A building already contains much of the information needed to understand how it was designed, how its systems are connected, and how it should operate. The problem is that this knowledge is rarely available in a form that a machine, or sometimes even an engineer, can readily use.
HVAC diagrams describe system relationships. Equipment schedules contain capacities and design values. Floor plans locate spaces and assets. Energy performance certificates describe building characteristics. BIM models encode geometry and system information. Commissioning reports, maintenance records and technical specifications add another layer of operational knowledge.
The problem
Individually, these documents are useful. Together, they could provide a detailed account of the building. In practice, they rarely do.
Research consistently describes building information as fragmented across sources, represented using incompatible structures and terminology, and difficult to transfer from design and construction into operation. Studies of BIM-to-facility-management workflows report incomplete, inaccurate, inconsistent and sometimes unintelligible information, while reviews continue to identify interoperability and information handover as persistent barriers.1234
AI changes what can be extracted from these documents. But extraction alone does not solve the problem. The harder question is what happens after the information has been extracted.
Buildings accumulate information rather than knowledge.
A single property may have hundreds or thousands of files created by architects, engineers, contractors, equipment manufacturers, commissioning teams and facility managers over decades.
The information may exist as text, tables, diagrams, scanned drawings, images, BIM objects or handwritten annotations. Naming conventions change between documents. The same piece of equipment may have several identifiers. Units differ. Some documents describe the original design while others describe later modifications. Some information is missing. Some is simply wrong.
Eight source types, eight different interpretation problems. Each is individually tractable. The difficulty is that they describe the same building and do not agree on how to name it.
This fragmentation has direct operational consequences.
An ethnographic study examining more than 2,000 BIM-for-facility-management compliance issues identified information that was incomplete, inaccurate, inconsistent or unintelligible, limiting its usefulness for facility-management decisions.1 Research reviews similarly identify fragmented workflows, incompatible schemas and weak information exchange during commissioning and handover as persistent barriers to effective operation and maintenance.234
The problem is therefore not simply that building data are unavailable. They are difficult to connect. That distinction matters.
Extracted data
"AHU-03"
A drawing contains this text. The system can find it, index it and return the page it appeared on. Nothing about the building follows from it.
Building knowledge
AHU-03 → serves Zone B
→ 2.2 m³/s design airflow
→ feeds HC-03
→ live BMS point AHU03_SF
The same string, placed. It has a zone, an airflow, downstream equipment, consistent appearances across several documents and a corresponding live point.
The first is a search result. The second is the beginning of an account of the building that a machine can reason over.
Why current approaches fall short
The obvious response to document fragmentation is digitisation. Scan the drawings. Apply OCR. Search the PDFs. Put the documents into a database. More recently, ask a large language model to read them.
Each step improves accessibility. None, by itself, establishes that the resulting information is correct.
Research on AI-assisted building information processing is increasingly promising. Domain-adapted multimodal retrieval systems can process combinations of text, images and tables while preserving more document structure than generic approaches. One recent building-focused pipeline reported approximately 30% improvement in document coverage and structural preservation and reduced semantic-model preparation from days to minutes.5
LLMs have also shown strong performance in narrower building tasks. Recent BIM compliance research reported approximately 97% classification F1 and similarly high performance for dependency identification, tool selection and rule execution.6 Other studies have demonstrated AI-assisted semantic enrichment of BIM models and extraction of building characteristics from imagery.78
These results establish capability. They do not establish universal reliability.
Reviews of multimodal AI in architecture, engineering and construction continue to identify limited external validation, weak robustness to missing modalities, inconsistent data quality and a lack of standardised benchmarks.9 Closely related engineering benchmarks have found that multimodal models can perform well at equipment identification and standardised extraction while still struggling with spatial reasoning, cross-modal synthesis and multistep inference.10
This creates an important distinction for building applications. Reading a document is not the same as understanding a building. And understanding a building is not the same as proving that what was extracted is true.
The core concept: document extraction is not the goal
The useful output of document AI is not another database of extracted text. It is a validated representation of the building.
Consider a simple statement extracted from an HVAC document. A conventional extraction pipeline may stop once those fields have been identified. A building-knowledge system should ask additional questions.
Extracted
AHU-03
design airflow
2.2 m³/s
What a building-knowledge system asks next
- Does AHU-03 appear elsewhere?
