An AI can build you a fast, clean, internally consistent underwriting model and still miss the one risk that decides whether the deal survives. The output will not flag the gap. It will look finished.
That is the pressure point worth sitting with as artificial intelligence moves from the edge of the underwriting process to the center of it. The adoption question is settled. Roughly two out of three commercial real estate professionals now use AI on a weekly or daily basis. What is not settled, and what actually matters, is where judgment lives once the machine is in the room.
What AI does well, and what it does not
AI is strongest where risk is standardized and legible. It extracts numbers from a rent roll faster than a person can open the file. It pulls comps, benchmarks expense lines, runs sensitivity tables, drafts the first version of a memo, and catches arithmetic that does not tie. On structured, repetitive, high-volume work, it is a genuine advantage, and the productivity gap is real enough to affect who wins competitive deals. This is where AI earns its place.
The trouble starts when confident, well-formatted output gets mistaken for a complete read on the risk. A model can be technically correct and still be a funhouse mirror: every number in its place, and the reflection distorted, because the machine can only reason about what it was shown. It does not know what is missing from the file. It has not walked the property. It cannot smell the deferred maintenance behind a fresh coat of paint, read a submarket that is turning before the data turns, or tell you how a sponsor behaves when distributions are at risk and the model breaks.
The industry seems to sense this. Firms have largely accepted AI for efficiency-driven work, while confidence in using it for underwriting and other high-stakes decisions remains limited. In one survey, sixty-six percent of professionals reported using AI weekly or daily, but only five percent said they trusted it enough to inform an actual deal decision. That is not hesitation. That is discernment.
Augmentation is not the same as outsourcing
The line to hold is between augmentation and outsourcing. Augmentation means the machine does the legible work faster, so you spend your judgment where judgment is required. Outsourcing means you let the output make the call because it arrived looking authoritative and you were short on time. The first makes you sharper. The second transfers your thinking to a system that carries none of the consequence.
And the consequence is the part that does not transfer. The model does not sign the offering documents. It does not sit across from your investors. It holds no fiduciary duty, to yourself or to the people whose capital you steward. When a non-standard risk is overlooked and the loss shows up, the machine bears none of it. You do. A tool that cannot be held accountable should not make the decision that creates accountability. One audit partner put it plainly in a recent interview: he would not hand an AI twenty million dollars to invest.
Where human expertise still decides
This is why the hardest risks are the ones that keep judgment human. Howard Marks calls it second-level thinking: not what the data says, which is available to everyone, but what the data does not say, what the consensus is missing, and what breaks first when the assumptions are wrong. AI is a first-level engine. It is fast and fluent at the surface read. Second-level work, on a complex or non-standard deal, is exactly where the machine is least reliable and where the person carries the liability. The non-standard risk is where AI adds the least and costs you the most.
None of this argues against using the tool. Refusing it is its own failure, slower and less competitive, and adoption is not reversing. The argument is for keeping the division of labor honest. Let the machine take the structured, repetitive load. Keep the human on the boots-on-the-ground knowledge, on the assumptions the deck presents that do not survive contact with reality, and on the final ownership of the output.
Closing thought
The question is no longer whether to let AI into your underwriting. It is which decisions you are still willing to own when the model is confident, fast, and wrong. On a standard deal, the machine may carry more of the weight. On the complex, non-standard risk, the one that decides whether the deal survives, the judgment and the liability stay yours.
Vessi Kapoulian
Breaking down multifamily underwriting one step at a time to create educated and empowered investors
P.S. If you would like a second set of eyes on a deal or want to sharpen your underwriting through a risk lens, feel free to connect with me.
P.P.S. And if you want to go deeper into analyzing multifamily investments step by step, my Mastering Multifamily Underwriting book and the Mastering Multifamily Underwriting program walk through this process in plain English, from acquisition to exit.
P.P.S. I recently did a pop up live on how I use AI in my own work. If you’d like the link to the recording, tap reply and let me know.
Sources:
• First American Data & Analytics and DealGround. CRE Industry Pulse Check. 2026 (reported by National Mortgage Professional).
• Keyway, in partnership with The Appraisal. 2025 State of AI Adoption in Real Estate Survey.
• Bisnow. Reporting on AI and underwriting practice, 2026.
• Howard Marks. Second-level thinking (The Most Important Thing).