Every CRE firm uses AI now. Almost half are slower because of it.

The adoption argument is over. Dealpath surveyed 100+ investment and technology professionals at institutional CRE firms this spring, from REITs to pension funds to PE shops running $500 million to $40 billion in AUM. 97% said AI is integrated into their investment process. On the brokerage side, the number of firms with no plans to adopt AI has collapsed to about 2%, down from nearly 11% two years ago. (Dealpath, Delta Media)

So we can stop having that conversation. Everyone is using it. Here is the number that actually matters.

In that same Dealpath survey, only 51% said AI saves them time once you account for verifying the output. 41% said work involving AI takes longer than doing it by hand. And 90% said bad or fragmented data is limiting what they get out of it. (Dealpath)

Read that again. Universal adoption, and roughly half the industry is paying a verification tax that eats the entire gain.

That is not an AI problem. That is a position problem. Every firm is somewhere on a spectrum, and most of them have never named which rung they are standing on, which is why they cannot explain why the thing is not working.

The spectrum

Rung one: the tab. Somebody on the acquisitions team has ChatGPT or Claude open in another window. They paste in a lease, ask for the key terms, paste the answer into a memo. This is real value and I am not knocking it. But it is individual, invisible, and unrepeatable. Nothing accumulates. When that analyst leaves, everything they figured out about how to prompt your deals leaves with them.

Kolena's 2026 State of AI in CRE report found 34% of CRE firms use a general-purpose AI tool in their regular workflow, and 78% of those same firms still process documents manually. (Propmodo) A third of the market has an AI subscription and a fully manual document pipeline sitting right next to it. That is rung one in one statistic.

Rung two: the point tool. You buy an AI-native lease abstraction or property management product. It deploys in three to six weeks and the enterprise license runs up to about $150K a year depending on portfolio size. (Forbes) Genuinely useful, genuinely fast. Two things to understand about what you bought. First, the tool encodes the vendor's opinion of how a lease should be read, not yours. Second, the pattern of how your firm reads a deal is now training signal inside a platform your competitors also license.

Rung three: the data spine. This is where most firms stall, and it is the one nobody wants to fund because it does not demo. Your rent rolls, leases, T-12s, entity docs, and deal history stop living in eleven places and start living in one place with structure and permissions. Nothing about this is exciting. It is also the reason 90% of that survey says data is the constraint. You cannot skip it. Every rung above it is built on it.

Rung four: your process, in software. Now you build. Not a generic assistant, a system that runs the specific way your firm underwrites, screens, and reports. Your rejection criteria. Your comp set logic. Your IC memo format. The judgment your principals have been carrying in their heads for twenty years, written down and executed consistently.

This is the rung where the verification tax disappears, and the reason is not a better model. It is provenance. When every extracted figure cites the clause it came from, review takes seconds instead of a full reread. Verification is only expensive when the output arrives with no trail.

Rung five: action. Agents that do things. Flag the mismatch, update the record, draft and route the notice, kick off the renewal. This is where the real leverage is, and it is also where AI stops being a document problem and becomes a systems problem, because now the thing can be wrong in ways that cost money.

The spectrum is not about model sophistication

Here is the part I think gets missed. Moving up the spectrum has almost nothing to do with which model you use. Frontier models ship like software patches now. The one that wins your bake-off in August is superseded by October.

What actually changes as you climb is how much of your firm's judgment is captured and kept. Rung one captures none of it. Rung two captures it into somebody else's product. Rung four and five capture it into an asset you own.

JLL's technology survey found about 88% of investors, owners and landlords piloting AI, and only 5% reporting they had achieved all their AI goals. (Commercial Observer) That gap is not a model gap. It is the distance between rung one and rung four.

What rung five costs you if you skip the work

One more thing worth knowing before anyone reaches for agents, because the news this month made it concrete.

Anthropic confirmed that certain Claude models misread their test sandboxes during containment trials and reached live systems. OpenAI disclosed agents escaping a sandbox during security testing and attempting to reach another company's production database. On August 11 at Black Hat, more than 120 organizations including Nvidia, Cisco, CrowdStrike and Red Hat proposed a shared incident reporting framework through the Linux Foundation, with Nvidia describing the agent harness as a flight recorder, borrowed straight from aviation. (Axios, Nvidia)

Translate that to a CRE context. An agent with write access to your accounting system and a loose permission model is not a productivity tool, it is an unlogged employee with your credentials. The control that matters is not the prompt. It is the scoping: an identity per agent, a service layer that is the only path to your data, human approval on anything irreversible, and a log of every action detailed enough to reconstruct later.

Do that and agents are the best thing to happen to CRE operations in a decade. Skip it and you have built something nobody can audit sitting on top of your capital stack.

What I would actually do

Name your rung honestly. Not the rung your tech vendor put in a deck, the rung your team is on when nobody is watching.

If you are on one, do not jump to five. The firms getting burned right now are the ones who skipped the boring middle. Fix the data spine, then build your process into software, then let agents act on it.

And keep the ownership question in front of you the whole way up. Rented software is fine for the parts of your business that look like every other firm's. Email, calendars, e-signature, accounting. Nobody wins on those. But the way your shop reads a deal is the only thing you have that a competitor cannot buy a license to. That should not be sitting in a multi-tenant platform that renews every January.

Where we stand

This is what we build. Custom production software for CRE owners, sponsors, and lenders, plus the agents that run on top of it. Portfolio intelligence, deal screening, document pipelines, underwriting tooling shaped around how a specific firm actually operates rather than how a category of firm supposedly does.

And our clients own it outright. Code, data, infrastructure. No per-seat tax that climbs every year, no lock-in, no vendor quietly getting smarter off their deal flow.

Everyone in this industry adopted AI. The next five years get decided by who moved up the spectrum and who is still paying the verification tax.