Amid growing complexity, the building review process hasn’t changed. AI can catch what human reviewers miss – but only if it understands how buildings actually go together, writes Ron Zaum of Structured AI
An architect at a leading UK practice told me recently that if he goes on holiday, no one else at the firm knows how to check fire compliance on the drawings. He is the firm’s oracle on fire compliance. That, in effect, is the role the Building Safety Act has formalised in the Principal Designer.
At another, larger firm working across multiple jurisdictions, every new project starts with a team reading the relevant fire regulations again, from the beginning.
Documentation has grown faster than the practices producing it. A decade ago, the team on a high-risk residential building delivered a fraction of what the same team must deliver today. Half of all Gateway 2 applications fail validation on first submission, according to Building Safety Regulator transparency data published in March. The honest question, asked rarely, is whether the tools in our offices have scaled in step with the work.
The drawings became databases
A Stage 4 submission used to be a finite stack of two-dimensional drawings. It is now a flattened extraction from a metadata-rich Common Data Environment, in which compliance is structured data, every fire rating, every fall arrow, every embedment depth that the design must satisfy.
A principal designer reviewing for Gateway 2 now works across three monitors at once, PDF set, calc pack, fire strategy with the BIM model running in the background and Approved Document B propped by the keyboard. The work is cross-referencing six documents at once, by hand, across hundreds of pages.
The rulebook expanded in steps. Approved Document B Volume 2 ran 172 pages in 2006 and runs 226 today. The Building Safety Act made Gateway 2 a hard stop in October 2023; PAS 8671 made the Principal Designer personally liable.
Each amendment adds something the principal designer must master: the second-staircase rule on buildings above 18m, expanded fire-stopping at compartment penetrations, the BS 8414 façade test. The Principal Designer regime now puts personal liability for that integration fire, structure, thermal, weatherproofing into a single role, where it used to be distributed.
The bar moved while we were still building
The bar those drawings now have to clear has shifted in the middle of live projects. A 10-storey residential development in north London started on site in 2014. The Building Safety Act passed in 2022, mid-build. The team has spent the years since working through Gateway 2 and Gateway 3 retrospectively, simultaneously, under a third-party engineer appointed by the Building Safety Regulator, who releases queries in successive RFIs.
A single bracket on a single drawing in that project’s structural pack now requires verification against fire performance above 18m, thermal bridging at the slab edge, weatherproofing of the fixing, the structural connection itself and the fire-rated mortar added to conceal exposed steel where erection order made the painting unfeasible. One detail, five regulatory conversations on the same drawing and a reviewer is now expected to hold all of them in their head.
The hard part is the domain
This is where the tools haven’t caught up. Design and collaboration software has scaled Revit, BIM 360, Bluebeam while the review side is still eyeballs on a PDF. A chatbot can answer a question about a sheet but it cannot reliably tell you whether the fall arrows on a waterproofing detail resolve, whether an anchor meets embedment specification, or whether a thermal-break label has been erased between revisions.
“Plausible is not the same as compliant,” as Clifton Harness wrote in AEC Magazine this March.
The reframe is precise: AI catches what human reviewers miss only if it understands how buildings actually go together. A stone bracket on a sheet is a structural connection that has to satisfy a fire rating above 18m, a thermal break and weatherproofing all at once.
Encoding that understanding is a domain problem before it is a model problem. The template is mathematical: for every X, find Y; verify Y against Z; repeat across the page set. The practitioner’s expertise is in defining what X, Y and Z stand for. One UK practice working at the edge of tall stone construction has had to write the engineering rules itself, because the codes do not yet exist.
The other day, walking back from lunch with an engineer on my team, I pointed out the weep holes on a building. We had spent the morning encoding the rules for finding them, finding the masonry course, finding the openings and verifying the spacing against spec. He had never noticed a weep hole before; by the time we got back, he was finding them on every wall. The understanding that used to live in one architect’s head can now run on every drawing in the set. That is what catching up to the buildings looks like.
The post AI catches what human building reviews miss – if it understands them appeared first on Planning, Building & Construction Today.
