Why Doesn't AI Actually Save CRE Teams Time Yet?

Dealpath's survey, published July 8, 2026, covers 100-plus investment and technology professionals, analyst through C-suite, at US and Canadian institutional CRE firms running $500 million to more than $40 billion in AUM.

Five numbers carry it. 97 percent have AI in the investment process. 90 percent say bad or fragmented data limits what it delivers. 51 percent say AI saves time once verification is counted. 41 percent say AI-involved work takes longer than manual work. And 43 percent blame fragmented data for AI falling short, ahead of hallucinated output.

A sixth number nobody quoted is the most revealing: 83 percent rated their own data infrastructure mostly or fully AI-ready. Most CRE firms believe their data is ready and also believe data is what broke their AI.

Name the interest first. Dealpath sells CRE deal-management software, so a survey blaming data infrastructure is a survey that sells Dealpath. Read the percentages as directional.

Run the arithmetic on one lease abstract. Abstracting a 60-page lease by hand takes an experienced analyst about an hour. A model does it in three minutes. Trust that output and you have cut 95 percent, which is where most vendor decks stop.

You do not trust it. Somewhere in those 40 fields a few are wrong and none are flagged. Commencement date right, escalation schedule right, option notice window quietly reading 9 months when the lease says 12. So you open the lease. Checking a field costs nearly what pulling it cost, because the expensive part was locating the clause. Call verification 70 percent of the hour: three minutes plus 42 is 45. You saved a quarter hour and paid with a second pass on a document you now read anchored to someone else's answer.

Improving the model barely moves this. Going from 90 to 95 percent accuracy halves the errors and changes your checking behavior by zero, because you still cannot tell which fields landed in the bad 5 percent. Accuracy gains pay nothing until they cross the threshold where a professional stops checking, and for anything a client signs that threshold is brutal.

Two exits, then. Drive errors under the threshold, or flag which outputs to distrust so verification narrows to a short list. The market is buying the second. IDC research for Sage, published July 2026 across 2,000-plus senior finance leaders, found roughly three in four would reject an AI tool with no human-readable reasoning trace, even at 99 percent claimed accuracy. Accuracy is the wrong unit of measure. Auditability is the one buyers price. That study also clocked the tax: finance teams lose close to 13 hours a week validating AI output, and 22 percent said verification eats over half of what AI saved.

The strongest evidence here is not a survey at all. METR published a randomized controlled trial in July 2025: 16 experienced open-source developers, 246 real tasks, randomly assigned to work with or without AI. The AI group finished 19 percent slower. Those same developers then estimated AI had made them 20 percent faster. Every CRE figure in this post is self-reported, so read that gap as a warning that self-reported savings run optimistic.

The CRE instruments agree from different directions. JLL's 2025 Global Real Estate Technology Survey (October 2025, 1,500-plus senior decision-makers across 16 markets) found 88 percent of investors, owners and landlords piloting AI and 92 percent of occupiers doing the same, while only 5 percent of occupiers hit all their program goals. First American Data & Analytics and DealGround surveyed 255 CRE professionals in April 2026: 66 percent use AI weekly or daily, 17 percent use it only with heavy verification, 5 percent trust it enough to inform a real deal decision.

Three instruments, three sponsors, one shape. Adoption is cheap. Trusted output is expensive.

A Seattle broker builds a market survey for a client with a Q3 2027 expiration in the Denny Regrade. The AI version takes fifteen minutes and looks like four hours of work: twenty buildings, availabilities, effective rents net of concessions, ownership flagged where there is loan stress.

Then the tax lands. Is that 14,000 square feet on floor 12 still available, or did it go under LOI in April with nobody updating the listing. Did the concession figure come from a signed comp or from the model smoothing between two numbers. Does the loan-stress flag attach to the real ownership entity or an affiliate three LLCs over. Each is a five to fifteen minute lookup, and you do all of them, because one fabricated free-rent number costs more credibility than the document saved in hours.

The cause is entity chaos. One tenant's footprint lives in a scanned lease PDF, an assessor record under a different legal entity, a listing with a stale suite number, and a broker's memory of the last renewal. Nothing resolves those into one tenant and one building, so a model reasoning across them produces confident synthesis resting on guesses you cannot see.

Four moves, all free.

Measure the tax before arguing about tools. For one week, log two numbers per AI task: minutes to generate, minutes to verify. Almost nobody separates those, and the ratio settles the argument for your desk. If verification runs past half of total task time, your files are the constraint.

Ask for transformation instead of facts. Models are unreliable at recalling what a rent was and reliable at restructuring material you hand them. Supply the document, ask for extraction into a fixed schema, and the output becomes checkable in seconds.

Make every extracted field carry its receipt: the exact quoted line and page number beside each value. That turns verification from re-reading a lease into skimming a column of quotes. Highest-return prompt habit in CRE right now, and it costs one sentence.

Keep an error log by field type rather than by tool. Three weeks in you will know the model is near-perfect on commencement dates and unreliable on option notice windows and CAM exclusions. Then you stop checking everything and start checking the two fields that break.

Three, and the third one bothers me.

Half a large sample reporting net savings from a technology adopted inside 24 months is historically fast, and calling 51 percent a failure requires a baseline nobody has stated. Fair. My answer is that the 41 percent posting net losses are paying real money today, and that is where the engineering work is.

Second, the data-quality diagnosis flatters everyone selling data infrastructure, including Dealpath and including BrokerHQ. There is a live scenario where models get good enough at messy documents that the resolution layer stops mattering. It has not arrived on CRE documents, and verification cost is what proves it.

Third, the honest one. The phrase verification tax is not mine. IDC and Sage were using it in finance by July 2026. A term converging across verticals independently suggests the dynamic is general and probably temporary, the way slow page loads were. Those get solved at the platform layer, which would leave a vertical specialist with a shrinking problem to sell against.

None of that changes the advice, because measuring your own verification ratio is free.

Casey Krueger, founder of BrokerHQ. My position, stated plainly enough to be held against me.

CRE spent 2025 and 2026 buying models and calling it strategy, and the bill arrived as unbilled hours. The 41 percent figure has little to do with adoption or training. It is what happens when you put a reasoning engine on top of four systems that disagree about who the tenant is.

The industry is also measuring the wrong thing. Every vendor conversation I have had this year benchmarks accuracy. No broker has ever asked me our accuracy rate and then behaved differently on the answer, because 92 versus 96 does not change whether they open the lease. What changes behavior is knowing which field to open the lease for. The product problem in vertical AI is making a model's uncertainty legible enough to safely skip most of the checking.

I build for Seattle tenant-rep rather than CRE generally because entity resolution is local. Knowing the entity on the King County assessor record is the same operator as the one on the lease is knowledge, and it goes stale in a market you do not walk. A general-purpose model can read anything and confirm nothing.

Disclosure with a test attached. BrokerHQ is building the resolved-data layer for Seattle tenant-rep, so treat this as an interested party's read, and here is the falsifiable version: deploy any AI tool on your existing document mess and measure the verification ratio before and after. If verification time does not drop, the model was never your constraint, and nobody selling you a better one can help.

Run it. If the number comes back against me, send it to me. I would rather be wrong in August than wrong in a roadmap.