- Is 2.2 m³/s dimensionally and physically plausible?
- Does another schedule give the same value?
- Which spaces does the unit serve?
- Is the document describing the current installation or an obsolete design?
- Can the extracted value be traced back to its source document and location?
- Does operational data from the BMS correspond to the equipment described in the documents?
The value becomes useful only when its context travels with it.
That requires three layers. First comes extraction: what does the source appear to say? Second comes validation: how much evidence supports that interpretation? Third comes connection: how does the extracted information relate to everything else known about the building?
01 · EXTRACTION
What does the source appear to say?
Interpretation of text, tables, symbols and geometry into candidate entities, attributes and relationships.
02 · VALIDATION
How much evidence supports it?
Schema, dimensional and engineering checks, plus agreement or disagreement between independent sources.
03 · CONNECTION
How does it relate to the rest?
Placement in a semantic structure where equipment, systems, spaces, sensors and documents reference each other.
Three layers, and the second is the one most pipelines skip.
This is where AI-assisted document processing starts becoming building intelligence.
Two structural points carry most of the weight.
A claim is not a value, and it should not be stored as one
An extracted number carries no evidence about itself. A claim carries its source document, its location in that document, the method that produced it, a confidence derived from evidence rather than requested from a model, and a date.
The difference is not bookkeeping. It is what makes the information inspectable rather than merely searchable, and it is the only thing that lets a later system, or a later engineer, decide whether the value should still be believed.
Sometimes the correct output is not a value
Buildings contain contradictions. A BIM model may describe the original design, a later drawing may contain a modification, a maintenance record may identify replaced equipment, and the BMS may use a completely different identifier. A system that silently picks one has manufactured certainty it does not have.
The honest output is "these sources disagree", with all of them shown and the reason the conflict could not be settled. That is substantially more useful than false certainty.
How it works
The process can be understood as a chain. Each stage produces information that the next stage depends on.
Ingest heterogeneous building information
The starting point is not a clean database. It is the building as it actually exists: HVAC diagrams, architectural drawings, EPCs, BIM files, equipment schedules, technical specifications, commissioning reports, maintenance records and other documents accumulated throughout its lifecycle.
Research consistently identifies this heterogeneity as one of the central barriers to building information integration.34
Understand document structure, not only text
Different documents require different interpretation strategies. A specification is primarily linguistic. An equipment schedule depends on table structure. An HVAC diagram depends on symbols, labels and spatial relationships. A floor plan combines geometry, text and graphical conventions.
Multimodal AI is increasingly capable of processing these information types together, and building-focused research has demonstrated systems combining images, tables and text while preserving hierarchical and spatial relationships.5 But current evidence does not support treating multimodal models as dependable engineering reasoners: spatial reasoning and cross-modal inference remain documented weaknesses.910
The system should therefore use AI to interpret documents, not assume that interpretation is ground truth.
Extract claims, not isolated values
A number without context is difficult to validate. Instead, extracted information can be represented as claims: AHU-03 has a design airflow of 2.2 m³/s. AHU-03 serves Zone B. Pump P-04 belongs to the heating circuit. Drawing X identifies AHU-03 as serving rooms 201-218.
Each claim can carry its source, document location, extraction method, confidence and date. This makes the information inspectable rather than merely searchable.
Normalize terminology and units
Building information rarely uses one vocabulary. The same asset may be named differently across drawings, BIM models, BMS exports and maintenance records. Semantic inconsistencies are repeatedly identified as a barrier to interoperability and automated building-data integration.311
Normalisation maps these representations into a consistent vocabulary while preserving the original source. Units require the same treatment. A value should not merely be extracted as "2200". Its quantity, unit and engineering meaning must be established before another system is allowed to use it.
Validate before accepting
This is the step that separates document extraction from trustworthy building knowledge.
Research outside and within engineering shows that AI extraction can perform strongly on simple, discrete fields while deteriorating on more complex information. Reported failure modes include partial extraction, hallucination, incorrect interpretation and confident errors.121314
The appropriate response is not to abandon generative AI. It is to constrain its role. Schema validation can check whether required information exists. Dimensional rules can detect incompatible units. Engineering constraints can identify implausible values. Multiple sources can be compared. Conflicting claims can be surfaced instead of silently resolved.
