Why CRE AI Pilots Fail—and What Actually Works

TL;DR

• 88% of CRE investors and owners are piloting AI. Only 5% have achieved all their program goals (JLL, 2024).
• The failure mode is almost always a data problem, not a model problem. AI cannot create structure from unstructured inputs.
• Firms seeing real AI returns link every AI investment directly to a P&L outcome and redesign workflows end-to-end.
• Zoning intelligence is the most underinvested structured data layer in the CRE stack—and the one with the clearest ROI case.

---

The Gap Between Piloting and Winning

The numbers from JLL's 2024 commercial real estate technology survey are striking in both directions.

On adoption: 88% of CRE investors, owners, and landlords are actively piloting AI technology. 92% of corporate real estate occupiers are running AI pilots. 73% of CRE professionals describe themselves as early AI adopters. 72% of global real estate investors have committed hard capital to AI-enabled solutions.

On outcomes: Only 5% of organizations achieved all their AI program goals. Just 47% achieved 2 to 3 goals. More than 60% of CRE firms remain primarily dependent on legacy technology despite active AI investment.

This is not a story about an industry moving slowly. Nearly every major CRE firm has a live AI initiative. It is a story about an industry that has invested heavily in AI and is not yet getting the returns it expected.

---

What the Research Shows About Why

McKinsey's analysis of AI adoption across industries—including real estate—identifies a consistent pattern separating high-performers from the majority. It is not model quality, compute resources, or even the sophistication of the use case.

It is two things:

1. AI investments are connected directly to P&L impact. Firms that see material AI returns define success in financial terms from the start: faster time-to-close, lower per-deal research cost, higher portfolio screening volume, reduced entitlement risk loss. They do not define success as "deployed AI successfully" or "employees using the tool."

2. Workflows are redesigned end-to-end, not patched. Layering a language model on top of an existing manual process—asking an AI to summarize a PDF an analyst would have read anyway—produces marginal improvement. Redesigning the workflow so that structured AI outputs replace manual steps produces transformation.

The firms in the 5% are not using better AI than the firms in the 95%. They are using AI differently, against better-structured data, in workflows that were purpose-built for AI-native analysis.

---

The Real Constraint: Data Quality

CRE is a data-rich industry in some dimensions and data-poor in others. The dimensions where enterprise-grade structured data exists have already been automated effectively:

The pattern is clear. Where structured data exists, AI has been applied successfully. Where data is unstructured, manual, or fragmented across 40,000+ jurisdictions, AI pilots fail to deliver—because no model can reliably extract consistent, accurate, queryable intelligence from inconsistent, inaccessible inputs.

Zoning is the largest remaining gap. It is also the one with the most directly measurable business impact.

---

What $430–550 Billion in AI Value Actually Requires

McKinsey's estimate of AI value potential in real estate ranges from $430 to $550 billion across the full value chain. Agentic AI specifically—systems that can execute complex, multi-step property research and analysis autonomously—could generate an additional $430 to $550 billion in annual value across real estate, construction, and development globally.

These numbers are not theoretical. They reflect measurable improvements in operating efficiency, asset selection quality, capital deployment speed, and risk characterization. McKinsey documents a retail HVAC optimization case where AI across 600+ stores generated $1.38 million in electricity savings, producing a $24.46 million valuation impact.

The specific applications that generate value are not generic AI deployments. They are purpose-built systems applied to specific, high-leverage data layers where structured intelligence replaces manual research.

Zoning analysis fits this profile precisely. The cost of manual zoning research is measurable (3 to 8 hours per parcel, $48,000 to $128,000 annually for a team screening 200 sites). The AI replacement is technically feasible. The P&L impact—faster underwriting, more deals screened, lower entitlement risk loss—is directly quantifiable. It is a textbook high-ROI AI application, blocked only by the absence of structured input data.

---

Why Generic AI Tools Don't Solve It

Some CRE firms have attempted to use general-purpose large language models to analyze zoning documents. The approach fails in predictable ways:

Hallucination on dimensional standards. Language models generate plausible-sounding text. When asked to extract a specific FAR value from a 200-page zoning ordinance, a general-purpose model may return a number that appears in the document in a different context, or may generate a value that is not in the document at all. For a financial decision, a hallucinated FAR value that inflates buildable square footage by 20% is not a minor error.

No cross-jurisdiction consistency. A general-purpose model asked about zoning in Boston and then in Dallas will apply different terminology, make different inferences, and produce results that cannot be systematically compared. Enterprise underwriting requires consistent, comparable data across jurisdictions.

No temporal awareness. General-purpose models are trained on data with a cutoff date. They cannot tell you whether the zoning rule they are citing has been amended in the past 18 months, or whether it has been preempted by a new state law.

No compound query support. A language model can answer a single question about a single parcel with reasonable accuracy on a good day. It cannot efficiently answer "which of these 500 parcels meet these 7 zoning criteria simultaneously?" without purpose-built data infrastructure underneath.

The firms achieving real AI returns in CRE are not using generic models against unstructured sources. They are using AI systems built on structured, maintained, domain-specific data.

---

The Infrastructure Investment That Unlocks the Rest

The 5% of CRE firms that achieve all their AI program goals share a common foundation: they have invested in the data layer before the application layer.

The sequencing matters. An AI underwriting system is only as good as the data it processes. An AI site selection tool is only as fast as the underlying data allows. An AI portfolio risk analyzer can only surface risks that are represented in structured form in the data it accesses.

For CRE firms currently in the 95%—investing in AI and not yet seeing the returns—the diagnostic question is not "which AI model should we use?" It is "which data layers in our stack are still unstructured, and which of those has the highest P&L impact if we fix it?"

For most institutional investors, developers, and enterprise property platforms, zoning intelligence answers that question.

---

ZoningGraph Team

ZoningGraph builds the structured zoning intelligence infrastructure that enterprise CRE platforms, institutional investors, and developers need to convert AI investment into measurable returns.