Why the next challenge for real estate AI is keeping digital information connected to physical reality

A property company can now ask an artificial intelligence system a question that would once have required days of analysis:

Which buildings in our portfolio have the greatest opportunity for space optimisation?

Within seconds, the system can compare floor areas, occupancy levels, leases, operating costs and asset information across dozens or hundreds of properties. It can identify patterns that would be difficult for a human team to see and present its conclusions with extraordinary clarity. But there is a question that should come before all of this.

What if the floor areas behind the analysis are wrong?

Or, more subtly, what if they were perfectly accurate when they were produced five years ago, but the buildings have changed since?

This distinction matters because commercial real estate is moving rapidly towards artificial intelligence. JLL’s 2025 Global Real Estate Technology Survey, covering more than 1,500 senior real estate decision-makers, found that 88% of investors, owners and landlords were already piloting AI, as were 92% of occupiers. Yet only a small minority had achieved all of their AI programme objectives.

The industry is understandably focused on models, platforms and use cases. But the quality of AI reasoning is only one part of the equation. The other is the quality of the reality being presented to it. For real estate, that creates a particularly difficult problem.

Buildings change. Data ages. AI scales whatever sits between them.

The challenge is therefore not simply to make property data available to AI. It is to maintain a trustworthy relationship between a changing physical asset and the digital information used to represent it.

I refer to this as Reality Fidelity: the degree to which digital information about a building can still be trusted to represent its current physical state, with sufficient understanding of where that information came from, why it was created and what may have changed since.

A building can be digitised and still be poorly understood

Data quality is already recognised as a major constraint on AI adoption. In its work on AI-ready real estate data, JLL highlights completeness, correctness and consistency as fundamental requirements. Is the information available? Is it accurate? Can it be interpreted consistently across systems?

Those conditions are essential. But buildings introduce another dimension: time.

A building may begin its operational life with excellent information. Its drawings are current. Its BIM model is coordinated. Its areas have been measured. Its equipment has been commissioned and entered into an asset register.

Then the building begins to live. A tenant moves a partition. Two offices become one. A retail unit is divided. An HVAC system is replaced. A refurbishment changes circulation. New services are installed. A leasing consultant produces an updated plan. A facility manager records a maintenance intervention.

None of these events is exceptional. They are normal consequences of operating property. The problem is that physical change and information change do not always happen together.

Over time, different departments can end up holding different versions of the same asset. Leasing has one floor plan. FM has another. A BIM model reflects the original construction. A newer PDF captures a refurbishment. An asset register contains equipment that no longer exists. An area schedule remains commercially active even though the physical configuration has changed.

The organisation may therefore have more data than ever before while becoming less certain about which information should be trusted. Research into BIM for facilities management demonstrates how real this problem is. A 2023 study published in Frontiers in Built Environment examined three projects and analysed 2,177 documented information-compliance issues across 111 Revit models. The researchers found recurring problems involving incomplete, inaccurate, inconsistent and unintelligible information, including failures to adequately capture deviations from design during construction.

This leads to an important conclusion:

Digitalisation is not the same as truth.

A digital record can be perfectly structured and still describe a building that no longer exists in that form.

Good data can become bad without anyone making a mistake

Imagine a measured drawing produced professionally five years ago. At the time of survey, everything was correct. The geometry was verified, the methodology was appropriate and the quality control was rigorous.

Two years later, the floor was refurbished and several partitions moved. Nothing has happened to the original drawing. The file has not been corrupted. The surveyor did not make a mistake. Yet the drawing is no longer a reliable representation of the current asset. The data has become temporally unreliable.

This is why the usual dimensions of data quality are not quite sufficient for the built environment. We need to ask not only whether information is complete, correct and consistent, but whether it is current enough for the decision being made.

Area measurement illustrates the problem particularly well. The same property can legitimately have several different area figures because the numbers were created for different purposes, under different standards or at different points in the asset lifecycle. One may support planning, another leasing, another valuation and another property registration.

An AI system may encounter three different values for the same property and correctly identify a conflict. It may examine associated documents and infer which one is most likely to apply.

But inference and verification are not the same thing. A number becomes far more valuable when the system also understands when it was measured, by whom, according to which methodology, for what purpose and what changes have occurred since.

In other words, provenance matters as much as the value itself.