Recent research also supports provenance-grounded extraction, parallel extraction and reconciliation, uncertainty estimation and human review as safeguards against unreliable outputs.14151617
Confidence should therefore come from evidence around the claim, not simply from the model that generated it.
Reconcile sources
Buildings contain contradictions. A BIM model may describe the original design. A later drawing may contain a modification. A maintenance record may identify replaced equipment. The BMS may use a completely different identifier.
A useful system cannot simply choose one source and discard the others. It must preserve provenance, compare claims, establish source authority where possible and identify unresolved conflicts.
Build a machine-readable knowledge layer
Once validated, building information can be connected into a semantic representation. Research on building ontologies and knowledge graphs has demonstrated integration of BIM, building automation systems, IoT devices, sensors, documents and operational data. Such representations have supported automated HVAC topology extraction, building-energy-model generation, control applications and digital twins.1819202122
Instead of storing isolated records, the system can represent relationships.
Now a machine can traverse the building rather than merely search its files.
Connect documents to operational data
Documents describe what the building is supposed to be. Sensors describe what it is doing. Connecting the two creates a much more useful analytical layer.
A temperature anomaly becomes more informative when the system knows which AHU serves the room, which heating circuit feeds that AHU, what design conditions were specified, what equipment is installed and whether maintenance records indicate a recent intervention.
This connection between relatively static building knowledge and dynamic operational data is still an open technical challenge. Reviews identify weak representation of time-series information in existing BIM and ontology workflows and continuing integration difficulties with legacy building-management systems.232425 It is also where much of the practical value lies.
Move from evidence to decisions
Once building information is structured, validated and connected, analytics no longer operate on anonymous columns. They operate in the context of the building.
A system detecting unusual nighttime heating can ask not only whether consumption is high, but which systems could physically produce that load, which zones they serve, what their intended schedules are, whether the relevant equipment has changed, and which documents support that conclusion.
That difference is the distance between an anomaly and a decision.
Why it matters
Facility managers do not need another repository containing more building data. They need less time spent finding, checking and reconciling information before they can act.
The research evidence for the operational impact of semantic building information is promising but still limited. Studies report applications in automated model generation, model-predictive control, facility management, fault detection and digital-twin workflows.1819202122 Other research reports substantial reductions in model-development effort when BIM metadata and semantic standards are used.22
But the evidence does not yet justify a broad claim that converting documents into semantic knowledge automatically reduces energy consumption or maintenance costs. Those outcomes depend on what is subsequently done with the information.
The defensible value proposition is narrower and more important: structured building knowledge makes information easier for machines and people to retrieve, connect, check and use.
improvement in document coverage and structural preservation, with semantic-model preparation cut from days to minutes, in one building-focused multimodal pipeline.5
classification F1 for LLM-based BIM compliance checking, with similarly high performance for dependency identification, tool selection and rule execution.6
Industry perspective
Three developments are converging. AI can interpret increasingly heterogeneous technical information. Semantic technologies can represent the relationships that conventional databases often lose. And operational analytics can increasingly connect those relationships to live building data.
The missing layer is reliability.
The literature repeatedly identifies interoperability, semantic inconsistency, missing metadata, legacy-system integration, data quality and scalability as unresolved barriers.39232425 Current research therefore supports promising prototypes, not dependable automated understanding of arbitrary buildings at portfolio scale.
That distinction matters commercially as much as scientifically. The winning system is unlikely to be the one that can extract the most information from a PDF. It will be the one that can show which extracted information should be trusted, where it came from, what contradicts it, how it relates to the physical building, and whether operational evidence supports it.
What the evidence does not yet support
There is not yet strong evidence that AI can autonomously convert arbitrary collections of building documents into complete and reliable digital representations of buildings.
The literature is fragmented by application. Strong results exist for particular tasks such as BIM compliance checking, semantic enrichment, document retrieval and building-characteristic extraction.5678 But cross-building benchmarks remain limited, external validation is uncommon, and document diversity makes generalisation difficult.9 Evidence for full multimodal engineering reasoning is weaker still: systems can identify equipment and extract standardised information while failing on spatial relationships and multistep reasoning.10
Nor is there sufficient empirical evidence that better document extraction directly produces measurable energy or maintenance savings. Studies generally support the operational usefulness of integrated information, but controlled field evidence remains limited.
