TL;DR: Foundation AI models are commoditizing fast. Frontier and open-source models are nearly equivalent in performance, and cost per task is falling five to ten times per year. The competitive moat for CRE owners has moved from the AI model to the owner-controlled data layer it runs on. The model can change tomorrow; the data layer underneath should not.
After a week of industry news, one pattern is clear. The AI conversation in commercial real estate has moved past which model is best, and most ownership groups have not caught up yet.
Stanford's 2026 AI Index compressed the performance gap between the best frontier AI and the best open-source AI from 8% in 2025 to 1.7% in 2026. DeepSeek shipped V4-Pro at roughly one-seventh the cost of comparable frontier models. According to a16z's 2026 enterprise AI report, 81% of enterprises now run three or more model families, up from 68% a year ago. The cost to reach a performance benchmark is dropping five to ten times annually.
For CRE owners, the practical meaning is straightforward: whichever AI vendor you evaluated last year is already interchangeable with several alternatives at similar or lower cost. The model layer is becoming a commodity input. What remains scarce, and what only your portfolio can produce, is the operating data, governance, and workflow design that sits above the model.
What This Means for Building Owners
CREtech's recent webinar framed the shift clearly: "AI is already doing the work in real estate." Operators described AI agents handling lease administration, broker research, internal operations, and enterprise workflows. Realcomm's featured conversation with executives at Primaris REIT and MIH Advisors carried the same message. The industry is no longer struggling to identify AI use cases. It is struggling to make them produce reliable results.
JLL's 2025 Global Real Estate Technology Survey spelled out the gap. 90% of CRE companies are piloting AI. Only 5% have achieved all of their program goals. The MLQ State of AI in Business 2025 report found that 95% of AI pilots fail across industries, with the inability to integrate AI with the right data and workflows cited as a major contributor.
That failure pattern is not a model quality problem. It is an owner data problem: the model cannot produce useful outputs when the inputs are fragmented, incomplete, or locked inside vendor systems that define basic terms ("vacant," "occupied," "made ready") inconsistently across the portfolio.
When the model layer commoditizes, three capabilities start to matter more than which vendor you bought.
Proprietary operating data. Every model on earth trains on the public internet. What is not on the public internet is what your buildings generate: tenant behavior, work-order patterns, energy and connectivity telemetry, leasing histories, vendor performance over time. Those signals become your competitive advantage only when they live in a form you own and can apply a new decision engine to. If they live inside vendor platforms on vendor terms, the vendor holds the differentiated asset.
Operational workflows with guardrails. A strong model dropped into a poorly designed workflow produces poor decisions faster. The Anthropic Claude incident reported by Tech Digest this week is instructive. An AI agent on a routine maintenance task wiped the entire database and all backups of PocketOS (a SaaS business) in nine seconds. The model executed its instructions correctly. The system around the model had no boundary preventing that action. In CRE, the parallel is granting AI agents access to building systems, work-order queues, or leasing workflows without clear permission boundaries, approval checkpoints, and rollback rules. That is an operational design problem, not a vendor selection problem.
An orchestration layer the owner controls. With 81% of enterprises running multiple model families, "we standardized on one AI vendor" is becoming a constraint rather than a strategy. The useful architecture is one that routes tasks to whichever model is appropriate, with the right data context, the right permission scope, and the right cost ceiling. That layer includes identity, lineage, retention, and audit. It does not come prebuilt inside any single vendor product.
What This Looks Like in a Portfolio
If a portfolio's data lives in a fragmented set of vendor systems, each with its own taxonomy, AI produces confident-sounding answers built on inconsistent inputs. An occupancy query returns different numbers depending on which system you ask and which definition of "occupied" that system uses. A utility analysis cannot compare assets because metering intervals, rate structures, and allocation methods are recorded differently building to building. The AI model works. The data underneath it does not support reliable portfolio-level decisions.
If a portfolio's data is consolidated, normalized, and owner-controlled, the picture reverses. Different decision engines plug in to the same governed data layer. Vendors can be swapped without losing operating history. Intelligence compounds across buildings because the data model is consistent. The portfolio becomes an environment where AI tools deliver returns rather than confident noise.
The Peak Property Performance® methodology addresses this sequence directly. The PPP 5C™ plan starts with Clarify (define what data matters, who owns it, what is portable), then moves through Connect (owner-controlled connectivity repeatable property to property), Collect (operating data normalized into a consistent model), Coordinate (govern identity, access, privacy, lineage, retention, and rules of use), and Control (enable any decision engine or workflow to act under owner permissions).
That last step, Control, is where most owners will encounter difficulty next. Control assumes the four steps before it are complete. If the data plane is a patchwork of vendor silos, granting an AI agent permission to act on the owner's behalf means delegating authority into a system the owner cannot fully audit or govern. OpticWise's Building of Things® (BoT®) provides the owner-controlled approach to data & digital infrastructure: consolidated, governed building connectivity so every device or system runs on a single, secure, segmented foundation. Property Brain™ is the owner-controlled intelligence layer at the property level. Portfolio Brain™ is the same model at portfolio scale. The AI model behind any of these can change. The owner-controlled data layer underneath persists.
If you don't own your data & digital infrastructure, your vendors do.
The Conversation This Week
If the last three executive AI discussions at your firm centered on which vendor to standardize on, the conversation has moved past that frame. The relevant questions now: which operating data is yours and which is your vendors'? Where does it live? What governance is in place when an AI agent takes action based on it? If you have not reviewed your data & digital infrastructure posture across the portfolio in the last 18 months, that review is the starting point.
Own your data & digital infrastructure. Operate with strategic foresight. Build for the long game.
Bill Douglas is the CEO of OpticWise, where he leads the company's mission to help commercial real estate owners take control of their data and digital infrastructure. With over three decades of entrepreneurial experience and a track record of leading companies onto the Inc. 5000 list, Bill brings a systems-minded, owner-first approach to everything OpticWise delivers. He holds a mechanical engineering degree from Georgia Tech and is a graduate of MIT's Enterprise Forum Entrepreneurial Masters Program. Bill is the co-author of Peak Property Performance (Fast Company Press).
Your Next Step
Complimentary CRE Data & Digital Review Session
One building. Map who owns what, where data lives, and where operational burden stacks up vs your KPIs.
Topic clusters
This article is part of the following OpticWise topic clusters. Each pillar page summarises the topic and links to related Insights pieces:
