Balcony, a Hoboken proptech, told Commercial Observer that its Keystone platform for digitizing U.S. county land records has nearly doubled since January — from about $243B and 318,000 parcels to $400B+ and 600,000+ parcels — assembled through direct government partnerships rather than scraping. Chief business officer Alexander McGee says Keystone “indexes records around the way each office actually works, so those answers take seconds instead of hours.” The unglamorous point for anyone running AI site selection or feasibility: the model is only as good as the authoritative, machine-readable parcel and deed data beneath it — and that layer is being built one county at a time. Balcony expects to reach 3,000+ of the roughly 3,143 U.S. counties. (Self-reported growth; a traction milestone, not an audited deployment.) Commercial Observer →

In Common Edge, Chad Reineke argues that AI now lets a junior “test massing, produce renderings, summarize precedents, diagram environmental factors, and assemble a presentation” at speeds unimaginable a few years ago — and that this is precisely the hazard. Polished test fits, zoning summaries and feasibility diagrams are reaching client meetings and public hearings looking finished before anyone has interrogated what they leave out. His line for the file: “A zoning summary is not an entitlement strategy. A code search is not a life-safety approach. A rendering is not a building.” The fix he proposes is procedural — review AI output for assumptions and omissions first, not polish, and show the rejected options next to the finalist. (A bylined critical essay, not reporting; no firm names or metrics.) Common Edge →

A new preprint, ARCHER, takes aim at the black-box problem in automated code checking: most commercial checkers are proprietary, so a firm “cannot verify whether their regulatory intent can be accurately captured.” ARCHER instead runs a fixed planner–generator–evaluator loop to write inspectable Python checkers from a plain-language rule interpretation plus labeled BIM models. Two findings matter for practice: the deterministic multi-agent setup lifts mean accuracy 82% over naive single-pass prompting, and — the part that should interest any firm wary of sending project data to a vendor cloud — a self-hosted open-weights model reaches 97.8% of frontier-API accuracy at roughly a quarter of the cost, which the authors say makes “data-sovereign compliance checking practical.” (Preprint; dataset drawn from real Singapore compliance scenarios; code released on publication.) arXiv →

Procore agreed to buy reality-capture firm DroneDeploy for about $845M cash, announced the same afternoon it posted its first GAAP operating profit and Q2 revenue of $375M (up 16%). The logic is data: DroneDeploy brings 3M+ jobsites and roughly 20 trillion sq ft of visual built-world data plus 100,000+ labeled safety issues; Procore adds nearly 400M photos, 126M+ drawings and 10M+ RFIs and inspections from the past year alone. VP of corporate development Meg Baldini told Construction Dive that M&A is “one tool we use to move faster… accelerate our AI strategy” (Procore also bought Datagrid in January). The open question Construction Dive flags — and neither company would answer — is what happens to DroneDeploy’s integrations with Procore’s competitors. Either way it extends the incumbents-buying-the-AI-layer run: Carrier–75F, RealPage–Cherre, now this. (Figures company-stated; closes later in 2026 pending approval.) Construction Dive → · ENR →

At the APPA conference, NAU CIO Steve Burrell described a two-year push to pull hundreds of Flagstaff campus buildings into an AI-assisted platform built on a Willow digital twin — over the facilities team’s initial objection that “we have bad data; we can’t do this.” They started anyway, with utility bills, meter data and work orders. The payoff was concrete: after a frozen pipe did millions in damage in the forestry building, root-cause analysis flagged the same condition the following week in the business building, and the crew repaired it in real time, as it was happening. Burrell’s framing: “Energy savings pays the bill. But operational efficiency and risk avoidance is really the long-term game.” Note the counterpoint to the week’s data-moat stories — you don’t need a pristine corpus to start; you need to start. (Second-pipe save not dollar-quantified; insourcing ROI cited without figures.) Facilities Dive →