The end-of-year AI roundups for commercial real estate all read the same way: long lists of what the technology can now do. On December's ChatCRE year-end edition ("The Top AI for CRE Questions of 2025, Pt. 1"), the questions ran to capability — things like "Can AI read a lease?" and "Can AI underwrite a deal?" Those are fair questions. They're also the wrong ones to lead with.
Capability stopped being the constraint a while ago. The system can read the lease. The real question — the one that decides whether AI helps your firm or embarrasses it — is what the system is allowed to do once it has read it.
What AI questions should CRE firms actually ask in 2026?
Stop asking what AI can do and start asking what it's allowed to do. The capability questions are mostly answered. The questions that matter in 2026 are about control: who approves an action, where it's logged, and how you catch a mistake before it reaches a tenant or a lender. Governance, not capability, separates firms that get value from firms that get burned.
Here's the shift in plain terms:
- 2025 question: "Can AI draft my tenant outreach?"
- 2026 question: "Can it draft the outreach and hold it for me to approve before anything sends?"
The first is about power. The second is about control. Only the second one protects your reputation.
Should CRE AI run autonomously?
No — not for anything that touches a counterparty, a document, or money. An autonomous agent is software that decides and acts on its own, without a human approving each step. A governed system is software that stages the action and waits for a person to approve it. For a firm whose name is on the deal, the second model is the only defensible one.
This is Vantrow's spine: propose, never commit. The system does the reading, the drafting, the matching, the flagging — the slow, tedious work. Then it stops and hands you a proposal. You approve or you don't. Every step lands on an audit trail.
The industry data supports the caution. McKinsey's The State of AI in Early 2024 survey found that 63% of organizations using generative AI regard inaccurate output as a relevant risk to their business — yet only 32% were actively working to mitigate it. In CRE, an inaccurate output isn't an abstract risk. It's a wrong number in a lender package.
How do I stop the AI from making things up?
You don't stop it entirely — you contain it. AI hallucination is when a system generates confident but false output: an invented clause, a plausible-but-wrong square footage, a citation to a document that doesn't exist. You reduce it two ways: feed the system only the context a task needs, and require human approval before any output leaves the building.
Two ideas do the heavy lifting here:
- Layered context — the practice of giving the system only the specific records a task requires, instead of dumping your entire archive into its memory.
- System of record — the single, authoritative place where a fact lives, so the system reads from one source instead of guessing across many.
Narrow context plus a required approval step means a hallucination gets caught at the desk, not discovered by a tenant. The mistake stays inside the building.
Is my firm's data ready for AI?
Almost certainly yes — you have more usable data than you think. A mid-sized CRE firm typically sits on years of executed leases, rent rolls, and email threads, plus whatever the county has recorded on every parcel it touches. That's a working corpus. The problem is rarely too little data; it's data scattered across drives, inboxes, and one partner's memory.
Concrete footprint for a typical firm:
- Executed leases going back 10–20 years, often as PDFs
- Rent rolls and CAM reconciliations by property
- Public county records — deeds, permits, assessments — for every parcel
- Years of deal correspondence in email
None of it needs to be perfect. It needs to be reachable. Pulling it into one system of record is the unglamorous work that makes everything downstream reliable.
What should a CRE firm automate first in 2026?
Start with the narrow, repetitive tasks where a mistake is cheap to catch — intake, follow-up drafting, record matching, document flagging. Avoid anything that commits your firm without review. The best first automation saves an hour a day and never sends, files, or signs anything on its own.
Good candidates:
- Lead and voice-note capture — everything logged in one place so details don't fall through.
- Governed outbound — the desk drafts the follow-up; a human sends it.
- County-record enrichment — the system pulls public records and attaches them to the right parcel.
Each of these proposes. None of them commit. That's the line to hold.