The question is no longer simply, “Do we have the area?”

It becomes: “How confident are we that this area is appropriate for this decision today?”

That is where Reality Fidelity becomes useful.

AI will help repair data, but it cannot eliminate the need for evidence

There is an obvious counterargument. If AI is becoming so powerful, why should it not solve these problems itself?

To an extent, it will. AI can already extract information from drawings and documents, classify records, detect anomalies, compare datasets, identify missing fields and make previously inaccessible information searchable. Multimodal systems are increasingly capable of reasoning across images, text, geometry and structured databases. Future systems will become substantially better at reconciling conflicting information. But intelligence cannot remove the distinction between what is inferred and what has been observed.

Suppose an AI discovers that an FM drawing shows six rooms while a newer leasing plan shows five. It finds an invoice for partition works completed two years ago, but there is no updated verified survey. The system can form a strong hypothesis about which configuration probably exists. What it should not do is silently convert that probability into certainty.

A trustworthy AI system should instead be capable of saying:

Based on the available evidence, the layout may have changed since the last verified survey. Confidence in the current geometry is limited and targeted verification is recommended. That is not a weakness. It is a better form of intelligence. The most valuable AI systems in property may eventually be those that understand not only what they know, but why they believe it and where uncertainty remains.

From information debt to Reality Fidelity

There is a useful analogy with software engineering.

Technology companies speak about technical debt: shortcuts and unresolved issues that accumulate over time until systems become expensive and difficult to maintain. Buildings accumulate something similar. We might call it information debt. Every physical modification that is not properly reflected in the digital record adds a small amount of that debt. A moved partition. A replaced chiller. An undocumented fit-out. An area recalculation living in one spreadsheet. A new drawing that never reaches the central system.

Each discrepancy may initially appear insignificant.

Over years, however, they compound. Eventually an organisation reaches a familiar situation: people stop trusting the central information. An architect requests another survey. The facility manager phones the person who “knows the site”. Leasing uses its own plans. Project teams visit the building to measure again. Consultants recreate information that supposedly already exists.

The organisation has not necessarily lost its data. It has lost confidence in it.

Work from organisations including Fraunhofer, CSTB and A3D illustrates the same broader principle: digital building information has to be maintained as the physical asset changes if it is to remain useful for operational and decision-making purposes. The answer is not continuous surveying.

If buildings continually change, must owners continually resurvey them?

No.

A better approach is to shift from continuous capture to continuous confidence assessment. Consider two assets. One building was verified nine months ago. No refurbishment has occurred. No significant equipment changes have been recorded. The current drawings agree with leasing and FM information.

Its Reality Fidelity is likely to remain high.

Another property was last fully verified four years ago. Since then it has undergone two fit-outs, a major HVAC replacement and several leasing changes. New drawings have appeared in different systems and some contradict one another.

That does not automatically mean its existing information is wrong. It means the organisation has a rational reason to place less confidence in it. The principle is straightforward: establish a trusted baseline, preserve the provenance of important information, monitor events that could alter the asset and use those events to determine where confidence in the digital representation may have degraded.

When uncertainty becomes material, verify the affected part of the building rather than resurveying everything.

A major HVAC replacement might require an update to asset information but not to the building envelope. A tenant fit-out could affect internal geometry without invalidating external elevations. A subdivision could require new area verification and leasing plans without requiring an entirely new building survey. The objective is not a perfect digital model frozen in time. It is a model whose degree of trust is understood and actively maintained.

This is what a digital twin should ultimately become

This perspective also clarifies an increasingly blurred term: the digital twin.

RICS emphasises that synchronisation between a virtual representation and its real-world counterpart is a defining characteristic of a digital twin. Accurate and reliable information needs to remain connected to the physical asset at an appropriate level of fidelity and frequency across its lifecycle.

That distinction is important. A BIM model can represent what was designed. An as-built survey can represent what existed when it was captured. A digital twin should represent a maintained relationship between the two worlds.

If that relationship stops being maintained, the digital representation may still be technically excellent. But gradually it becomes a historical model rather than a trusted operational twin. This is particularly relevant to existing portfolios. Much of the discussion around AI and digital twins implicitly assumes that organisations begin with clean BIM environments and well-structured databases. Real estates rarely do.

