The Seam — Daily — July 30, 2026
THE SEAM · DAILY
Daily Briefing
Thursday, July 30, 2026 · Where AI meets the built environment
This week the thread has been about who owns the layer the AI runs on. Today it sharpens into a build-or-buy question — and an honest limit. Across all five segments, the firms pulling ahead aren’t the ones with the cleverest model; they’re the ones training on data they already own and keeping a human on the last decision. The counterweight comes from a founder who’d profit by saying otherwise: the models still can’t do the one thing design most needs — geometry.
01 · Pre-Development / Feasibility
A 57-year-old landlord builds its own underwriter
Security Properties, a Seattle multifamily investor, has put an in-house AI underwriting tool called FirstPass into live acquisition use. It ingests rent rolls and operating statements, pulls current Treasury rates, layers in deal-specific assumptions, and returns a first-pass operating model in under five minutes — work that took an analyst four to five hours. The tell is what they refused to do: rather than license a vendor, they trained it on 57 years of their own portfolio data, “not scraped market averages,” and every output is reviewed before any decision. The pitch isn’t headcount — it’s turning over more stones. And the line that should travel across every AEC discipline: the machine never has the last word. Multifamily Executive →
02 · Pre-Design
The generative-design CEO who says AI still can’t do geometry
In an interview worth reading against the hype, Zenerate co-founder Benji Shin — an ex-architect who ran high-rise feasibility studies — argues the headline models (ChatGPT, Claude, Gemini) “really can’t support geometry.” They can describe a floor plan in words but can’t guarantee that rooms don’t overlap, that the unit mix fits the footprint, or that setbacks and FAR all hold at once — the gap between a plan that sounds plausible and one that’s buildable. Where AI does pay off, he says, is collapsing the feasibility loop: weeks of architect–developer back-and-forth become a one-hour session, because the architect can finally see the pro forma driving the ask. Editing beats generation for now; whoever cracks true geometry wins the next round. New AvalonBay and Brazil (Tenda) deals ride on the editing, not the magic. KeyCrew interview →
03 · Design
Endra rebuilds MEP on its own data model — and demotes Revit
Stockholm’s Endra ($50M Series A, major engineering consultancies as clients) is the engineer-side version of the same bet. CEO Niklas Lindgren says Revit’s underlying data model is “too coarse” for deep automation, so Endra built its own granular 3D model and a “spatial AI” that routes conduit and ductwork clash-free, modeling entire electrical systems receptacle-to-transformer in a single source of truth. Two things matter for engineers: the workflow keeps them in the loop rather than handing back a black box, and Revit — in his telling — becomes an “orchestration layer” startups build around, not the authoring engine. Same lesson as the underwriting desk: control the model of your domain, don’t rent it. The electrical module launches Sept 14. AEC Business →
04 · Construction
McKinsey puts a number on it — and it’s about who owns the data
The report the industry has spent two weeks digesting: McKinsey finds AI could automate roughly 39% of nonphysical work in construction (and about 50% in architecture and engineering), across 150 workflows in 25 domains, with data entry, invoicing and equipment inspection shifting most by 2030. The framing is the useful part. Advantage accrues not to firms that “add AI” but to those that control proprietary project data, own their decision workflows, and can charge for outcomes rather than hours — and it points to Suffolk and Turner building in-house. Its build-vs-buy rule is quotable: build where your expertise is the product; buy where a vendor will out-invest you. It was never a tooling question — it’s a data-and-commercial-model one. Construction Dive →
05 · Post-Occupancy / Operations
A whole building-ops stack, retrofitted into an ordinary Tokyo tower
Naver has installed its “1784” building-operations platform in Tokyo Midtown Yaesu — its first overseas deployment, with NTT East and Mitsui Fudosan. What’s notable for owners: it isn’t a single robot but an integrated “spatial intelligence” stack — autonomous delivery robots, the ARC Brain control system that talks to elevators and doors, camera-based positioning that skips costly sensors, and a digital twin — dropped into a building with none of the robot-ready infrastructure of Naver’s HQ (the robots ride the freight elevators). Digital-twin facility management and congestion prediction come next. The post-occupancy pattern to watch: the operations “brain” is becoming a portable product you install — echoing Carrier’s move on 75F last week. Seoul Economic Daily →
General AI, in brief
Washington is putting real money behind AI permitting. The Technology Modernization Fund opened a call worth up to ~$200M for agency AI-and-permitting projects; the new 21st Century Road to Housing Act carries a $200M innovation fund rewarding jurisdictions that reform permitting and zoning; and HUD is seeding AI code-review adoption with grants. The proof it works is already local — Honolulu’s CivCheck cut small-residential permit decisions from 73 days to 32.5. For AEC, the slowest actor in the whole lifecycle — the authority having jurisdiction — is the one now being paid to adopt AI. TMF → Honolulu →
Watching
Bolt Graphics “Zeus.” A GPU pitched squarely at AEC/DCC rendering that deliberately skips AI tensor cores, claiming up to 17x cheaper path-tracing and targeting Revit, SketchUp, AutoCAD and Twinmotion viewports. A contrarian bet against the AI-driven GPU squeeze — silicon isn’t due until Dec 2027. Architosh →
MCP reaches the field. A week after the Open Design Alliance and Revizto opened BIM data to AI agents via Model Context Protocol, Trackunit’s new IrisX MCP exposes equipment and fleet data to ChatGPT, Claude and Copilot, while PM platforms (Linarc, Simpro) bolt on AI scheduling. The “query your project data in plain language” pattern is spreading from the model to the machine. For Construction Pros →
The Seam · AI + AEC · A personal project by Bruce Lanier
Daily · July 30, 2026
Don't miss what's next. Subscribe to Bruce Lanier: