Three of the most powerful figures in artificial intelligence spent last week warning that the productivity base underpinning capital markets is being rebuilt faster than most economic models assume - and the sponsors raising capital today have yet to show their investors they have noticed.

Sam Altman, CEO of OpenAI, called for a national wealth fund, a robot tax, and automatic safety net expansions calibrated to AI displacement metrics, framing the moment as requiring a social contract on the scale of Roosevelt's New Deal. Demis Hassabis, CEO of Google DeepMind, told Fortune the commercial chatbot race is a distraction: the real event is AGI capable of genuine planning and reasoning, and the harder problem is arriving there without catastrophic outcomes. And Mustafa Suleyman, CEO of Microsoft AI, published an essay in MIT Technology Review projecting a 1,000x expansion in effective compute by 2028, arguing that humans chronically underestimate exponential change because we evolved for a linear world.

They disagree on priorities. They agree on scale and direction.

Every CRE pro-forma rests on assumptions about employment density, occupancy, rent growth, and the durability of demand in a given submarket. Those assumptions were calibrated against a labor market and a productivity curve that may no longer hold if Suleyman's exponential is directionally correct, if Hassabis's agentic systems arrive on his timeline, or if Altman's redistribution mechanics get political traction.

A deal closing today carries a minimum three-to-five year hold. The demand assumptions written into this quarter's underwriting will be tested against a market that has absorbed several further turns of the AI curve before the asset is refinanced or sold.

Investors are reading the same coverage and forming their own views. Sponsors presenting 2026 underwriting built on 2022 demand logic are signaling something about their analytical posture, whether they intend to or not.

The instinct, confronted with a fast-moving environment, is to reach for faster tools. AI workflows can process market data, generate rent analyses, and produce polished acquisition memos in a fraction of the time a junior analyst requires. The temptation to treat that speed as a proxy for accuracy is, in the current climate, close to irresistible.

It is also the field's most consequential error.

Gary Marcus, the cognitive scientist who has spent 25 years arguing that pure language models would never be reliable enough for serious analytical work, published three pieces last week making this point from a fresh angle. The AI systems showing genuine progress, he observed, are not the ones that scaled the largest. They are the ones that paired pattern recognition with explicit, classical reasoning that constrains and checks the output. Pure language models remain too probabilistic and too erratic to trust on their own.

In CRE operational terms, the trap has a name: The Wizard Fallacy.

The fallacy is the assumption that because an AI workflow produces a polished, formatted, professionally laid-out output, the output is also correct. It is the same error that has surfaced in every previous wave of CRE enthusiasm - from syndication to crowdfunding to the current AI moment. A confident presentation is mistaken for a verified result.

The mechanism is simple: AI models do not flag uncertainty. They generate the most plausible answer given their training, format it cleanly, and deliver it with the appearance of competent analytical work. In a multi-step workflow, each step's output feeds the next. Small errors compound silently into a final result that looks finished but is materially wrong. In a business where outputs carry fiduciary weight and shape capital allocation, the cost is not a rounding error.

The two problems are connected.

The macro environment is accelerating faster than most underwriting models have registered. The tools designed to help sponsors keep pace are capable of producing authoritative-looking errors at scale. The sponsors who navigate this well will be the ones who treat AI adoption as a verification discipline rather than a production exercise - who ask not how much the workflow can do unattended, but where the human checkpoints sit and whether they are adequate for the fiduciary stakes involved.

That discipline is not new. It is the same standard that has always applied to work of this kind. The current moment simply makes it easier to set aside.

***

Every cycle finds its evangelists. The syndication (crowdfunding) era had, and still has, its gurus - those who made the complexity of real estate investment look deceptively simple, and whose promises proved rather more compelling than their results.

AI has its own version: fluent, confident, and equally adept at making the difficult look effortless. The technology itself is not the problem. The incantations surrounding it are.

​The AI in Real Estate Accelerator, the executive program I run, is built on a different premise - that sponsors are better served by learning to implement AI in deliberate, sequential steps than by chasing solutions that look impressive in a demonstration and prove unreliable in practice.

Sponsored by the National Apartment Council of Canada, the next cohort is scheduled. Enrollment is open.

​Learn more here.​

Adam

P.S. Enroll before we start on May 26 and you will have free access until then to the GowerCrowd AI in CRE inner circle - a working session held every Thursday at noon where participants bring real workflows and leave with real solutions.

​More details here.​

This Week's Podcast

Guest: Michael Mandel, CEO, CompStak

Data Is the Real AI Advantage

In brief:

• AI in commercial real estate is only as powerful as the underlying data.
• Crowdsourced lease and sales comps create defensible market intelligence.
• Institutional asset managers are using AI primarily to benchmark portfolio performance.
• AI improves completeness and speed, but human validation remains critical.
• The next frontier is proactive deal origination using predictive signals.

