Can I use ChatGPT or Claude to write Dynamo scripts for Revit today?

Yes, with caveats. GPT-4o and Claude 3.5 Sonnet generate functional Dynamo Python scripts for well-defined, common tasks — filtering elements by type, creating views, setting parameter values, generating schedules. Success rate is 70–85% for common operations. Complex tasks involving custom families, linked models, or multi-phase operations often require debugging. The recommended workflow: generate the script, paste it into Dynamo's Python node, run it on a test model, and validate the output before using it on a live project file.

Will AI replace CAD drafters and BIM modellers?

AI will automate significant portions of repetitive modelling and documentation tasks — estimated 30–50% of current drafter hours over the next 5 years according to McKinsey Global Institute's 2024 AEC productivity analysis. However, roles are evolving rather than disappearing: BIM professionals who can direct AI tools, validate outputs, manage model quality, and coordinate across disciplines will be in high demand. The skill shift is from manual geometry creation toward model governance, AI tool management, and higher-level design coordination.

What is the best way to use AI for Revit family creation?

Family creation remains one of the harder tasks for AI because it requires understanding parametric constraints, geometry types (solid extrusion, sweep, blend), and hosting behaviour. The most effective current approach: use an LLM to generate the Revit API code (C# or Python via pyRevit) that programmatically creates the family, rather than trying to drive the GUI. For standard family shapes (rectangular column, simple structural connection), this works well. For complex adaptive components or irregular geometry, sketch-to-family tools like Finch3D or Hypar are more reliable than pure LLM generation.

How can I protect client BIM data when using AI tools?

Key steps: (1) Review the AI tool's data processing agreement — most enterprise tools (Autodesk AI, Azure OpenAI, AWS Bedrock) offer DPAs that prohibit training on your data. (2) Check your BIM Execution Plan (BEP) and client NDA for data residency and third-party sharing restrictions. (3) Use on-premise or private cloud LLM deployments for projects with strict confidentiality requirements. (4) Anonymize or reduce model data before sending to external APIs — send element types and parameters rather than complete project geometry when the full model is not necessary. (5) Document AI tool usage in your project quality management plan.

Natural Language Interfaces for CAD and BIM: What's Possible in 2026

A practical guide to conversational AI in CAD and BIM — which workflows are production-ready, which are emerging, and how to evaluate tools for your engineering practice in 2026.

The Shift Toward Conversational Design Tools

For decades, CAD and BIM tools required engineers to master complex command hierarchies, parametric modelling logic, and proprietary scripting languages (Dynamo, Grasshopper, Revit API, AutoLISP). The emergence of large language models capable of generating code and understanding design intent is beginning to break that barrier. In 2026, natural language interfaces exist at three maturity levels: production-ready features inside mainstream tools, emerging research prototypes nearing commercial release, and speculative capabilities that remain 3–5 years away.

Production-Ready: AI Features in Major Platforms (2026)

Several major AEC software vendors have shipped AI-assisted features that are in active production use:

Autodesk AI (Project Bernini / Forma): Autodesk Forma includes AI-generated massing studies from site constraints and performance targets. Natural language prompt input ("maximize south-facing floor plate area while maintaining 35% open space") generates multiple massing options evaluated against wind, daylight, and energy metrics.
AutoCAD AI Commands: Autodesk has integrated natural language command search and AI-assisted drafting suggestions. Users can type "add a 2-meter offset from this wall" and the system interprets and executes — reducing reliance on memorizing exact command names.
Revit Copilot (Autodesk AI / third-party integrations): Plugins using the Revit API + LLM backends (GPT-4o, Claude) can generate Dynamo scripts from natural language descriptions, query model elements ("list all rooms below 15 m² on Level 3"), and create basic parametric families from text descriptions.
Trimble / SketchUp AI: SketchUp's AI sidebar generates 3D geometry from text prompts for concept massing, and suggests material assignments based on element context.
Speckle + LLM integrations: Speckle's open-source BIM data platform enables custom LLM chatbots to query model geometry, element properties, and version diffs through its API — useful for project status queries ("how many structural columns changed between revision 4 and revision 5?").

