Building code compliance review has long been one of the most labor-intensive checkpoints in the design and permitting process. A plan reviewer or design professional has to cross-reference a building model against thousands of provisions, many written in dense, exception-laden legal language rather than the structured logic a computer can easily parse.
As firms adopt BIM more broadly, Automated Compliance Checking (ACC) tools, some now enhanced with AI and large language models (LLMs), are being positioned as a way to close that gap. The research on these tools is consistent on one point: they speed up parts of the process. It’s just as consistent on a second point, that they aren’t a substitute for professional review yet.
Why Automation Matters Here
This isn’t just a convenience question. A NIST-commissioned study estimated that inadequate interoperability among software systems costs the U.S. capital facilities industry roughly $15.8 billion annually in avoidance, mitigation, and delay activities tied to data-exchange problems.
Manual code review, where a human interprets natural-language provisions against a design by hand, sits squarely inside that inefficiency. Automating even part of it has obvious appeal for firms working under tight schedules and thinner margins than they’d like.
Where the Automation Genuinely Helps
Rule-based ACC systems, which translate structured provisions into machine-readable logic and check them against a BIM model, have matured considerably. A review published in Scientific Reports describes a BIM-and-knowledge-graph-based system that maps design specifications against fire protection and civil design standards, then generates a structured review report, addressing what the authors call the “manual dependency” problem in traditional drawing review.
An IEEE review of BIM-based ACC systems traces this approach back decades, noting that neutral data formats like Industry Foundation Classes (IFC) have made rule interpretation, model preparation, and automated rule execution far more reliable, at least for well-defined, quantifiable provisions such as minimum ceiling heights, egress width, or occupancy load.
Machine learning has pushed this further in narrower domains. One large language model-driven ACC study cites a machine learning system built for accessibility compliance, using a convolutional neural network to flag issues like problematic ramp slopes, that achieved roughly 95% accuracy identifying violations. For high-volume, clearly quantifiable checks like that, automation is already carrying real weight.
Where It Still Fails
The gap opens around interpretation, not calculation. A framework study published in Automation in Construction benchmarked a general-purpose LLM directly against a purpose-built ACC system and found the LLM struggled with nested, multi-level logic, measurement units, and domain-specific exceptions.
It couldn’t reliably tell the difference between general Python and IronPython, the variant Revit actually uses, a small detail with real downstream consequences if it makes it into generated rule code. Researchers described the output as fluent, but short on the precision and validation traceability compliance work demands.
That pattern shows up elsewhere too. A study on fine-tuning LLMs for building regulation compliance points out that natural language processing still struggles with qualitative rules built on implicit assumptions, phrasing like “shall be ventilated,” where getting the interpretation right depends on expert domain knowledge, not pattern matching.
A systematic literature review in the Journal of Information Technology in Construction lands in the same place: real progress, but the ambiguity of legal text, the complexity of cross-referenced regulations, and thin training data still cap how far these tools scale unassisted.
What This Means for Design and Survey Firms
For architects, engineers, and other design professionals, the practical takeaway is that AI-assisted compliance tools work best as a first-pass filter, not a final signoff. They catch high-volume, well-defined violations quickly and free up staff time for the calls that still need a licensed professional: ambiguous provisions, cross-referenced exceptions, and project-specific context no rule set fully anticipates.
Firms leaning on automation for those interpretive calls without a documented human review step are carrying risk the tool itself can’t absorb, which is exactly the kind of exposure a professional liability policy is meant to cover.
