149980-KeyzLogo-2-3-112717
Main Menu

How AI & Machine Learning Are Shaping CRE Valuation & Forecasting

/ Articles
Last Updated: October 2025 | Reading Time: 8 minutes | Author: KEYZ Commercial Real Estate Advisory Team

Almost every commercial real estate firm is now running AI pilots. Very few can tell you what those pilots produced. That gap is the real story in CRE valuation right now, and it matters more to property owners than any vendor demo.

This article covers what machine learning genuinely does well in commercial valuation, where it falls down, what the regulatory picture actually looks like, and what an owner should ask before trusting a number that came out of a model.

Adoption is broad but shallow

The headline numbers are real and they are worth understanding properly.

JLL’s October 2025 research found that 88% of investors, owners and landlords are already piloting AI, up from under 5% in 2023. That is one of the fastest technology adoption curves this industry has seen.

The same research found that over 60% remain strategically, organizationally and technically unprepared to scale AI beyond pilots. JLL’s separate 2025 Global Real Estate Technology Survey puts it more bluntly: 92% of CRE teams have started piloting AI or plan to this year, but only 5% report having achieved most of their program goals.

Deloitte’s 2026 Commercial Real Estate Outlook, based on 850+ senior executives at firms with at least $250M in assets under management, found 19% still in the early stages of their AI journey and 27% hitting implementation problems ranging from technical issues to lack of expertise. Only about 22% are using industry-specific software platforms, with another 20% relying on publicly available large language models.

Read those together and the picture is clear. Nearly everyone is experimenting. Almost nobody has finished. Anyone selling you certainty about AI valuation in 2026 is ahead of the evidence.

What machine learning is genuinely good at

The strongest evidence for ML in commercial valuation comes from peer-reviewed work, not vendor marketing.

A study published in The Journal of Real Estate Finance and Economics analyzed 7,133 individual commercial properties in the NCREIF Property Index across apartment, industrial, office and retail, covering 1997 to 2021. Two findings are worth every owner’s attention.

First, human appraisals carry more error than most people assume. The study found a mean absolute percentage error of 11.1% across all property types and years when appraised values were compared against subsequent sale prices. Accuracy was highest for apartments at 8.6% error and lowest for industrial at 12.5%.

Second, appraisals lag the market in a predictable direction. The researchers found appraisals “consistently lag actual sales prices, falling short of sales prices in bullish markets and remaining in excess of sales prices in bearish markets.” That second half is the one that stings right now. In a correcting market, appraised values tend to sit above what buyers will actually pay.

The useful part: that lag is structured, not random. Machine learning models can identify and correct for it. This is the clearest, best-documented case for ML in commercial valuation, and it is a genuinely valuable one. It is also considerably narrower than “AI predicts property values.”

Where it falls down, and why commercial is harder than residential

Residential automated valuation models work reasonably well because the data supports them. Thousands of comparable homes trade in a metro every year, with standardized attributes and public price records.

Commercial is a different problem. Comparable sets are thin. A 12,000 square foot flex building in Cerritos may have three or four honest comps in eighteen months, and each will differ on clear height, office finish, power, loading, and lease structure. Sale prices are often not public. Lease terms that drive value, including concessions, escalations and tenant improvement allowances, are usually confidential. A model trained on asking rents rather than effective rents will systematically overvalue assets in a market like the current one, where concessions are standard.

Deloitte’s outlook identifies the same constraint from the practitioner side, noting that the challenge is “finding usable, significant data without extensive extract, transform, load efforts,” and that real estate data often cannot be fed into models for training without substantial preparation.

There is a further point that deserves more attention than it gets. There is no public, standardized accuracy benchmark for commercial AVMs. Residential AVMs are independently tested and scored. In commercial real estate, no equivalent scoreboard exists from the Appraisal Institute, the Appraisal Foundation, or the major data providers. When a vendor claims their commercial model is accurate, there is no independent body verifying that claim, and no published methodology you can inspect. Treat accuracy claims accordingly.

The regulatory picture, stated accurately

There is a lot of confusion here, including on brokerage websites, so it is worth being precise.