​
Data First, AI Second

The central thesis of my conversation with Michael Mandel, co-founder and CEO of CompStak, is straightforward: AI does not create an edge in commercial real estate unless the underlying data is clean, normalized, and defensible.

Comstak is a commercial real estate data company, one that has spent 14 years building a crowdsourced database of lease comps, sales comps, loan data, and property-level intelligence. And the firms that will benefit from AI in this cycle will be the ones with the best structured data and the clearest understanding of how to apply it.
​

The Real Problem: Decision-Making Under Uncertainty

Every investor, lender, and asset manager faces the same structural question: How do I make better decisions, faster, with greater confidence? Historically, that process required:

• Calling brokers for comps.
• Manually parsing PDFs and rent rolls.
• Building internal Excel models.
• Writing market narratives from scratch.

AI changes the mechanics of that workflow, but not the objective. CompStak’s core innovation was crowdsourcing hard-to-find lease and sales comps from nearly 50,000 brokers, appraisers, and research professionals. That alone created market transparency in asset classes where true rent intelligence is notoriously opaque. AI then becomes the acceleration layer.
​

Where AI Actually Shows Up

There are two distinct categories of AI use inside CompStak:

1. Back-End Data Normalization

Thousands of comps arrive daily in inconsistent formats - Word documents, Excel sheets, scanned PDFs, even physical mail. Historically, much of that required manual abstraction. Today, document abstraction models dramatically reduce internal processing time while retaining a human-in-the-loop validation model.

2. Front-End Intelligence and Benchmarking

The more visible use case is AI-driven querying and benchmarking. Users can ask semantic questions:

• How are grocery-anchored retail rents trending?
• Where are Class A office rents outperforming?
• How do fried chicken restaurant rents in New York compare to Boston?

Behind the scenes, CompStak combines LLM classification with deterministic filters. That hybrid architecture limits hallucination risk while enabling flexible search across 1.3 million lease comps and 2.6 million sales comps. But the more sophisticated application is portfolio benchmarking.
​

The Institutional Use Case: Benchmarking Performance

For large institutional players - private equity funds, banks, sovereign wealth funds - the workflow is increasingly two-step:

1. Centralize and understand internal portfolio data.
2. Benchmark that performance against the market.

That second step is where external data becomes critical. Asset managers can now see:

• Whether they are underperforming market rents.
• Whether concession packages exceed competitive averages.
• How expiring leases align with above- or below-market positioning.

CompStak’s rent estimation models further extend this. For historical leases, the system estimates what the rent would be today based on hundreds of comparable comps, weighted by similarity. For underwriting teams, that materially improves modeling assumptions.
​

AI-Generated Market Narratives

One of the more subtle but powerful features is dynamically generated market reports. Rather than relying on analyst teams to manually draft quarterly summaries, the system synthesizes CompStak’s data into structured market narratives - complete with references to specific deals and market movements. For sponsors and advisors preparing investment memos, board updates, or lender packages, that reduces time while maintaining data-backed credibility.
​

A Question Sponsors Should Ask

Are you benchmarking your portfolio against the market, or against your own assumptions? Many mid-market sponsors still rely on broker-provided comps or isolated transaction history which can create blind spots. In contrast, institutional players increasingly integrate external data via APIs and Model Context Protocol (MCP) connections directly into internal AI systems. Asset managers can query portfolio performance conversationally while pulling in live market context.

​
The Next Frontier: Predictive Origination

Perhaps the most forward-looking initiative is lead generation for brokers. By analyzing past deal history, lease expirations, tenant growth signals, and market expansion announcements, CompStak is building tools that proactively identify high-probability opportunities. For tenant rep brokers, that means:

• Identifying expiring leases two years out.
• Cross-referencing tenant growth.
• Matching past relationships to new markets.

This shifts AI from passive analysis to proactive opportunity sourcing. For sponsors, the implication is broader: AI driven market intelligence will increasingly surface opportunity signals before traditional broker outreach begins.
​

Bottom line

The biggest impact of AI in commercial real estate is compression of workflow and expansion of insight. The firms that win will be those that:

• Treat data quality as infrastructure.
• Benchmark performance continuously.
• Integrate external market intelligence into internal decision systems.
• Use AI to surface signals, not substitute for judgment.

This discussion is particularly relevant for multi-cycle sponsors, institutional allocators, and senior asset managers seeking structural advantages rather than tactical efficiencies. The message from Mandel is simple; clean data is the moat - AI is the multiplier.

​Listen to or watch the full demo here.​
​

​​If this is your first time seeing this newsletter, please click here to subscribe.​

***

Connect with me on LinkedIn

​

Subscribe to my YouTube channel

​