The evidence therefore supports a narrower conclusion. AI can substantially reduce the effort required to turn building documents into structured information. Whether that information becomes trustworthy building knowledge depends on what happens after extraction.
How Struxiva applies these principles
Struxiva treats generative AI as an interpretation layer, not as the database of record. Documents enter the system as evidence. AI identifies candidate entities, attributes and relationships from heterogeneous sources, and those outputs remain claims until they pass validation.
The objective is not automated document reading. It is a traceable path.
Document → Claim → Validation → Building knowledge → Operational evidence → Decision. The dashed box is the only stage where a generative model decides anything, and nothing leaves it without a source attached.
The platform keeps the connection between a claim and its source, including where the information came from and how it was extracted. Terminology and units are normalised so that information from different documents can be compared rather than simply accumulated.
Cross-source agreement strengthens a claim, and contradictions remain visible. Engineering and dimensional rules provide deterministic checks where the problem allows them. Information that cannot be resolved automatically is routed for review rather than converted into artificial certainty.
Validated information becomes part of the building's machine-readable knowledge layer. This is where documents meet operational data. A sensor is no longer simply a time-series identifier. It belongs to equipment. Equipment belongs to systems. Systems serve spaces. Documents describe design intent. Operational measurements provide evidence of actual behaviour.
Analytics then move in both directions: from an anomaly to the equipment and documents that may explain it, and from a documented requirement to operational data that can test whether the building behaves accordingly.
Every arrow matters. The value emerges when all four are connected.
Key takeaways
- Building information is abundant but fragmented across drawings, BIM models, certificates, schedules, specifications, reports and operational systems.
- AI can substantially accelerate extraction and structuring, but current evidence does not support treating AI-generated interpretations as engineering ground truth.
- Extraction should produce traceable claims with provenance, not anonymous values.
- Validation, normalisation and cross-source reconciliation are necessary before extracted information becomes trustworthy building knowledge.
- Semantic representations allow machines to connect equipment, systems, spaces, sensors and documents rather than treating them as isolated records.
- Connecting this knowledge layer to time-series data turns static documentation into operational context.
- When sources contradict each other, surfacing the disagreement is a more useful output than a silently chosen value.
- The objective is not to make buildings produce more data. It is to make the information they already contain usable, connected and trustworthy.
Evidence & references
Fragmented building information and handover
- Tsay, G. S., et al. (2023). Ethnographic study of more than 2,000 BIM-for-facility-management compliance issues, identifying information that was incomplete, inaccurate, inconsistent or unintelligible. Frontiers in Built Environment. doi.org/10.3389/fbuil.2023.1117066
- Singh, J., & Anumba, C. (2023). Information sharing during commissioning and closeout, and its effect on operation and maintenance. International Journal of Construction Management. doi.org/10.1080/15623599.2023.2211409
- Chatsuwan, M., et al. (2025). BIM and facility-management information integration, with semantic inconsistency identified as a barrier. Architecture, 5(3), Article 74. doi.org/10.3390/architecture5030074
- Altwassi, E., et al. (2024). Building information interoperability and lifecycle information management. Buildings, 14(3), Article 611. doi.org/10.3390/buildings14030611
AI-assisted extraction and building information processing
- Li, H., et al. (2025). Domain-adapted multimodal retrieval and semantic building information processing, reporting approximately 30% improvement in document coverage and structural preservation and reducing semantic-model preparation from days to minutes. Advances in Applied Energy, Article 100261. doi.org/10.1016/j.adapen.2025.100261
- Iversen, O., & Huang, L. (2026). LLM-based BIM compliance checking reporting approximately 97% classification F1, with similarly high performance for dependency identification, tool selection and rule execution. Automation in Construction, Article 106707. doi.org/10.1016/j.autcon.2025.106707