What Natural Language Can Do Today in BIM

In practical terms, 2026 natural language BIM interfaces excel at:

Model querying and reporting: "Generate a door schedule for Level 2 showing fire rating and hardware group" — reliable and widely deployed.
Script generation: "Write a Dynamo script that copies all Type A light fixtures from Level 1 to Level 2 at the same XY coordinates" — GPT-4o and Claude 3.5 Sonnet generate functional Dynamo Python nodes for well-defined tasks.
Code generation for Revit API: generating C# or Python snippets for common Revit API tasks (element filtering, parameter setting, view creation) with high reliability.
Clash detection summaries: interpreting Navisworks or Revit clash reports and generating natural language summaries prioritized by severity.
Specification linking: matching model element types to relevant specification sections from a project spec database.

Emerging Capabilities: The 2025–2027 Frontier

Research prototypes and early commercial products are demonstrating capabilities not yet in mainstream production:

Sketch-to-BIM: hand-drawn floor plan sketches (or photos of white-board diagrams) converted to parametric Revit models. Startups like ARX and research groups at ETH Zurich have demonstrated this at room layout granularity. Accuracy for complex multi-room floor plans with structural elements remains insufficient for production use.
Multi-modal design dialogue: systems that accept a combination of image, text, and voice input to iteratively refine a design ("make this wall load-bearing, shift the staircase 500 mm east, and show me the impact on the schedule").
Automated code compliance checking: LLMs parsing model geometry against building code requirements (IBC, Eurocode, NBC Canada) — research-stage, with multiple commercial products (Archilyse, AI Clearing) targeting this in 2025–2027.
Structural pre-sizing from architectural intent: from a natural language description of span, occupancy, and seismic zone, AI proposes preliminary member sizes for engineering review.

Limitations and Realistic Expectations

Engineers evaluating AI-driven CAD/BIM tools should understand current constraints:

Geometry accuracy: LLMs do not inherently understand 3D coordinate geometry. Generated scripts must be validated; generated geometry often requires manual correction.
Design intent vs. design correctness: AI tools generate options that look plausible but may violate structural logic, egress requirements, or coordination constraints. Engineer review is mandatory.
Proprietary model complexity: the Revit API exposes a fraction of model data; many complex families and linked models remain outside LLM query reach without custom tooling.
Hallucinated parameter values: AI tools querying model properties can misreport values if the underlying API call fails silently — implement validation checks for any AI-generated schedule or report used in production.

How to Evaluate a Natural Language BIM/CAD Tool

A practical evaluation framework for your engineering team:

Scope test: submit 20 representative tasks from your actual workflow; measure success rate and time savings versus manual approach.
Error rate audit: for each generated output, quantify errors requiring manual correction. An error rate above 15% usually negates time savings.
Integration depth: does it operate inside your existing tools (plugin) or require data export/import? Round-trips add friction and version control risk.
Data privacy: does the tool send model geometry and project data to external LLM APIs? Verify with your client NDA and BIM Execution Plan (BEP) data security requirements.
Training requirement: how long does it take a typical team member to become proficient? Tools requiring more than 2 hours of onboarding have low adoption rates in busy engineering offices.

The Future: Integrated Design AI Agents

The trajectory points toward persistent AI agents that maintain context across an entire project — understanding the design intent, structural system, code constraints, and coordination history. Rather than a single query-response interaction, these agents will monitor design changes, proactively flag issues ("this beam span increase exceeds the pre-sized section capacity — do you want me to re-run sizing?"), and maintain a living record of design decisions. Autodesk's Forma platform and Bjarke Ingels Group's custom internal tools are early examples of this direction.

🤖 Natural Language Interfaces for CAD and BIM: What's Possible in 2026