Legal tech spent late 2025 arguing about whether general-purpose models like ChatGPT can replace legal-specific AI, and whether law firms and in-house teams even want the same thing. According to Legal Tech Trends (#48, Q4 2025), those debates topped the quarter. CRE operators are next in line for the same argument — and it's the wrong one to have.
The vertical-vs-horizontal question feels important because it's easy to demo. But for an owner-operator deciding whether to let software touch the rent roll or send a broker update, the model matters far less than two other things: does the tool know your deals, and does it act on its own or wait for a human. That's the question legal tech mostly skipped, and the one CRE should start with.
Why is legal tech arguing about general-purpose vs. legal-specific AI?
Legal tech's fight is about whether a general model trained on everything can beat a tool built for one domain. General-purpose AI means a broad model like ChatGPT or Claude. Legal-specific AI means a tool fine-tuned and wrapped for legal work — clause libraries, citation checks, matter context. According to Legal Tech Trends #48, this was a defining Q4 2025 debate.
The appeal of the general model is obvious: one tool, cheap, improving fast. The counter-argument is that law runs on precision — a wrong citation or a missed clause carries real liability. So vendors argue their vertical wrapper adds the guardrails the base model lacks. Both sides are describing a real gap. Neither is naming what actually closes it.
What's the CRE version of that same fight?
The CRE version is "should I just use ChatGPT for my leasing work, or buy a CRE-specific tool?" It's the identical debate with different nouns. A general model can draft an LOI (letter of intent) or summarize a lease. A CRE-specific tool claims to know RSF (rentable square feet), NNN (triple-net) terms, and your stacking plan.
But the honest answer is the same as in legal: the model isn't the deciding factor. A general model that has your rent roll, your broker updates, and your deal history in context will outperform a generic vertical tool that doesn't. What decides quality is context and control — not which brand of model sits underneath.
Does the underlying model actually matter?
Less than the demo suggests. Model quality is now table stakes; the frontier models are close, and they all improve monthly. The difference between a useful answer and a wrong one is usually the input, not the intelligence. A model that doesn't know your NNN structure or which lease expires next quarter will confidently invent an answer regardless of how it was trained.
That's why "general vs. specific" is a distraction. The real variable is the context layer — the connective tissue that feeds the system your actual deals, tenants, and terms. Feed a general model your leasing tracker and it reasons about your portfolio. Leave it blind and even a legal-grade vertical tool guesses.
What should CRE operators ask instead?
Ask the governance question. "Governed AI" means software that proposes an action and waits for a human to approve it before anything is sent, changed, or committed — everything landing on an audit trail. Vantrow's spine phrase is propose, never commit. That's the question the legal debate skipped.
Ask, in order:
- Does it know my deals? Can it read my rent roll, leasing pipeline, and broker updates — or is it guessing?
- Does it stage or send? Does a human approve the LOI, the tenant follow-up, the CRM update before it goes out?
- Is there a trail? Can I see what it proposed, who approved it, and what changed?
- Can I turn it off cleanly? No mystery actions taken while I wasn't looking.
Those questions cut across general and vertical tools alike.
How does a governed CRE setup handle a leasing task?
It drafts, you approve, the trail records it. Say a broker texts an update on a deal. A governed desk reads it, updates the deal record as a proposal, drafts the tenant follow-up, and stops. You review the change, edit the draft, and approve the send. Nothing left your firm without a human signing off.
Compare that to an autonomous agent that updates the rent roll and emails the tenant on its own. The first time it misreads "5,000 RSF" as "50,000," you find out from the tenant. Governance isn't slower for its own sake — it's the difference between a mistake you catch in review and one your prospect catches for you.
General-purpose vs. CRE-specific vs. governed — who's each for?
Plainly, by fit:
- General-purpose model (ChatGPT, Claude), used raw: fine for drafting emails, brainstorming, and one-off summaries where you paste the context yourself and check the output. Not for touching systems of record.
- CRE-specific point tool: useful when it genuinely holds your data and one workflow — lease abstraction, say. Weak when it's a thin wrapper that still can't see your portfolio.
- Governed operating layer: for operators who want the system to act across leasing, CRM, and outreach and keep a human in the approval seat. This is Vantrow's category.
Candidly: most CRE firms start with a general model, hit the wall the day it invents a number, and then want governance.