- Chen, N., et al. (2024). Hybrid deep-learning, LLM and ontology framework for regulatory information in buildings. Buildings, 14(7), Article 1983. doi.org/10.3390/buildings14071983
- Forth, K., & Borrmann, A. (2024). AI-assisted semantic enrichment of BIM models. Journal of Building Engineering, Article 110312. doi.org/10.1016/j.jobe.2024.110312
Limits of multimodal AI and semantic consistency
- Ahmed, A., & Mohsen, O. (2026). Review of multimodal AI across the AEC lifecycle, identifying limited external validation, weak robustness to missing modalities, inconsistent data quality and a lack of standardised benchmarks. Advanced Engineering Informatics, Article 104863. doi.org/10.1016/j.aei.2026.104863
- Alampara, N., et al. (2024). MaCBench: multimodal reasoning for engineering information, with strong equipment identification and standardised extraction alongside weak spatial reasoning, cross-modal synthesis and multistep inference. arXiv. doi.org/10.48550/arXiv.2411.03782
- Jia, J., et al. (2024). IFC and ontology integration: review of semantic representation and enrichment. Buildings, 14(4), Article 911. doi.org/10.3390/buildings14040911
Extraction reliability and safeguards
- Daraqel, B., et al. (2025). Reliability of AI-assisted information extraction, including partial extraction and incorrect interpretation on more complex fields. Journal of Dentistry, Article 105846. doi.org/10.1016/j.jdent.2025.105846
- Gougherty, A. V., & Clipp, H. (2024). Evaluation of AI-assisted structured data extraction, with performance strong on simple discrete fields and weaker on complex information. npj Biodiversity. doi.org/10.1038/s44185-024-00043-9
- Shobanke, M., et al. (2026). AI extraction errors and their effects on engineering reliability estimates, with human review proposed as a safeguard. Advanced Engineering Informatics, Article 104803. doi.org/10.1016/j.aei.2026.104803
- Shorinwa, O., et al. (2024). Uncertainty quantification for detecting confident hallucinations in generative models. ACM Computing Surveys. doi.org/10.1145/3744238
- Kong, D., et al. (2026). Schema- and provenance-grounded information extraction. Sensors, 26(4), Article 1081. doi.org/10.3390/s26041081
- Chawla, R., et al. (2026). Parallel extraction and reconciliation for reducing hallucinations. IEEE Transactions on Computational Social Systems. doi.org/10.1109/TCSS.2026.3691181
Semantic representation and operational integration
- Donkers, A., et al. (2024). Semantic integration of building information and operational data. Building and Environment, Article 112433. doi.org/10.1016/j.buildenv.2024.112433
- Wang, M., et al. (2024). BIM-linked knowledge graphs for HVAC topology extraction and building energy modelling. Energy and Buildings, Article 115035. doi.org/10.1016/j.enbuild.2024.115035
- Wan, L., et al. (2025). Semantic building information for model predictive control, reporting room-temperature predictions within approximately 0.3°C. Applied Energy, Article 125597. doi.org/10.1016/j.apenergy.2025.125597
- Bjørnskov, J., et al. (2025). Semantic building models for hospital operation, reporting approximately 0.40°C average error across 45 sensors. Advanced Engineering Informatics, Article 103049. doi.org/10.1016/j.aei.2024.103049
- Pereira, F. de Andrade, et al. (2025). BIM metadata, Brick and SAREF for building control development, reporting substantial reductions in model-development effort.
Remaining integration barriers
- Du, S., et al. (2024). BIM/IFC interoperability and AI-ready building information. Buildings, 14(10), Article 3305. doi.org/10.3390/buildings14103305
- Benfer, R., & Müller, J. (2024). Semantic digital twins and heterogeneous building-system information. Energy and Buildings, Article 114637. doi.org/10.1016/j.enbuild.2024.114637
- Bampoulas, A., et al. (2025). Ontologies and building energy-management systems, including legacy-system integration difficulty. Energy and Buildings, Article 116817. doi.org/10.1016/j.enbuild.2025.116817
A building does not usually suffer from a lack of data. It suffers from a lack of usable knowledge. AI can help close that gap, but only when the path from document to decision remains visible. Everything in this article stops one step short of the person who needs the answer: the knowledge layer is built, validated and connected, and it is still a structure you have to know how to query. The last step is the one that makes it ordinary, where an engineer asks a question in their own words and the system answers from the validated graph rather than from a plausible-sounding guess, with the evidence for the answer attached to the answer.
In future articles, we'll explore how conversational interfaces, continuous digital twins, fault detection and predictive maintenance build on resolved identity, validated data and a semantic knowledge graph to create truly intelligent buildings.
Ready to see this in practice?
Talk to us about your building's documents.