Say hello to an increasingly common scenario in our field: A CRE firm buys a round of AI licenses, sends out a rollout email, and calls it a strategy. What happens after that is mostly ad hoc and improvised.
According to MRI Software's Commercial Real Estate Industry Pulse Check, North America | 2026, which surveyed more than 550 CRE professionals, 62% of firms say they're preparing for AI in some way. Yet 54% offer no AI training at all. And across every role surveyed, from property managers to CRE executives, more than half say their main plan for AI over the next year is improving their own individual performance, not participating in any firm-directed workflow.
That's the actual failure mode behind most rollouts: Buying the license becomes the whole plan. Nobody decides who's accountable for how it gets used, what it's allowed to touch, or who checks the output before it reaches a client or an investor, so every employee ends up deciding those things alone, one prompt at a time.
Getting ahead of that means making three calls on purpose: what to hand to AI right now, what to keep away from it for now, and what has to be true across the firm before any of it goes out the door.
What to automate
The safest starting point is work that's repetitive, well-defined, and easy to check against a source document. This is where AI earns trust fastest, because a human can verify the output in less time than it would have taken to do the work by hand. These tasks include:
- First-pass document summaries: Leases, appraisal reports, purchase and sale agreements, offering memoranda, and other documents where someone still needs the two-minute version before digging into the source.
- Data extraction and organization: Pulling information out of PDFs and scanned files into a structured format someone still reviews before it's used.
- Routine internal drafting: Meeting notes, internal correspondence, first-pass market summaries—anything nobody outside the firm sees before a person edits it.
- Cross-document flagging: Catching inconsistencies, like a rent roll that doesn't match the lease abstracts behind it, and routing them to a person to investigate.
In all instances, the difference is that AI now takes the first pass: drafting the summary, extracting the numbers, flagging the inconsistency. A person still reviews the result before it goes anywhere—this task is just faster now that there's a draft to check instead of a blank page to fill.
What to leave alone
Some tasks shouldn't go to a general-purpose AI tool at all, at least not without a defined review step standing between the output and anything external. These include:
- Numbers that reach an investor or a committee: Anything that becomes a figure in an investor report or an investment committee memo without a human tracing it back to the source.
- Valuation judgment calls: Cap rate assumptions, exit pricing, anything an appraiser or underwriter is professionally accountable for signing off on.
- Deal terms and negotiation strategy: Anything client-facing that leaves the firm without a second set of eyes.
- Confidential data handling: Any workflow touching data a client or lender expects to stay confidential, unless the firm has actually confirmed where that data goes.
The pattern above comes down to which mistakes are cheap to catch early and which ones are expensive to catch late. A wrong first draft of a market summary costs an afternoon to fix. A wrong number that reaches a lender, or shows up in front of an LP, costs the firm's credibility, and sometimes the deal itself.
What has to be true before licenses go out
This is the step most firms skip, and it's the one that determines whether the first two categories hold up in practice.
- Someone owns AI use at the firm. Not "everyone," not IT by default, but a specific person or small group who decides what's approved, what isn't, and who to ask when a use case falls in between.
- There's a short, specific list of approved and off-limits use cases. Communicated in plain language, not buried in a policy document nobody reads twice.
- A review step exists before AI-assisted work reaches anyone outside the firm. The same way a junior analyst's model gets checked before it goes to a lender or an LP.
- The source data is actually reliable. A rent roll in one system, lease abstracts in another, and an ARGUS file nobody's touched since the last amendment mean an analyst can write a flawless prompt and still get an answer built on whichever version happened to be open. AI doesn't fix bad data spread across systems that don't talk to each other—it reflects that mess back, with an additional layer of unearned confidence.
Ask a general-purpose AI tool what cap rate an asset like this should trade at, and it'll return a plausible-sounding range pulled from whatever it's absorbed about the market generally. Ask the same question with the firm's own recent comps and current asking terms, pulled together from wherever they actually live and reconciled against each other first, and it can name the specific deals supporting that number, and where they diverge from the generic range. Same tool, same prompt. The difference is what's sitting underneath it.
Firms that skip straight to buying licenses are, in effect, asking every employee to make those calls alone. Some will make the right call, but the organization shouldn't be betting on that.
Naming an owner, writing the use-case list, and building in a review step are decisions a firm can make in a single meeting. Getting the underlying data clean is different. It's closer to infrastructure than policy, and it's usually the one still unresolved months after everything else got signed off.
That's the piece Truss was built to solve. Truss takes the data a firm already has, spread across ARGUS, Yardi, MRI, and Excel, and turns it into a foundation clean enough to actually put AI on top of, without asking a firm to rip out the systems it already runs.