Technology · 10 July 2026
Why generic AI does not really understand building data
Construction has good standards for geometry and models. For day-to-day operations one is missing, and that is exactly where generic AI fails.
By Aidocr
Key takeaways
- The physical side of a building has standards (IFC, Brick, SAREF); the operational side does not.
- Without that structure, a language model does not know that “heating”, “HVAC” and “heating system” mean the same thing.
- Without modelled relationships, every AI answer remains a guess.
- A knowledge graph raises accuracy measurably: from 16.7 to 54.2 percent in an enterprise study.
BIM models, maintenance records, contracts, fault reports. Anyone working in construction and real estate knows the sheer number of data types that have to come together every day. Yet most AI systems only understand this data on the surface. The reason is not the language model, but a gap that hardly anyone names.
Standards exist, but not everywhere
For the physical side of a building, established standards are in place. Geometry and topology are described through BOT, sensors and equipment through Brick or SAREF, BIM models through IFC. That side of the industry is well modelled.
The operational layer is not. Maintenance histories, fault reports, leases, invoices: for this part of the building lifecycle there is no continuous standard. Every company builds those relationships by hand, usually over years.
Why that becomes a problem for AI
A language model can read text. What it does not automatically understand is that „heating“, „HVAC“ and „heating system“ refer to the same thing, or that a component belongs to a particular contract and a particular maintenance history. Without that link, every answer remains a guess.
A benchmark study by Sequeda, Allemang and Jacob shows the difference clearly: a language model answers questions over raw SQL databases correctly in only 16.7 percent of cases. With a properly modelled knowledge graph, accuracy rises to 54.2 percent, more than three times as high.
Source: Sequeda, Allemang, Jacob (2023), A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model's Accuracy for Question Answering on Enterprise SQL Databases, arXiv:2311.07509.
The answer is a structure of its own, not a better prompt
This is why a clever prompt is not enough. Aidocr builds its own knowledge structure for the operational side of buildings, mapping trade terms, synonyms and the relationships between component, contract and maintenance.
The result: questions are not just answered, they come with their source. Traceable for audits, for compliance and for every follow-up question that builds on the answer. What that structure looks like in detail is covered in the basics: What is a knowledge graph?
Frequently asked questions
- Why does generic AI only understand building data on the surface?
- Because there is no continuous standard for the operational side of buildings, with contracts, maintenance and fault reports. A language model can read text but does not know the relationships between component, contract and maintenance.
- Which standards exist for building data?
- For the physical side, established vocabularies such as IFC for building models, Brick and SAREF for equipment and sensors, and BOT for geometry and topology. For the operational side, no such standard exists.
- Does a better prompt help?
- No. The problem is structural, not linguistic. It needs a dedicated knowledge structure that maps trade terms, synonyms and relationships, not a cleverer wording.
- How much does a knowledge graph improve accuracy?
- In the enterprise benchmark by Sequeda, Allemang and Jacob (2023), answer accuracy rose from 16.7 percent on raw databases to 54.2 percent with a knowledge graph.