A new MIT survey, summarized in commercialobserver.com's coverage, finds that the overwhelming majority of organizations are still getting nothing back from generative AI, even after a year of heavy enterprise spending. The pattern hits commercial real estate especially hard, where pilots are everywhere and production wins are rare.
The reason isn't the model. It's what the model doesn't know about your business.
The short version
- MIT, 2025: 95% of organizations report zero return on generative AI; reported enterprise investment sits between $30 billion and $40 billion.
- JLL, 2025 Global Real Estate Technology Survey: roughly 88% of investors, owners and landlords are running AI pilots, while only 5% say they've cleared their targets.
- The blocker most often named is fragmented data without context — what some call data debt.
- MIT says the tools that win are the ones that live inside workflows and remember prior work.
- McKinsey finds staff are using generative AI about three times more than managers think.
Background: the AI wave that didn't reach the P&L
AI budgets across the corporate world have been climbing for two years, and real estate has been a willing test bed. Lease abstraction, underwriting memos, comp pulls, tenant screening — every one of those jobs looks like the kind of task a large language model should eat for breakfast. So firms signed the contracts, paid the invoices, and waited.
The result, as both MIT and JLL now document, has been quietly disappointing. Not catastrophic. Not dramatic. Tools get adopted, then shelved. Subscriptions get renewed, then ignored. The actual work still happens in Excel. The gap between activity and outcome is the whole story.
Our analysis
The MIT report's diagnosis is the one to take seriously. Off-the-shelf foundation models don't keep a running memory of your firm. They forget last week's deal. They never learn how your shop weights tenant credit, reads a submarket, or handles recoverable expenses. A tool with no business history can't sharpen with use, so it doesn't.
The cleanest illustration is the lease abstract. Two firms can review the same document and reach different conclusions for entirely valid reasons. One prioritizes rollover risk; another focuses on replacement cost. One calls a submarket early-cycle; another calls the same submarket overbuilt. A general-purpose chatbot cannot make those calls. That is not a model failure. It is a missing context.
The next stop on the failure curve is wrapping a generic model inside an "agent" and calling it precision. An agent built on top of a foundation model is fancier than a raw prompt, but fancier is not the same as right. If the central intelligence was never trained on your work, no layer of orchestration around it will produce real judgment.
The McKinsey point is the one that gets overlooked. People are already using generative AI at roughly three times the rate their leaders estimate. The resistance inside firms isn't ideological. It's operational. Staff reject tools that slow them down or force them to re-explain the same background every session. A system with institutional memory kills that re-explaining and earns its seat.
A worked example
Say you run a mid-sized investment shop closing 30 deals a quarter. You license a top-tier foundation model, point it at your lease PDFs, and request abstracts. The output looks tidy. Then your senior analyst spends two extra hours per deal correcting what the model missed: tenant credit nuance, renewal probability, and recoverable expense treatment. By deal four, the model has already forgotten everything. By deal thirty, your team is back in Excel.
Now run the same shop with a memory layer fed by every prior abstract, every correction, and every decision flag. By deal thirty, the system is surfacing renewal risks your analyst would've missed on the first pass. Same underlying model. Different outcome. That's the bet the survivors are making.
What to watch next
Two signals matter in the next few quarters. First, watch the gap between announced AI partnerships and disclosed production outcomes in earnings calls — once companies start breaking out AI cost lines in filings, you'll know who's spending without shipping. Second, watch which CRE-specific vendors begin publishing per-seat productivity metrics rather than vague adoption stats. That's the first sign the context-first thesis is winning.
Bottom line: the firms pulling ahead with AI in real estate are not the ones licensing the biggest foundation model. They are the ones building the deepest context layer and letting the system keep learning. Everyone else is funding the other 95%.
Key takeaways
- MIT, 2025: 95% of organizations see zero return; spend sits at $30–$40 billion.
- JLL, 2025: 88% of real estate firms are piloting AI, 5% are hitting goals.
- The recurring failure mode is missing context, not a weak model.
- Survivors build memory; laggards buy the biggest foundation model.
Frequently asked questions
What does MIT say about generative AI ROI?
MIT's 2025 State of AI in Business report finds that 95% of organizations are not getting any measurable return on generative AI, despite reported enterprise investment of $30 billion to $40 billion.
How many real estate firms are piloting AI?
JLL's 2025 Global Real Estate Technology Survey reports about 88% of investors, owners and landlords are running AI pilots, while only 5% say they've fully achieved their AI goals.
Why do AI pilots keep failing in real estate?
Both MIT and JLL point to fragmented data without context — sometimes called data debt — rather than limitations in the underlying model.
What do the AI winners actually do differently?
The MIT study finds the successful tools are the ones that embed inside workflows, remember prior work, and adapt to context over time — what some in the industry call a company brain.
How much are employees really using AI?
McKinsey research cited in the article suggests staff are using generative AI at roughly three times the rate their leadership assumes.
