Deal Intelligence Is Only Valuable If Nobody Else Has It
The Sovereign Institute | Week 17
Protecting Proprietary Insights in Real Estate AI
---
CoStar Group is worth $35 billion. That valuation rests on a database of commercial real estate transactions, lease comps, and market analytics that brokers and firms uploaded freely over three decades, believing the tool worked for them. CoStar aggregated it, normalized it, and sold it back to the market as market intelligence. The firms that built the database had a service. CoStar had an asset.
AI providers are building the same database — on an accelerated timeline. Every acquisition analysis, every cap rate model, every off-market sourcing pattern uploaded to a US-hosted AI platform is contributing to a shared intelligence layer. The firms doing the uploading have a productivity tool. The providers have something more durable.
Deal intelligence is only valuable until someone else has it. Cloud AI is the fastest mechanism ever invented for ensuring someone else has it too.
---
How the Leak Works
Real estate firms do not believe they are disclosing deal intelligence. They believe they are using a tool. The distinction matters less than it should, because the data flows identically regardless of intent.
When an analyst at a major fund queries an AI platform to model cap rate sensitivity for a specific market, the query encodes the fund's acquisition focus, the target market, the analytical framework, and the assumptions the team is stress-testing. The platform's standard enterprise terms — OpenAI's terms publicly permit use of "aggregate anonymized data to improve our services" unless the customer is on an API plan with an explicit opt-out, not the default ChatGPT Teams tier — allow that usage pattern to inform model improvement.
Stripped of the firm's name, the cap rate model is still a cap rate model. The analytical patterns it represents still inform completions for every other user querying that platform about that market. The intelligence isn't exfiltrated in a way that shows up as a breach. It accumulates through normal operation.
92% of enterprise AI usage converges on OpenAI infrastructure, either directly or through tools that embed it — this is the documented finding from Kiteworks and LayerX's 2025 enterprise AI analysis. In commercial real estate, this means the deal analysis queries from competing firms are flowing through the same model, the same logging infrastructure, and the same terms of service. The statistical picture of multiple firms' acquisition strategies sits in one place, under one set of terms, subject to one company's decisions about how to use it.
The RealPage case made the structural risk visible. The DOJ investigated the property management software firm for antitrust violations after its AI rent-optimization tool allegedly coordinated pricing across competing landlords using aggregated data that all landlords had uploaded. The firms that contributed their proprietary rent data didn't just lose privacy. They became raw material for a government investigation. The AI tool that processed their data — and trained on it — was the mechanism.
---
The Terms Don't Protect What You Think They Protect
Real estate investment professionals sign enterprise AI contracts with data protection addenda and assume the protections are complete. Two mechanisms undermine that assumption.
The first is model training clauses. "No training" agreements govern what the provider does voluntarily with your data. They govern nothing about what the provider must do under a government order. The CLOUD Act — signed in 2018 — lets US law enforcement compel any American company to produce data stored anywhere in the world. The clause in your enterprise agreement cannot override a federal statute. An acquisition thesis analyzed through a US-headquartered AI provider is subject to US government access, regardless of what your data processing agreement says.
The second is provider acquisition. Every major real estate firm's deal pipeline is encoded in its cloud AI interaction logs. When that AI provider gets acquired — a routine event in the AI sector — the new parent inherits those logs. Terms of service governing data use typically include exceptions for corporate restructuring. The deal intelligence uploaded in 2023 follows the asset, not the original agreement.
The RealPage investigation, the CLOUD Act's reach, and acquisition exposure operate independently. All three can activate simultaneously, through different mechanisms, for different reasons, on different timelines. An organization that has addressed one has not addressed the others.
---
The Compounding Consequence
The intelligence ratchet turns in one direction. Once an AI provider's model has been trained on your deal analysis patterns, that training is irreversible. Canceling your subscription stops new contribution. The model that already learned your underwriting approach, your cap rate thresholds, your geographic acquisition focus continues informing completions for every other user who queries it. There is no mechanism to retrieve what was contributed.
This creates a timing problem that most real estate firms have not assessed. The decision to use cloud AI for deal analysis in 2022 is still generating competitive exposure in 2026. The queries about acquisition targets that preceded transactions three years ago are part of a shared training pool that competitors query today. The investment in proprietary deal origination — the broker relationships, the direct sourcing, the proprietary analytics — competes with a continuous data egress that the same organization funds through its AI subscription.
Blackstone manages a $326 billion real estate portfolio. Its competitive edge is not its capital — it's the deals it sees before anyone else. When analysts use AI to model potential acquisitions, they encode that proprietary deal flow into queries. The AI provider holds a statistical picture of what the firm is evaluating, months before any transaction is announced. The intelligence advantage that justified decades of relationship-building is on someone else's servers.
