Large Language Models (LLMs) are no longer a novelty in construction. Tools like ChatGPT, DeepSeek, and their successors are being tested in real workflows by researchers and practitioners alike. At Aalto University’s Building 2030 Summer Seminar, doctoral researchers Tuomas Valkonen and Roope Nyqvist demonstrated what happens when these models are asked to perform tasks directly tied to engineering and design.
Their experiments echo what we at Construction AI have observed across the sector and what we explored more deeply in our recent technical paper, From Chatbots to Project Insights: The Role of LLMs in Construction Data Processing - Link.
Why Out-of-the-Box LLMs Struggle with Engineering Data
When Valkonen asked ChatGPT and DeepSeek to generate an HVAC system description for a residential building, both produced serviceable but incomplete outputs. DeepSeek even omitted two essential systems. Later, when fed a hospital floor plan, the models attempted to create ventilation zoning tables, only to produce conflicting results.
These failures aren’t surprising. LLMs excel at textual reasoning but lack the structured awareness of BIM, IFC, or CAD-based environments. In our own Construction AI tests, similar gaps appear when trying to link model geometry with engineering logic. Quantity takeoffs, load calculations, or duct sizing often veer into nonsense unless paired with purpose-built engines like Kreo, Togal.AI, or Buildots.
Where LLMs Shine
Nyqvist’s research confirms the other side of the story: with the right framing, LLMs can radically improve the efficiency of construction knowledge work. His custom GPT for hospital design, built on thousands of expert files, was developed in just 100 hours and now answers specialized queries instantly.
This aligns with industry use cases we track at Construction AI:
- Drafting RFIs, change orders, and meeting minutes in minutes rather than hours.
- Translating technical content for multilingual project teams.
- Accelerating bids and proposals by generating narratives and case studies.
- Assisting with compliance documentation and client-facing summaries.
In these contexts, LLMs act less like engineers and more like tireless assistants—turning complex material into clear, actionable insights. To put it into perspective, we asked GPT 5 to summarize the reliable and unreliable use cases of LLMs in construction. It generated the following list, which we’ve visualized in the diagram below.
The Hybrid Future: LLMs + Domain-Specific Tools
The real value lies not in asking LLMs to replace established software, but in orchestrating them alongside specialized solutions. Imagine a workflow where ChatGPT drafts a client-friendly report while Kreo verifies quantities, or where a custom GPT interfaces with IFC validation engines to flag missing data.
This hybrid model—LLMs for reasoning and communication, domain-specific AI for precision will define digital construction in the coming decade. At Construction AI, we are already mapping this convergence: LLMs as the front-end intelligence layer that coordinates, explains, and augments the outputs of trusted specialist platforms.
A Practical Rule for 2025
For now, the rule of thumb is simple:
- Use LLMs for language-heavy, reasoning-oriented, and low-liability tasks.
- Rely on specialist tools for geometry, measurement, and compliance.
- And always ask the model: “Are you sure?”
The next leap will come from embedding LLMs into construction-specific ecosystems where their reasoning complements the precision of digital twins, BIM workflows, and AI-driven monitoring. That is where reliability will finally meet scalability.
