The most informative number in commercial real estate AI right now is a gap, not a level.
Writing in Propmodo on November 11, editor Franco Faraudo set two findings side by side. Citing a CRE Industry Pulse Check, he reports that 66 percent of commercial real estate professionals use AI weekly or daily, while only 5 percent said they trusted it enough to inform an actual deal decision. More than half use it for support while explicitly excluding it from final decisions, and another 17 percent use it only with heavy verification.
Sixty-six to five is not a slow adoption curve. It is near-universal use paired with near-total refusal to rely.
The pressure is real, and so are the gains
It would be easy to read the 5 percent as institutional caution for its own sake. The same reporting makes that reading hard to sustain.
Faraudo notes roughly 875 billion dollars of commercial mortgages coming due in 2026 against originations forecast near 805.5 billion dollars, which puts a large volume of refinancing in front of teams that were not staffed for it. He reports that a market study found 79 percent of financial institutions running AI pilots or operating AI inside lending workflows by July 2026, up from 34 percent in late 2025. On results, he cites institutions running purpose-built underwriting platforms reporting 40 to 60 percent reductions in analyst time per commercial loan, with application-to-decision timelines moving from the traditional five to ten business days down to 24 to 48 hours, and banks deploying AI underwriting reporting time-to-decision compression of 50 to 75 percent.
Those figures come secondhand. The article attributes several of them to unnamed market studies, so they are worth treating as directional reporting rather than as measurements an institution could reproduce. Even discounted, they describe a technology that is working on the document-extraction front end of a loan while being walled off from the decision at the end of it.
The constraint is admissibility, not capability
This is the part worth being precise about. A firm that uses a tool daily has already satisfied itself that the tool is useful. The 5 percent is therefore not a judgment about output quality. It is a judgment about what the output can be used to justify.
Those are different tests. "Is this analysis good" is answered by reading it. "Can this analysis support a credit decision I will have to defend" is answered by the record behind it: which document produced which figure, which version of that document, what the agent was permitted to do, and where a person intervened. Most deployments were built to pass the first test. The 5 percent is what it looks like when they meet the second.
Seen that way, the gap is not a trust problem in the psychological sense. It is a missing evidence layer, and it will not close because a model improves.
The requirements are already written down
Usefully, the list of what that layer has to produce is no longer guesswork, and it is converging from two independent directions.
From supervisors: Reuters reported in October, citing six people familiar with the matter, that regulators are asking whether banks can clearly track and audit AI agent decisions, what safeguards prevent agents from exceeding their authority, whether humans remain involved in key decisions, and how banks can shut down autonomous systems. The agencies declined to comment, and the reporting notes regulators are still gathering information rather than being prescriptive.
From the examination side: the Conference of State Bank Supervisors published its AI Supervisory Framework in September, explicitly to give institutions clarity on the types of questions and information a state examiner may request. It is discretionary, each state decides how far to adopt it, and it reaches state-chartered banks and state-licensed nonbanks rather than national charters. It is still the clearest published question set available.
Those two lists are close to identical, and they are close to identical to what a credit committee would need in order to move a tool from the support column into the decision column. The compliance artifact and the reliance artifact are the same artifact.
What this implies for the next year
If the binding constraint is evidence rather than capability, then the firms that convert heavy usage into actual reliance will not be the ones that licensed the strongest model. They will be the ones that built provenance, authority limits, intervention records, and a tested shutdown path around whatever model they already have.
That work is unglamorous and it is mostly plumbing. It is also the only thing standing between 66 percent and 5 percent.






