rule-based generators remain more dependable in production.
- BIM integration: IFC (buildingSMART) and Revit round trips work for geometry and areas; semantic richness and edits returned from Revit are still limited [2][6].
- Limits and failure modes: surrogate models can be wrong outside their training range (unusual terrain, tall buildings, dense contexts); rules encoded incorrectly give illegal schemes; optimisers maximise yield and ignore quality of life, social mix or heritage context; local rules change.
Shared resources & combinations
- Same stack: simulation, optimization-solver and 3d-vision building blocks are shared with SCI-006 (generative engineering design) and PRED-048 (building energy performance).
- Upstream: GEN-025 (photo/video-to-3D) and drone/GIS data for existing context models.
- Downstream: GEN-009 (rendering and visualisation of the chosen scheme), GEN-018 (staging), DOC-016 (permit document checks against rules), PRED-036 (physical climate risk of the site).
- Data pipeline: one BIM/IFC and GIS pipeline (Speckle/IfcOpenShell) serves all these cases.
Effort, cost & KPIs
- Path A: days to weeks to onboard; per-seat subscriptions (Forma comes with Revit access for some customers [2]); pilot on 3-5 past projects.
- Path B: 3-6 months for rules, connectors and UI; ongoing rule maintenance.
- Path C/D: 6-12 months with a computational-design team; simulation licences are low, staff cost dominates.
- KPIs: options evaluated per project, time from site to feasibility report, deviation of surrogate results from certified simulation, yield (GFA/net area) uplift versus baseline scheme, code-compliance rejections at permit stage, embodied carbon per m2.
People & change
Who and when. Start with management (budget, which typologies) and the department lead for design or project development, who owns the rules and signs off the options. Involve end-users (architects, planners, project developers) in the pilot from week one, because they define objectives and judge the layouts. IT handles BIM and GIS integration and accounts; data owners of cadastre, GIS and plan archives release data; legal-compliance clarifies liability and licence terms for site data and generated layouts. See GUIDE-001 for choosing pilot projects.
Works council. Usually no. The tool processes geometry and zoning, not employee behaviour, so §87(1) no. 6 BetrVG is typically not triggered unless the platform logs individual usage in a way suited to monitor staff (BAG: objective suitability suffices); check vendor telemetry. Inform the works council anyway under §90(1) no. 3 if work processes change noticeably. Public sector: involve the staff council.
DPO and DPIA. Mostly no personal data; involve the DPO only if cadastre owner or resident data enters the workflow, and keep it out of cloud prompts (GUIDE-005).
Literacy, communication, pilot. Support AI literacy (Art. 4 AI Act) with a workshop on what surrogate models can and cannot prove (GUIDE-006). Pilot on 3-5 past projects and compare against the manual scheme (GUIDE-009). Typical concern: "the tool makes design generic." Address it by showing it as an option generator; the architect chooses, refines and remains professionally liable, and certified simulation precedes any permit.
Risks, regulation & compliance
- EU AI Act: minimal risk for design tools; AI literacy (Art. 4) applies. No Art. 5 practice is involved. A system used as a safety component of critical infrastructure would differ, but building layout generation normally is not. The AI Act text: [7].
- GDPR: mostly no personal data (geometry, zoning). Cadastre owner data, neighbour information and resident data used for municipal scenarios may be personal; keep them out of prompts to cloud models.
- Professional liability and codes: architects remain responsible under national chamber law and HOAI-style fee and liability rules; generated layouts must be verified for building code, fire protection, structure, accessibility (e.g. DIN 18040) and energy law (GEG). Surrogate results are not certified proof.
- IP: licence rights for training drawings and for generated output; check vendor terms.
- Mitigations: documented rule sets with version dates, certified simulation of the final scheme, human sign-off, keep a log of options and parameters, avoid optimising for a single metric.
Getting started
- Pick 3-5 completed sites and record the real brief, zoning constraints and the final yield.
- Trial Forma Site Design and one feasibility tool (TestFit or Finch) on those sites; compare options and yield.
- Test surrogate wind/daylight results against one certified simulation to learn the error range.
- Encode your local zoning rules (BauNVO, state building code) as a checked rule set; start with one typology.
- Define the Revit/IFC export and the sign-off steps before using results with clients or authorities.
Sources
- Autodesk — What is Autodesk Forma Site Design — https://blogs.autodesk.com/forma/2026/08/20/what-is-forma-site-design/
- AEC Magazine — Autodesk Forma Building Design and Revit Connected client — https://aecmag.com/bim/forma-building-design/
- TestFit — product site (feasibility, unit and parking layout) — https://testfit.io
- Archistar — product site (property feasibility and code compliance) — https://www.archistar.ai
- Finch — product site (parametric floor-plan and facade optimisation) — https://www.finch3d.com
- buildingSMART — Industry Foundation Classes (IFC) — https://technical.buildingsmart.org/standards/ifc/
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Ladybug Tools — open-source environmental analysis — https://www.ladybug.tools