Why do most AI projects fail at CRE firms?
Most CRE AI projects fail because they start from the technology, not the problem. Adi Shavit — an engineer who built AI systems for 25 years — argues the same trap sinks startups: teams fall in love with capability, then hunt for a use case. For an owner-operator, that looks like buying a tool and asking what it might automate later.
The fix is old and unglamorous. Name a specific broken thing — the leasing pipeline living in a spreadsheet, the rent roll nobody trusts, the broker update lost in email — and solve that. Everything else is a demo. According to the study covered in Why 92% of CRE firms pilot AI and so few reach their goals, most firms pilot; few reach their stated goals. The gap is rarely the model. It's the missing problem.
What does "start from the problem" mean for an operator?
It means you write down the artifact that breaks before you evaluate a single tool. The problem is not "we need AI." The problem is "our leasing tracker — the spreadsheet listing available space, prospects, and stage — goes stale the moment a broker updates a deal by phone."
Shavit's point, translated for CRE:
- Name the artifact. Rent roll, stacking plan (the floor-by-floor map of who occupies what), LOI pipeline, lease-expiration calendar.
- Name who suffers. The asset manager reconciling numbers at month-end; the principal who missed a renewal.
- Name the current workaround. Usually a spreadsheet plus email plus memory.
- Then ask what a tool must do to retire that workaround — not the reverse.
If you can't name the artifact, you don't have a project. You have enthusiasm.
Why does capability alone keep failing to land?
Because generation is now cheap and judgment is not. A model can draft a tenant follow-up, summarize a lease, or fill a pipeline row in seconds. None of that is the hard part. The hard part is being right about the number, the date, and the tenant — and having someone accountable when it's wrong.
Shavit's veteran read is that the impressive demo and the reliable production system are different animals. In CRE the cost of "wrong on camera" is real: a mis-stated RSF (rentable square feet) in a listing, a hallucinated rent escalation in an abstract, a renewal date off by a quarter. The tool that ships is the one that treats output as a proposal, not a fact.
How does "propose, never commit" prevent the usual mess?
Vantrow's spine phrase is propose, never commit: software stages an action, a human approves it, and everything lands on an audit trail. Applied here, the system drafts the pipeline update or the broker packet — it does not send, book, or overwrite the rent roll on its own.
That single rule handles most of what makes operators nervous about AI:
- The desk drafts a tenant follow-up; a person hits send.
- The system flags a lease expiring in 90 days; a human decides the outreach.
- A broker texts a deal update; the system proposes the tracker change; the leasing lead confirms.
Nothing autonomous touches money, tenants, or the system of record without a name attached. Governance is not a brake on speed — it's what lets you move at all.
What should you actually pilot first?
Pick one artifact with a clear owner and a short feedback loop. Good first candidates:
- Broker updates into the tracker. Let brokers text or forward updates; the system proposes the row change.
- Lease-expiration tracking. Surface what's expiring across the portfolio before it surprises you.
- Lead and voice-note capture. Deals captured on the go, filed without losing the details.
Avoid the reverse pattern — buying a broad platform and searching for something to point it at. That's the failure mode Shavit describes, dressed in CRE clothing.
How is this different from hiring an "AI evangelist"?
An internal champion who loves the technology is a symptom, not a strategy. The evangelist optimizes for adoption of the tool; the operator should optimize for a fixed artifact. When the metric is "did we use AI," you get pilots. When the metric is "is the rent roll now trustworthy," you get a result.
Judge tools by fit, not novelty. Ask what breaks, who owns it, and whether the system proposes rather than commits. That's the whole test.
FAQ
Q: Who is Adi Shavit and why should a CRE operator care? A: Shavit is an engineer who built AI systems for roughly 25 years and writes on why most AI startups fail. His central lesson — solve a named problem, don't chase capability — is directly useful to owner-operators deciding where to spend on software.
Q: What's the single biggest reason CRE AI pilots stall? A: They start from the tool, not the artifact. Firms buy capability and then look for a use case. Naming the broken thing first — the tracker, the rent roll, the expiration calendar — is what turns a pilot into a result.
Q: Doesn't "propose, never commit" just slow everything down? A: No. Generation is already fast; the risk is being wrong about a number or date with no accountable human. Staging actions for one-click approval keeps speed while putting a name on every change and an audit trail behind it.
Q: What should we pilot first? A: One artifact with a clear owner and fast feedback — broker updates into the tracker, lease-expiration tracking, or lead and voice-note capture. Prove it retires a real workaround before expanding.
Q: How do we know a tool actually fits CRE? A: It speaks in your artifacts (rent roll, stacking plan, LOI pipeline), it proposes rather than acts on the system of record, and it can name who approves each change.