From Chatbots to Project Intelligence: The New AEC Workflow
AI assistants in AEC software are specialised digital agents embedded in design, BIM and preconstruction platforms that use structured project data, engineering rules, and automated reasoning to reduce manual tasks, accelerate decisions, and maintain traceability across design automation CAD and construction workflows. This is not a side feature anymore; it is a deliberate attempt to rebuild the everyday workflow of architects, engineers and contractors around machine-readable knowledge instead of scattered files and spreadsheets. Vendors learned the hard way that generic chatbots hallucinate dimensions and misread regulations, so the serious focus is now on AI grounded in authoritative project data, connected BIM models and domain-specific agents that can be trusted in safety‑critical environments. The real shift is that AI assistants are starting to sit where work happens: inside CAD, inside BIM, and inside the messy preconstruction inbox.
Design Automation: Neural CAD and Smarter Geometry
In design, the biggest change is that AI assistants are being taught to understand geometry, not just text. Autodesk’s emerging strategy is to move beyond chat overlays and build what it calls “project intelligence”, where AI reasons over structured BIM data and relationships instead of isolated drawings. Its Neural CAD work trains a foundation model directly on professional CAD data, enabling AI to generate editable boundary representation geometry from prompts, sketches, images and even voice, rather than crude meshes. That matters because most physical products and many building components begin life as CAD, and current generic image‑to‑CAD tools often produce shapes “too simple for practical engineering work”. Researchers from MIT, Red Hat and IBM push in the same direction with GIFT, a framework that converts failed CAD program attempts into new training data and delivers more accurate CAD programs with about 20% of the usual computation. Users can tune GIFT’s compute budget to match time and cost constraints, which makes AI design automation viable in real projects rather than only in labs.
Virtual Companions: Dassault Systèmes Treats Construction as Manufacturing
On the construction‑as‑manufacturing side, Dassault Systèmes is betting that AI assistants should behave less like search boxes and more like expert teammates. At its AEC Summit, executives framed a convergence of virtual twins, industrialised construction and AI, arguing that the real value lies in intellectual property and know‑how encoded inside platforms such as Catia. Their answer is “virtual companions” — AI agents designed as world models, not large language models, built to work alongside engineers rather than replace them. Aura focuses on project management, logistics and planning; Leo supports structural and mechanical design; Marie concentrates on scientific research. Across these companions, the company describes 19 competencies spanning design, simulation, manufacturing and operations, all grounded in decades of physics‑based models and virtual twins. This approach is unapologetically opinionated: AI is framed as the new printing press for industrial knowledge, and firms are urged to treat their configuration logic, data lineage and optimisation recipes as assets they own, not features rented from whatever vendor ships the shiniest assistant.
Preconstruction AI Tools: Where Projects Are Won or Lost
The most quietly radical AI assistants may be landing not in design, but in preconstruction, where projects are won or lost before a single foundation is poured. For years, bid leveling has been a painful mix of Excel, PDFs and an estimator’s ability to spot scope buried in the fine print. MeltPlan is built entirely around that bottleneck. Its flagship preconstruction AI tool, Melt Bid, ingests subcontractor proposals in any format — handwritten, scanned, freeform or standard — and produces a leveled comparison automatically. According to MeltPlan, “pricing compared side by side, scope gaps and exclusions flagged, in about two minutes per package … down from roughly six hours a package”, compressing a 30‑package list from about 180 hours to under 30. Verification of a full package now takes about 30 minutes, replacing the old six‑hour grind rather than adding yet another draft that needs checking. An Ask AI feature lets teams query line‑item differences, lead times or low‑tier options and returns precise answers sourced from the underlying documents instead of vague summaries.
Owning the Knowledge Stack: BIM AI Integration Beyond Vendor Control
The common thread across these moves is a clear rejection of vendor‑controlled black boxes. As BIM AI integration matures, serious AEC practices want AI assistants that reason over their own data structures, not the public internet. Vendors are responding by building around knowledge graphs, ontologies, retrieval systems and digital twins so assistants can pull only what is relevant while preserving engineering context and traceability. This is the quiet revolution: firms that win will be those that treat their configuration logic, rules and optimisation methods as a proprietary knowledge stack inside these platforms, compounding value over time instead of renting generic intelligence from outside. Meanwhile, research such as GIFT is already planning its next step: extending the framework so models can generate CAD programs that improve performance and manufacturability of 3D designs, and testing it with larger models and broader CAD tasks. The direction is clear. AI assistants AEC software are becoming deeply embedded, domain‑specific collaborators. The open question is whether each practice will own enough of the stack to shape those assistants around its way of working, rather than adapting its workflow to whatever the assistant happens to offer.
