The speed of AI adoption in commercial real estate has been remarkable by any measure. JLL’s 2025 Global Real Estate Technology Survey, which gathered responses from more than 1,000 senior CRE decision-makers across 16 markets, found that the share of teams piloting real estate AI jumped from under 5% to 92% in three years. Occupiers are running an average of five real estate AI pilots simultaneously across roughly 27 identified use cases. Technology budgets have been reorganized around AI initiatives, with the top five spending priorities all tied to implementing AI in commercial real estate or preparing for its impact.

The results have not kept pace with the enthusiasm. Only 5% of organizations report having achieved most of their program goals. The rest are still experimenting, still evaluating, or quietly shelving projects that never produced what the business case promised. The gap is rarely about whether AI can perform a task. It is about what happens after a pilot works, turning a promising proof of concept into a reliable, scalable, and economically viable production workflow. That gap is where the real failure patterns repeat, across companies, property types, and vendors.

Before any of that gets diagnosed, most teams skip a step, establishing a baseline. Without a clear read on current cost, turnaround time, accuracy, manual effort, and exception rates, it is hard to prove a pilot moved the needle, let alone build the case to scale it. That baseline should exist before the technology conversation starts, and it is the benchmark every failure pattern below gets measured against.

Real Estate Data Quality Can Make or Break an AI Pilot

The most frequent point of failure is not the model. It is what the model is reading.

“One of the biggest reasons that AI pilots fail is inaccurate data,” said Snehal Joshi, Head of Data Solutions at Hitech i2i. “Many solutions are not able to produce the level of accuracy that they claim, when this happens there will always be disappointment.”

The JLL data supports that diagnosis. More than 60% of companies must address fundamental technology issues such as duplicated functionality or dormant systems before they can fully leverage AI, and 81% report at least three existing systems that are not generating expected results. Those conditions do not disappear when an AI tool is layered on top of them. They propagate.

A pilot designed to prove that AI can extract data from property records will produce output quality that reflects the source material. When the inputs are inconsistent, incomplete, or structured differently across jurisdictions, the accuracy falls below what was promised in the demo, and the project loses credibility with the stakeholders who approved it.

Knowing When AI Is Wrong Matters More Than Being Right

Every AI system produces errors at some rate. That is documented, expected, and manageable. What determines whether a pilot survives is whether the system can identify which outputs are likely to be wrong.

“Some models will tell you that a certain amount of its answers are inaccurate but for it to be useful it has to tell you when of the values are the ones you have to check,” Joshi said.

An aggregate accuracy rate is not actionable. A model reporting 94% accuracy across a batch of lease abstractions tells a reviewer nothing about which 6% require attention, which means either every output gets reviewed or none of them do. Neither option works. Reviewing everything eliminates the efficiency gain. Reviewing nothing exposes the organization to errors in financial models, title work, and compliance filings.

When Verification Costs More Than the Manual Process

The verification problem is where a surprising number of pilots die on economics rather than performance.

“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,” Joshi said.

This is the calculation that rarely appears in a pilot proposal. A tool that processes documents faster than a human still requires someone to confirm the results, and if that confirmation takes nearly as long as the original work, the net savings approach zero. Add the cost of the software itself and the implementation effort, and the pilot ends up demonstrating that the old process was cheaper.

Budget pressure makes this worse. JLL found that 65% of organizations have experienced CRE tech budget constraints over the past two years, and that ROI expectations have tightened, requiring more extensive business case development before approval. A pilot that cannot show clear savings does not get funded for a second phase.

Confidence Scoring Turns Verification Into an Exception Queue

One practical way to solve the verification problem is field-level confidence scoring, paired with validation rules and cross-checks. Running multiple models against the same document set and flagging where they diverge is one input into that score, not the whole architecture.

When two independently trained systems produce the same value, confidence in that value is considerably higher than when either produces it alone. When they disagree, that disagreement becomes the review queue. Instead of examining everything, reviewers examine the specific fields where the models could not reach the same conclusion.

Over time, that process generates something more valuable than the individual corrections.

“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.”

The sample size accumulated through parallel processing allows an organization to calibrate its thresholds empirically rather than by guesswork. If values scoring above a certain level have proven accurate across thousands of documents, those values can pass through without review. The verification burden shrinks as the evidence base grows, which is what turns a pilot’s economics from marginal into compelling.

Generic Models Do Not Understand Real Estate Documents

The choice of model also affects how often the verification problem arises in the first place.

“Most of the models are generic, they are more likely they will hallucinate at some point,” Joshi said. “We have models specific for real estate documents to fine tune the model so it can really understand the nuances.”

Subdivision is a useful illustration. A property record may present what looks like a single property, and a generic model reading it will extract it as one. A model trained on real estate documents should recognize from context that the property has been divided into multiple parcels, each with its own legal description, its own tax identifier, and potentially its own ownership. Missing that distinction produces a record that is confidently wrong in a way that is difficult to catch downstream, because nothing about the extracted output signals that anything was missed.

Production systems need more than a capable model. They need domain context, real-estate-specific evaluation, validation rules, and training against actual edge cases like this one, built from the document types, terminology, and structural conventions real estate records actually use.

Production AI Requires an Operating Model, Not Just Software

The last failure pattern is organizational rather than technical. Companies treat an AI deployment as a product 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.”

The comparison to SaaS is instructive. Enterprise software purchases come with implementation support, configuration assistance, training, and ongoing account management, because the industry learned that shipping a login and walking away produces low adoption and high churn. AI deployments carry all of those requirements and add several more. Models need monitoring. Confidence thresholds need recalibration as document mixes change. Exception handling requires people who understand both the technology and the underlying real estate context.

JLL found that only 33% of the workforce feels adequately trained on AI, and that 70% of occupiers use multiple sourcing strategies including external partnerships to fill capability gaps. Organizations that buy the technology without the supporting service end up staffing those functions internally, usually without having budgeted for them.

What the Successful Pilots Have in Common

The 5% that reach their goals are not working with fundamentally better models. They are working with better preparation.

They have cleaned and structured the data before the pilot begins rather than expecting the AI to compensate for what is missing. They have built verification into the workflow through confidence scoring rather than bolting it on afterward. They have selected models suited to the documents they actually process. And they have treated the vendor relationship as ongoing rather than transactional, with support that extends past implementation into operation.

Before scaling a real estate AI pilot, leaders should be able to answer five questions. Can we measure the baseline? Can we trust the output? Can we identify exceptions efficiently? Does the economics still work after human review? Can the solution operate reliably within the production workflow?

None of that is glamorous, and none of it appears in a product demonstration. It is the difference between a pilot that produces a presentation and one that produces a process the organization keeps using after the pilot period ends. The expectations the real estate industry has placed on AI are not unreasonable. Meeting them mostly depends on work that happens before the technology ever gets switched on.

The post Why Over 90% of Real Estate AI Pilots Fall Short of Their Goals appeared first on Propmodo.

​