Proptech
Nine in Ten Commercial Real Estate AI Pilots Are Failing. Here Is Why.
By Real Estate Wire Staff, . Real Estate Wire.
Commercial real estate has rushed into AI faster than almost any other property sector, and the results have been largely disappointing. Propmodo reports that JLL's 2025 Global Real Estate Technology Survey, drawing on responses from more than 1,000 senior decision-makers across 16 markets, found the share of teams running AI pilots climbed from below 5% to 92% in three years. Occupiers are now running an average of five pilots at once. Yet only 5% of organizations say they have achieved most of what those programs set out to do.
The failure is rarely about whether the AI can perform the task. It is about what happens after a pilot works in a controlled setting and has to become a reliable, economically viable production workflow.
Snehal Joshi, Head of Data Solutions at Hitech i2i, told Propmodo that the most common breakdown starts before the model is even evaluated. "One of the biggest reasons that AI pilots fail is inaccurate data," Joshi said. "Many solutions are not able to produce the level of accuracy that they claim, when this happens there will always be disappointment." The JLL survey supports that: more than 60% of companies must resolve fundamental technology problems, including duplicated functionality and dormant systems, before AI can be layered on top. Roughly 81% report at least three existing systems that are not producing expected results. Those conditions do not disappear when a new tool arrives.
Even when a model performs well in aggregate, the verification problem kills the economics. A system that reports 94% accuracy across a batch of lease abstractions gives a reviewer no way to know which individual outputs are wrong, leaving two options: check every result, or check none. Joshi put it plainly: "Some of the AI solutions don't provide enough savings compared to the manual process and once you add in double-checking the model it might even cost more." The JLL data adds pressure: 65% of organizations have faced technology budget constraints over the past two years, and return-on-investment requirements have tightened. A pilot that cannot demonstrate clear net savings does not get funded for a second phase.
The practical fix Joshi describes is field-level confidence scoring, where multiple independently trained models process the same documents and flag disagreements rather than producing a single aggregate score. When two models agree on a value, confidence is high and the field passes. When they diverge, that specific field goes into the review queue. Over time, the organization builds an empirical record. "You can also create a confidence score at the value level and then benchmark those scores," Joshi said. "Over time you can start to not check data with a high enough score." As the evidence base grows, the verification burden shrinks, and that shift is what moves the economics from marginal to defensible.
Model choice matters too. Generic large language models are more prone to hallucination on real estate documents because they lack domain context. Joshi points to subdivision as a concrete example: a generic model processing a property record may return what looks like a single parcel, while a domain-trained model should pick up from context that the land has been divided, with each portion carrying its own legal description, tax identifier and potentially separate ownership. Missing that distinction produces a record that looks complete and is confidently wrong, with nothing in the output to signal the error.
The last failure pattern is organizational. Companies treat an AI deployment as a software purchase when it functions more like an ongoing service relationship. "Most solutions are only offering the tech but the service that goes along with the tech is so important," Joshi said. "What most people want is an end-to-end solution, the tech and the human in the loop." Models require monitoring. Confidence thresholds need adjustment as the mix of incoming documents changes. Managing exceptions calls for people who can work across both the technology and the real estate subject matter beneath it. The JLL survey found that only 33% of the workforce feels adequately prepared for AI-driven change, a figure the article was still developing when the published text ends.
What the reporting does not establish is which specific use cases, lease abstraction, property record extraction, valuation support, are producing the 5% of successful scaled deployments, or whether success rates differ meaningfully by property type or market. Those gaps matter for any operator trying to decide where to concentrate a pilot budget rather than spread it across the 27 use cases the survey identifies.
The practical question for any team currently running pilots is whether they have a baseline at all. Without a clear pre-AI read on cost, turnaround time, accuracy and exception rates, there is no way to prove the tool moved the needle, and no case to bring back to a budget committee that has already tightened its return-on-investment requirements.
Source: https://propmodo.com/why-over-90-of-real-estate-ai-pilots-fall-short-of-their-goals/
Real Estate Wire provides news and information, not real estate, legal, or financial advice. Consult a qualified professional about your specific needs.