How to Make LLMs Work in Construction

During the Building 2030 Week, I hosted three panel discussions, one on AI in construction (pictured). Artificial Intelligence was one of the week’s key themes.

LLMs are already delivering value at Jatke

Much of the communication in construction is verbal or written and unstructured. Therefore, Large Language Models are perfect for analyzing, summarizing, and alerting about what’s happening during a project.

In his seminar presentation, Mikko Kuusakoski, the CIO at a Finnish general contractor Jatke, explained how the construction company uses AI. It has three scopes:

1. Everyone at Jatke has Microsoft Copilot Chat Enterprise at their disposal. Employees can use AI chat as they would any free version, but with a paid version, Jatke ensures the security of the submitted data.
2. For power and office users, MS Copilot Pro and Teams Premium are tools for summarizing, translating, and recording meeting minutes. A critical use case is the analysis and summarization of customer feedback.
3. For construction sites, Jatke has developed a virtual work desk. One of its apps is a jobsite safety reporting. Over the last 12 months, the company had 1,300 workplace safety observations. AI analyzes the reports and summarizes the most critical issues, thus helping a site fix them immediately. Other jobsite apps include Onboarding and Wendy, which uses documents and daily data to answer questions.

Mikko believes that, within a year, construction companies will start extending LLM’s capabilities with internal policy, compliance, and other documents. The latest RAG (Retrieval-Augmented Generation) solutions make this feasible and affordable.

Need for industry-specific models.

The seminar presentations and discussions indicated that more than general-purpose LLMs are needed for AEC companies. The terminology, business processes, regulations, and so on are local, industry-specific, and discipline-specific.

Instead of having each company build their apps to extend LLMs, the industry should have a shared “construction LLM.”

Furthermore, the presenters showed how LLMs create more trustworthy results if the data they read are “AI friendly.” In other words, tabular data (CSV) works better than PDF files, and unambiguous statements outperform expressive language.

Also, the terminology should be consistent. You should avoid using different words and expressions for the same thing.  Knowledge graphs and examples further improve AI’s accuracy.

The research continues

Building 2030 is a research and development consortium of 21 Finnish construction sector firms and Aalto University.

AI will remain a central research theme in 2024-25. One of the research topics is “The Use of AI in Construction Management: Benefits, Risks, and the Necessary Capabilities.”