No two Building Information Modelling (BIM) models necessarily store information in the same way. This creates practical challenges for anyone trying to find what they need, particularly when they did not create the model themselves. My research looks at how AI can help address this problem, how we can assess the reliability of its answers, and how BIM and geospatial models can work together without losing the information each contains.
Why finding information in BIM models is difficult
The biggest challenge is what we call BIM heterogeneity: no two BIM models store the same information in the same way. The structure depends on the authoring software, the modelling conventions of each office, and even the language of the project.
A door width can be stored as “Width,” “NominalWidth,” or “Breite (B),” in a standard property set or a vendor-specific one. The gross floor area of a building might be an explicit attribute in one model and something that has to be computed from the geometry in another. In practice, the person looking for the information is often not the person who created the model. A facility manager or quantity surveyor may know exactly what they need but simply not know where to look.
Until a few years ago, the software available for information extraction was largely deterministic and had to assume a fixed data structure. This made it difficult to develop reliable extraction pipelines that could work across different models.
Letting AI explore the model
Large language models offer a different approach. Instead of translating a question into a fixed query, an AI agent can explore each model at runtime. It first discovers how that particular BIM model is structured, then searches for the answer, adapting each step to what the previous one returned.
This approach is known as agentic search. AI coding assistants already use similar techniques to find their way through large codebases.
In our recent paper in Automation in Construction, Agentic search for BIM information extraction: Systematic evaluation, we compared agentic search with a single-pass baseline, in which the model writes one query and cannot correct it after seeing the result. Agentic search won in every configuration we tested, by 29.5 to 48.5 percentage points.
To make these results reproducible, we published ifc-bench, a benchmark containing more than 1,000 questions across 37 IFC models from 21 projects, along with the code as open source.
Our latest prototype turns this approach into a tool. Users load an IFC model, ask a question in plain language, and receive an answer with the relevant elements highlighted in the 3D view. The aim is to make information in BIM models easier to access without requiring users to understand how every model is structured.
Can we trust the answers?
There will never be a guarantee that AI is always correct, but we can make its answers transparent and auditable.
AI-powered web search engines offer a useful example: they support their answers with sources, allowing users to check the information and judge for themselves whether it is correct. For building information, this means explaining how an answer was derived and linking it to the model elements on which it is based.
In our prototype, users can click on an element to inspect its properties or open the steps the agent took, including the code behind a computed value. They can then decide for themselves whether to rely on the answer.
The quality of the data matters, too. In our experiments, BIM model quality was the second strongest factor affecting accuracy, after the search approach itself. Models with more validation issues tended to produce fewer correct answers.
Trust also needs to be measured. We developed an evaluation framework for BIM question-answering systems and validated it using AI judges tested against human experts. One finding surprised us: on average, the AI judges agreed with each human expert more than the two experts agreed with each other.
This does not mean that AI is inherently more reliable than people. But without AI, the information we rely on comes from people, who are not fully reliable either. The reference answers in ifc-bench were written by people, and since its publication, researchers in the community have identified errors that we have corrected.
We should aim for highly reliable AI applications. At the same time, I think it is fairer to compare their performance with a realistic human baseline than to expect perfection from them.
Why converting BIM and geospatial models is not enough
BIM and geospatial models describe the built environment in fundamentally different ways, both geometrically and semantically. IFC describes a wall as a volumetric component, reflecting how it is built. A typical CityGML city model, by contrast, describes a building through the surfaces that bound it, often captured from the outside using aerial surveys or laser scanning.
A completely lossless conversion from IFC to CityGML is “practically impossible,” as the joint OGC and buildingSMART discussion paper on IFC, CityGML, and LandInfra puts it.
I find the alternative outlined in that paper more promising: linking instead of converting. Each model retains its own schema, while links connect objects that describe the same thing. A particular use case can then access whichever representation it needs.
Building permits provide a practical example. Placing a georeferenced BIM model in a 3D city model for viewing already works today. However, checking a proposed building’s height and its distances from neighboring buildings against the zoning plan is still largely done by hand. The EU project CHEK, in which OGC took part, developed tools to automate exactly these checks.
Connecting BIM and geospatial information in this way can help make data more useful in real workflows, without requiring different models to represent the world in exactly the same way.
Building interoperability through open standards
Open standards and continuous dialogue between the BIM and geospatial communities are key. The more standards we share and the more we talk to each other, the better interoperability gets.
The German Leitfaden Geodäsie und BIM (Guideline on Geodesy and BIM) is a good example of what this dialogue can produce. Published every year since 2017 by the German associations DVW and Runder Tisch GIS, it has grown into a reference of more than 300 pages on the practical use of BIM from a geodetic perspective.
Bringing these communities together is also one of the core missions of the Leonhard Obermeyer Center at the Technical University of Munich. The two fields are represented there by Professor André Borrmann in BIM and Professor Thomas H. Kolbe in geoinformatics, whose chair drives the development of CityGML.
Their communities have been exploring ways to connect their models for years. A 2017 research paper showed how IFC and CityGML models could be linked through linked data while both retained their original structures.
For me, this is an important principle: interoperability does not necessarily require everyone to adopt the same model. It requires open ways to connect information, preserve its meaning, and make it accessible across different applications and disciplines.
About the Author
Sylvain Hellin is an experienced real estate professional who holds a Master of Science in Civil Engineering and an Executive Master of Science in Real Estate. He has over 10 years’ experience in property development, most recently at Hines, where he managed multi-family and office projects of up to 100,000 m² in France and Germany. He currently serves as Managing Director of the Leonhard Obermeyer Centre, TUM’s Centre of Digital Methods for the Built Environment, while pursuing a PhD at the Chair of Computing in Civil and Building Engineering. His research focuses on the development of novel, AI-powered algorithms aimed at improving knowledge management within the AEC-O domain.