Existing buildings may begin with twenty years of PDFs, DWGs, spreadsheets, survey reports, photographs, CAFM records, maintenance systems and knowledge stored in people’s heads. The challenge is therefore not to wait for perfect information before using AI. It is to understand the evidential quality of the information that already exists and progressively improve it.

The next generation of property AI should be reality-aware

Today, most AI interfaces are designed to answer questions. Tomorrow’s property intelligence systems should also be capable of qualifying their answers.

Imagine asking: Which property in this portfolio offers the greatest potential for space consolidation?

A conventional system might answer: Building 12.

A more reality-aware system might answer:

Building 12 appears to offer the greatest potential based on current occupancy and recorded floor area. However, its spatial information predates a major refurbishment completed three years ago. Confidence in the comparison is therefore lower than for Buildings 7 and 9. Verification of the affected floors is recommended before a consolidation decision. The second answer is less absolute, but more useful because it makes the evidence and its limitations visible rather than presenting an uncertain conclusion as fact.

This is the difference between an AI system that simply processes building data and one that understands the evidence behind the data. A reality-aware system would ideally know where important information originated, when it was verified, whether relevant changes have occurred since and how much confidence should therefore be attached to the conclusion. It could identify information debt, highlight conflicting records and determine where uncertainty matters commercially. Eventually, it could also help prioritise the inspections, measurements or captures required to restore confidence. AI would no longer sit on top of the digital building record merely as an analytical layer. It would participate in maintaining the relationship between information and reality.

The competitive advantage may be beneath the AI

There is also a strategic implication for real estate companies. AI capability is becoming more accessible. Models will continue to improve, and many organisations will eventually have access to similar analytical and reasoning capabilities. Imagine two property owners using equally capable AI. The first runs it over fragmented plans, uncertain areas and incomplete asset information. The second runs it over verified spatial data with known provenance, recorded changes and a process for identifying where confidence has degraded. The underlying AI capability may eventually be comparable, but the quality of the decisions will depend heavily on the information and evidence available to it.

The long-term competitive advantage may therefore come less from owning a particular AI model and more from owning a trustworthy, proprietary understanding of the physical portfolio. AI may become increasingly commoditised. Verified knowledge of your own assets will not. This is also the principle shaping how we think about Spatial-IQ at Digital Reality Corp. The objective should not simply be to place a conversational interface over property files. The more meaningful opportunity is to make verified spatial information interrogable while preserving the relationship between an answer, the evidence behind it and the physical building itself.

A building can then move from being a collection of drawings, models, spreadsheets and reports towards becoming a living information asset that can be questioned, compared and continuously improved. That is when spatial intelligence becomes genuinely useful.

Trust is the real AI infrastructure

The real estate industry should continue investing in AI. But AI readiness should not begin with the question, “Which model should we deploy?”

It should begin further down the stack: understanding where critical building information came from, when it was verified, the purpose for which it was created and what has physically changed since. Most importantly, organisations need to understand where uncertainty has become significant enough to require new evidence.

That is the essence of Reality Fidelity.

The solution is not to endlessly capture every building, nor to assume that AI can reconstruct decades of physical history from incomplete records. It is to create a disciplined loop between reality, evidence, intelligence and change. Establish ground truth when it matters. Preserve provenance. Monitor change. Assess confidence. Let AI expose uncertainty. Verify selectively. Update the record.

Then repeat.

The next generation of real estate AI should therefore know more than what the data says. It should know why the data deserves to be trusted. Because the most important question may no longer be whether AI understands our building data. It is whether our building data still understands the building.

By Strategic Director at DRC, Fouad Zouak

Selected References:
JLL. Global Real Estate Technology Survey 2025 and Myth: AI can overcome low quality data.
Tsay, G.S., Staub-French, S., Poirier, E., Zadeh, P. & Pottinger, R. BIM for FM: understanding information quality issues in terms of compliance with owner’s Building Information Modeling Requirements. Frontiers in Built Environment, 2023.
RICS / Glodon. Digital Twins from Design to Handover of Constructed Assets.
Fraunhofer Institute for Building Physics. Digitaler Gebäudezwilling: Intelligente Verknüpfung von Messdaten und Simulation.
CSTB. C Data: Données unifiées, bâtiment maîtrisé and research on BIM, digital twins and building-data management.
A3D. Spanish case studies on as-built BIM and continuous digital model maintenance for existing facilities.