AI pilots are everywhere in real estate. Almost none of them are working.

Property companies are piloting AI at a rate almost no industry can match. Ask how many of those pilots are actually delivering, and the number falls off a cliff.

What the data actually shows

• 88–92% of real estate investors, owners, and landlords have started piloting AI (JLL, 2026), up from under 5% in 2023
• Only 5% of those companies report achieving all their program goals
• 60% of companies across industries get almost no measurable value from AI investment (BCG)
• ~1/3 of enterprises have moved past piloting into actually scaling AI (McKinsey)
• 78% of CRE leaders expect AI to reshape their strategy within 3-5 years, but only 15% have moved past exploration (JLL Future of Work Survey, 2,200+ leaders across 21 countries)

That's the shape of the problem in five numbers: near-universal experimentation, and a strategy gap wide enough that almost nobody has closed it.

The adoption curve went vertical. The results curve didn't.

JLL's latest global technology survey puts real estate's AI pilot rate at 88-92%, a jump from under 5% just three years earlier. That's about as fast as an adoption curve gets, faster than cloud adoption, faster than mobile.

Then comes the second number, and it's the one that matters more: only 5% of those piloting companies report hitting all their program goals. Not most goals. All of them. The other 95% are somewhere between "partial results" and "nothing changed."

JLL's own diagnosis of the gap is direct: the constraint isn't the AI itself. It's whether the organization had quality data, working systems, and a change-management process ready to actually receive what the AI produced. Most didn't, and no amount of piloting fixes a readiness problem. It just runs the pilot on top of the same gap.

The same pattern shows up outside real estate too

This isn't an industry-specific failure. BCG's cross-industry research found 60% of companies get almost no measurable value from their AI spend, minimal revenue or cost impact despite real investment. Only 5% have built what BCG calls "future-built" capability: organizations that can turn AI into results reliably, not by accident. Another 35% are still on the way there.

MIT Sloan's interviews with finance leaders describe the mechanism behind that number in granular detail. AI pilots in finance functions routinely stall right after the proof-of-concept stage. A model looks promising in a demo, gets built out, and then sits unused the moment quarterly reporting pressure actually hits, because the underlying way people close the books, forecast, and explain variance doesn't change just because a new tool exists next to the old process. Dashboards get produced and refreshed. Decisions don't change.

McKinsey's own enterprise-wide survey backs this from a different angle: 88% of organizations report regular AI use in at least one business function, up from 78% a year earlier. But at the enterprise level, the majority are still experimenting or piloting, and only about a third report having actually begun scaling AI across the organization. Everyone is trying it. Not many are running it as something they can depend on.

What the 5% actually have in common

McKinsey's data isolates one factor more than any other as the thing that separates AI "high performers" from everyone else: a defined process for when a model's output requires human validation before anyone acts on it. Not a vague "use AI carefully" policy. A specific, working rule for when a person checks the output and when they don't.

That rule only works if there's something reliable to check the output against. If a model is confidently wrong because it was trained on a document version that's three revisions out of date, that's not a human-oversight problem. It's a "which version is actually current, and does anyone know" problem, and no validation process, however well-designed, fixes a question the organization can't answer in the first place.

This is the part most AI rollouts skip. Teams reach for a pilot before they've confirmed that the underlying documents, records, and processes the AI will read are actually organized, current, and structured. The pilot then either surfaces the gap in the worst possible way, a confidently wrong answer shipped to a customer or a team, or quietly fails to produce anything usable, and gets written off as "the AI wasn't good enough" when the real issue was upstream of the AI entirely.

What this means if you're one of the 88%

If your organization is piloting AI on top of documents, property records, and processes nobody has fully organized yet, the order that actually works is the reverse of how most rollouts happen: get the underlying information structured and current first, then let AI work on top of something solid. Skipping straight to the pilot, hoping the AI will somehow compensate for messy inputs, is exactly why most pilots stay pilots instead of becoming something the business actually runs on.

Frequently asked questions

Does this apply to smaller property portfolios, or only large enterprises?

The research cited here (JLL, McKinsey, BCG) is enterprise-heavy, but the underlying mechanism isn't a scale problem. It's a readiness problem. A 40-unit portfolio with disorganized handover documents has the same "which version is current" gap as a 4,000-unit one; it's just smaller in absolute cost, not different in kind.

Isn't this just an argument for hiring more IT staff?

No. More staff maintaining a disorganized system produces a better-staffed disorganized system. The gap JLL and McKinsey describe is structural: data quality and findability, not headcount. A smaller, well-organized data set that AI can actually trust outperforms a large team manually patching around a messy one.

If our pilot already failed, does that mean AI doesn't work for us?

Not necessarily, and this is the mistake the research warns about directly. A failed pilot usually means the underlying data wasn't ready, not that the AI itself is the wrong tool. Fixing the readiness gap and re-running the same use case is a very different proposition than concluding AI doesn't apply to your business.

How do we know if our data is "AI-ready" before piloting again?

The consistent signal across this research: can a person find the current, correct version of a given record in under a minute, without asking someone else? If the honest answer is no, that's the gap to close before the next pilot, not after.

---

CroVIN exists for exactly this step: organizing what your projects already produce into a structured, current, findable record, so that when AI (or anyone else) needs an answer, there's something trustworthy to give it.