---
What the Regulatory Timeline Adds
The European enforcement layer is now set. TikTok's €530 million GDPR fine — assessed by the Irish Data Protection Commission in May 2025 for cross-border data transfers without equivalent protections — established the enforcement template for large-scale cross-border data flows. Real estate firms with European portfolios processing deal intelligence through US-headquartered AI infrastructure are structurally exposed to the same category of violation.
EU AI Act enforcement begins in August 2026. For real estate AI that makes or influences investment decisions, the high-risk classification requires complete audit trail infrastructure, bias monitoring, and human oversight mechanisms — all of which must be demonstrable by the deploying organization, not certified by the vendor. An EU real estate investment firm using US cloud AI for deal analysis has approximately 12-18 months to build compliant infrastructure before enforcement begins.
---
The Architecture That Keeps Intelligence Exclusive
The SIA standard's Hybrid Sovereign level — Level 1 — addresses the deal intelligence exposure without requiring a choice between AI productivity and data protection.
The Router classifies every query before it goes anywhere. A general research query about publicly available market conditions — vacancy rates from public reports, published cap rate averages, macroeconomic indicators — can use a cloud model at full speed and standard cost. A sensitive acquisition analysis — cap rate sensitivity modeling for a specific undisclosed target, deal structure analysis for a transaction in process, competitive positioning for an off-market bid — stays on sovereign infrastructure, processed by a validated open-weight model the firm controls.
The Vault stores the firm's proprietary market data, historical deal analytics, and underwriting models on infrastructure the firm operates. The AI uses that data for analysis. None of it reaches the provider's infrastructure, none of it contributes to training pools, and none of it is subject to the CLOUD Act's extraterritorial reach because it never sits on an American company's servers.
The Recorder logs every inference — what was queried, what data was accessed, what model responded, what was produced. When a regulator asks for the reasoning chain behind an investment decision, when an audit requires documentation of analytical processes, when a compliance review needs evidence of information barriers: the answer exists in infrastructure the firm controls.
Open-weight models — Mistral, LLaMA, and their successors — have reached performance parity with proprietary models for structured deal analysis tasks. The choice is not between competitive AI and sovereign AI. It is between renting intelligence from a shared platform and owning infrastructure that keeps the intelligence exclusive. For an industry whose business model is knowing what others don't know, the choice has a clear direction.
---
The Path Forward
Three specific actions address the exposure in order of immediacy.
The first is a query audit: map what your analysts are querying cloud AI platforms with. Not what they're supposed to query — what they actually query. Most real estate technology teams have never completed this inventory. The audit will identify which deal analysis workflows are generating continuous data egress and which are genuinely low-sensitivity research tasks. The differentiation between the two is the basis for a sovereign routing policy.
The second is enterprise agreement review. Standard cloud AI terms permit model improvement use unless you've opted out through specific configurations that most firms haven't requested. Review your current agreements for data use clauses, training opt-out mechanisms, and acquisition exceptions. The review is a legal project, not a technology project, and it establishes what protections you actually have versus what you assume you have.
The third is sovereign infrastructure deployment for the highest-sensitivity workflows. Deal analysis, acquisition modeling, and underwriting for specific undisclosed transactions are the categories that justify the infrastructure investment. General market research and publicly available analysis can remain on cloud models at Level 1 for the workflows that don't carry competitive risk. Sovereign deployment for the sensitive tier typically takes 8-10 weeks for a real estate investment firm with a defined scope.
---
Two Trajectories
The next deal cycle will produce two categories of real estate AI outcomes.
Firms that build sovereign infrastructure this year accumulate a data moat that compounds. Every deal analyzed stays exclusive. Every underwriting model stays proprietary. Every acquisition thesis encoded in AI stays inside the firm's perimeter. The intelligence advantage that justified the analytics investment remains theirs.
Firms that continue operating on shared cloud infrastructure donate their intelligence to a shared pool that competitors also draw from. The analytical frameworks they've developed, the market models they've built, the acquisition patterns they've encoded — all of it accessible to the same model their rivals query next week. The investment in deal intelligence continues. The exclusivity it was supposed to create does not.
The architecture decision is a compounding one. Made in 2025, it determines which trajectory the firm is on for the next decade of deal cycles. Firms that understood what information was worth — and built the infrastructure to protect it — will be the ones whose advantage persists when AI becomes table stakes for every firm in the market.
---
The Sovereign Institute publishes the SIA standard for AI deployments that keep organizational intelligence within the governance perimeter. Certified practitioners implement SIA-compliant infrastructure for real estate investment firms and financial services organizations requiring deal intelligence sovereignty.