AI in commercial real estate: why 92% pilot and only 5% ship

Commercial real estate has never piloted anything this fast. It has also never shipped this little. JLL’s 2025 survey put numbers on the gap: 92% of corporate real estate occupiers are running AI pilots, and 5% have achieved all the goals they set. AI in commercial real estate isn’t stuck because the models are weak. It’s stuck because the agents are reading lease records and maintenance tickets the industry never made legible.

The money says the ceiling is real. Capturing it is a data problem nobody scoped.

TL;DR: JLL’s 2025 Global Real Estate Technology Survey found 92% of CRE occupiers piloting AI but only 5% reporting they’d met all their program goals. Morgan Stanley Research estimates AI could automate roughly 37% of real estate tasks and unlock up to $34 billion in efficiency by 2030. The opportunity is genuine. The reason pilots stall before production is that the two highest-value use cases — lease abstraction and maintenance triage — run on data the firm never standardized. Leases live as scanned PDFs, amendments, and side letters with no shared structure; maintenance requests arrive as free text. The agent inherits that mess and acts on it confidently. Fix the lease and maintenance data contract first, then turn the agent on.

The pilot wave is real, and so is the wall it hits

The enthusiasm is not the problem. JLL surveyed more than 1,500 senior CRE decision-makers across 16 markets for its 2025 Global Real Estate Technology Survey, and found both sides of the market deep into AI: 92% of occupiers and 88% of investors and owners are piloting. JLL counted 56 distinct AI use cases across the value chain, with the average firm running about five pilots at once.

Then the same survey reports the cliff. Only 5% of occupiers say they’ve achieved all of their AI program goals. Forty-seven percent have achieved two or three. The activity is everywhere; the finished, in-production, value-capturing systems are rare. JLL named the reason in plain language — firms have to fix data quality before AI can influence real decisions.

This isn’t a real-estate quirk. Gartner’s June 2026 forecast has spending on purpose-built AI agent software hitting $206.5 billion this year, up 139% from 2025 — and in the same breath projects that more than 40% of agentic AI projects will be cancelled by the end of 2027. The spend is racing ahead of the readiness. CRE is just an unusually clear example, because its two best use cases sit directly on top of its messiest data.

The opportunity is real — Morgan Stanley priced it

Be honest about the prize: it justifies the rush. Morgan Stanley Research analyzed the tasks performed across 162 REIT and commercial real estate firms — a combined $92 billion in labor costs and 525,000 employees — and concluded that about 37% of those tasks could be automated, worth as much as $34 billion in efficiency gains by 2030. The fattest pools are management, sales, administrative support, and installation and maintenance.

McKinsey, in research published March 4, 2026, pointed at the same upside from the operations side: early agentic-AI implementations saving more than 30% of time on maintenance tasks, lifting renewal rates 3 to 7%, and cutting lead response times by more than 90%. McKinsey’s framing matters more than its numbers, though. The argument was to redesign whole business domains rather than bolt AI onto isolated tasks. Read that as the operator read in a consultant’s voice: the value isn’t in the model, it’s in rebuilding the process the model plugs into.

So the ceiling is high and the analysts agree on it. The 92-to-5 gap is the distance between that ceiling and what firms are actually getting. That distance is data.

The broken layer: the lease isn’t a document, it’s a pile

Take the single most-cited CRE use case — lease abstraction. Pull the key terms out of a lease automatically and the downstream wins are obvious: cleaner rent rolls, faster due diligence, no analyst burning hours on data entry. Industry estimates put manual abstraction of one commercial lease at four to eight hours of skilled work. The case writes itself.

Here’s what the case skips. A commercial lease is not a document. It’s a base lease, plus amendments, plus side letters, plus estoppels, accumulated over years, often scanned, with clause language that varies by landlord and by the law firm that drafted it. An abstraction model reads the base lease beautifully. Then a renewal option that the base lease set at a 12-month notice window gets quietly changed to 9 months in a third amendment that lives as a scanned attachment nobody indexed. The agent extracts 12. It’s confident. It’s wrong. And nothing flags it, because from the model’s point of view it answered exactly the question it was handed.

Maintenance triage has the same shape from the other direction. The agent routes work orders based on the request — but the request is a tenant typing “heat not working” into a free-text box at 11 p.m. If severity is a free-text field instead of a defined value, there’s nothing stable for the agent to triage on. It guesses. McKinsey’s 30%-time-saved figure is real, but it’s earned on intake that was structured first.

The pattern is the one that shows up wherever automation lands on an un-fixed substrate: speed amplifies whatever the data already is. This is the same failure mode as autonomous field-service dispatch, where autonomy amplifies whatever the field data already is. A lease abstraction agent run across a portfolio of inconsistent leases doesn’t fix the inconsistency. It industrializes it.

High-value CRE use case	What the agent reads	Why it stalls before production
Lease abstraction	Base lease + amendments + side letters, often scanned	No shared structure; an amended term in a non-digitized document silently overrides the base lease the agent read
Maintenance triage	Tenant request, usually free text	”Urgent” isn’t a defined value; the agent has no reliable severity or asset field to route on
Renewal / critical-date tracking	Notice windows across many leases	Each lease defines the window differently; one missed amendment shifts a date by months
Portfolio analytics	Aggregated rent roll, lease data	Inherits every abstraction error upstream; one bad base year propagates into the model’s conclusions
Underwriting / due diligence	Lease + financials across systems	Source systems disagree on the same figure; no contract says which one is authoritative

None of these are model failures. Every one is a definition the firm never wrote down.

The working version: write the data contract, then point the agent at it

The fix isn’t to wait for better models. The models are ready. The fix is to stop treating the lease and the maintenance ticket as if the agent can make sense of them on its own.

Decide what the data means before the agent reads it. For lease abstraction, that means defining the fields you actually need, deciding which document controls when a base lease and an amendment conflict, and digitizing the side letters that override the terms hiding in plain sight. That’s the unglamorous work — and it’s the same integration-and-data-contract layer that decides whether any AI deployment works, applied to leases instead of CRM records.

Structure the intake, don’t parse your way around it. A maintenance triage agent is only as good as the severity and asset data it’s handed. Turn the free-text box into structured fields at the point of capture. It’s less exciting than the agent and it’s what makes the agent worth deploying.

Make drift fail loud. When a property management system renames a status, or a new landlord’s lease template uses different clause headings, the pipeline should stop and flag — not let the abstraction agent keep filling a rent roll with confident wrong values. A silent mismatch in a lease term is the kind of error that surfaces during a sale, which is the most expensive moment to find it.

Start narrow, with a human in the loop. One use case, one property type, one region, with someone checking the agent’s output before it touches a rent roll or dispatches a tech. Earn the autonomy on data you’ve reconciled. The 5% who finished their pilots didn’t have better models than the 92% who are stuck. They did this part first.

The CRE firms pulling ahead aren’t the ones who bought the most agents. They’re the ones who made their leases and their maintenance intake legible enough that an agent could be trusted with them. That work has no demo. It’s also the entire reason the demo ever turns into production. Making the process legible to a machine is the work worth doing first, and it’s the conversation worth having before the next pilot.