The binding rule does not cover you. Federal regulators adopted Quality Control Standards for Automated Valuation Models, effective October 1, 2025. It requires AVMs to meet standards for accuracy, data integrity, independence and nondiscrimination. It applies to “a mortgage secured by a consumer’s principal dwelling.” Commercial real estate is outside its scope entirely. The quality-control floor that residential AVMs must meet does not apply to the models valuing your industrial building.

The appraisal guidance is new, and it is guidance. The Appraisal Standards Board adopted Advisory Opinion 41, “Use of Technology in an Appraisal or Appraisal Review Assignment,” on April 23, 2026. It addresses automated valuation models, regression and statistical software, and generative AI. Two things about it are commonly misstated. It was adopted in 2026, not earlier. And Advisory Opinions are explicitly non-binding: they do not establish new standards or interpret existing ones. AO-41 clarifies how an appraiser’s existing obligations apply when using these tools. Responsibility for the analysis and the conclusion stays with the appraiser, exactly where it was before.

The practical takeaway for owners: in commercial real estate, there is no regulatory body checking the model. There is only the professional standing behind the number.

What this means for your property

Southern California owners are making pricing decisions in a corrected market, which is precisely the condition under which the appraisal lag documented above works against you.

Inland Empire industrial asking rents sit around $0.98/SF/month NNN per Kidder Mathews for Q2 2026, down 4.85% year over year, with direct vacancy at 7.6%. CBRE reports $1.08/SF/month NNN at 7.4% vacancy for its IE Core submarket. CoStar data prepared for the Riverside County Economic Development Agency puts market asking rents at $1.03/SF/month, roughly 23% below the 2023 peak.

Three credible sources, three different numbers, because each defines the market and the building set differently. No model resolves that disagreement for you. Someone has to decide which comparable set actually describes your asset, and that is a judgment call.

For context on capital markets, Marcus & Millichap’s 2025 National Multifamily Investment Forecast reported a national average cap rate of 5.9% for trades completed between October 2023 and September 2024, up 120 basis points from the 2022 low.

Questions to ask before you trust a model-generated value

  • What data trained it, and does that data include effective rents? A model trained on asking rents will overvalue your asset in a market with standard concessions.
  • How many genuine comparables sit behind this number? If the answer is four, the model is doing interpolation, not statistics.
  • Can you explain why the number moved? If nobody can trace the drivers, the output is not usable in a negotiation or a dispute.
  • Does it account for lease structure? Two identical buildings with different escalation schedules and remaining terms are not worth the same amount.
  • Who is professionally accountable for this figure? For any formal purpose, that has to be a licensed appraiser, not a platform.

How we use these tools

We use data platforms for what they are good at: surfacing comparables faster, tracking lease expirations across a submarket, monitoring absorption, and flagging when an asking rent has drifted away from where deals are actually closing.

We do not hand a client a model output and call it a valuation. In a market where the published sources disagree by 10% on the same submarket, the work is deciding which set of comparables genuinely describes your building. That is judgment, informed by data. It is not the other way around.

If you want a straight read on where your asset sits against current comps, get in touch, or see what we currently have on the market.

Key takeaways

  • Adoption is near-universal but shallow. 88% of investors are piloting AI, yet over 60% are not ready to scale it and only 5% report hitting most of their goals.
  • Peer-reviewed research puts commercial appraisal error at 11.1% mean absolute percentage error, and shows appraisals run above sale prices in falling markets. That is the current market.
  • Commercial valuation is harder than residential for structural reasons: thin comp sets, private pricing, and confidential lease terms.
  • No independent accuracy benchmark exists for commercial AVMs. Vendor accuracy claims are unverified by any third party.
  • The binding federal AVM quality-control rule covers consumer dwellings only. It does not apply to commercial property.
  • USPAP Advisory Opinion 41 was adopted April 23, 2026 and is guidance, not a mandate. Responsibility stays with the appraiser.

Sources

Market data current as of Q2 2026. Figures vary between sources because each firm defines market geography and building sets differently. Nothing in this article is an appraisal or a valuation opinion for any specific property.

Scroll to Top