Inside two days, two commercial-real-estate platforms shipped conversational, inbox-driven underwriting. Diald relaunched a plain-English engine that pulls zoning changes, permit activity and neighborhood sentiment into a pro forma, with a “Confidence Score” rating how well the cap-rate assumption holds up (on $1M in follow-on funding, $4.75M total; data licensed from Moody's, Green Street and LightBox). A day later Altrio launched Altrio Pro, which builds an AI-scored deal pipeline straight out of an acquisitions team's email inbox. Both are vendor launches with self-reported claims — read the numbers as marketing. But the design convergence is the story: the pitch is no longer “AI decides,” it's “AI assembles the pro forma and shows its work, and you adjust every assumption.” Diald →   Altrio →

Two scans of the packet's design sources turned up only content farms, vendor blogs and Midjourney-prompt listicles — nothing that clears the non-advertorial bar. This is the recurring pre-design gap, empty again for the fourth time in five weeks. Naming it beats padding; the freshest real pre-design development remains yesterday's TestFit-drives-itself-by-plain-language piece.

Architosh's SIGGRAPH 2026 report frames the question every firm principal should be asking: can we use capable AI without shipping our drawings to someone else's cloud? Two answers were on the floor. Nvidia's RTX Spark can run a 120-billion-parameter model with up to a million tokens of context locally on a laptop — enough that a small studio could keep an LLM, and the project data it sees, entirely in-house. And Steg AI showed forensic watermarking to prove provenance and protect drawings, images and documents from misuse. The summer's “own your data” thread — AIA Honolulu on training-data rights, CSI on machine-readable specs — is reaching the hardware: on-prem inference as an IP-sovereignty move, not just a speed one. Architosh →

The $2.5B Australian firm whose digital checklists became standard issue on construction sites has renamed itself Mitti and reorganized around two AI bets: computer vision — cameras on workers or in workspaces watching for human error and environmental risk — and insurance, now ~$70M in gross premiums and headed toward a third of the business. Founder Luke Anear says the company is now building its own AI infrastructure and training models in-house (against $197M revenue and a $49M FY25 loss, with profitability targeted by September). The tell for contractors: the value is migrating from recording what happened (the checklist) to watching in real time and pricing the risk — the same observe-then-act logic surfacing across the lifecycle this week. Forbes →

R-Zero launched Continuous Building Savings, layering onto any BACnet building-management system to flag operational faults continuously — but prioritized by financial impact, each finding grounded in ASHRAE Guideline 36 / 90.1 / 62.1 and described “in terms a contractor can act upon,” with the cost to fix it attached (company claims 20–40% HVAC savings; $170M raised, backers include La Caisse and John Doerr). The same day, Johnson Controls' OpenBlue added an agentic assistant that surfaces recommendations and routes work orders under human-in-the-loop control (a DATEV trial cited ~10% airside energy savings). Two launches, one architecture: the model doesn't run the building — it points at the ASHRAE-defined fix and lets an engineer sign off. R-Zero →   Johnson Controls →