Why the industry’s AI ambitions will depend on stronger asset identity, better data foundations and a clearer view of return on investment, and a much clearer view of what AI is quietly costing.
Dan Holland | Commercial & Operations Director, Nimbus
Following PropTech Connect 2026, Nimbus Commercial & Operations Director Dan Holland reflects on one of the questions sitting beneath the property industry’s growing AI ambitions: how useful can increasingly sophisticated AI really be if the property data underneath it is fragmented?
From asset identity and AI economics to the Golden Record and Nimbus ID, Dan looks at why the foundations behind property AI matter just as much as the models themselves.
“AI is an amplifier for the underlying data. Before making the amplifier more powerful, the industry must be able to trust the signal it is feeding into it.”
I spent last week at PropTech Connect, and one thing was impossible to miss: there is an extraordinary amount of technology being built for the property industry.
Some of it is genuinely exciting.
AI assistants, agentic workflows, automated research, data enrichment, document analysis and increasingly sophisticated models capable of carrying out tasks that, until recently, required significant human input.
But I think the commercial property industry is now entering a more important phase of the AI conversation.
The question is no longer simply: “What can we build with AI?”
It is becoming: “Which technology genuinely solves a problem, improves an operational workflow or creates measurable commercial value?”
And underneath that sits an even more fundamental question: is the data foundation good enough for AI to work properly in the first place?
One expression I heard there has stuck with me: AI is an amplifier for the underlying data.
I think that is particularly relevant to commercial property.
AI can interrogate information faster, connect datasets, identify patterns and automate increasingly complex workflows. Agentic models are beginning to move beyond answering questions and start completing tasks on our behalf.
But none of that fixes a weak underlying data structure. In fact, it can make the problem worse.
If an organisation has duplicated assets, inconsistent addresses, conflicting property records, incomplete information or multiple systems referring to the same building differently, putting an AI layer over the top does not magically resolve those problems.
It allows you to process them faster.
AI can therefore amplify good data, but it can also amplify inaccuracies, inconsistencies and structural weaknesses at enormous speed. That distinction matters.
Property data is particularly complicated because there is rarely a single, universally adopted identifier sitting underneath every dataset and workflow.
The same commercial property can appear differently across ownership records, planning data, EPCs, lease information, transactions, internal CRM systems, valuation records and third-party datasets.
One system may identify an asset by address. Another by title. Another by UPRN. Another through an organisation’s own internal property reference.
The residential industry’s instinct is to resolve all of this back to legal title, and for good reason. Title is the one record with a defined legal basis, and most of the property data tools in the market are built on exactly that assumption.
But title alone does not solve it. Titles frequently carry no address at all. Some land remains unregistered, despite land registration having been with us for so long. And in commercial real estate, the transaction very often does not happen at title level.
A ground floor retail unit is let separately from the office block above it. A multi-let industrial estate is let unit by unit. One title can carry several buildings, and one building can contain many lettable demises, each with its own occupier, lease event, rent and EPC. Assuming that every transaction happens on a whole title is a residential model applied to a commercial market where it does not hold.
Humans are remarkably good at looking at several slightly different records and understanding that they relate to the same building. Machines need considerably more certainty.
This becomes particularly important as the industry moves towards agentic technology.
Imagine asking an AI agent to assess an asset, identify ownership, review comparable evidence, interrogate planning history, analyse environmental constraints, understand lease events and combine that information with an organisation’s proprietary data.
The intelligence of the model is only part of the equation. Before it can reason across those datasets, it needs to know with confidence that every piece of information relates to the correct asset.
Otherwise, we risk building incredibly sophisticated technology on surprisingly fragile foundations.
There is another part of the AI conversation that I do not think receives enough attention: what does all of this actually cost?
AI can look deceptively inexpensive during experimentation. A proof of concept works. A small group of users tests it. The results are impressive. The organisation then starts thinking about deployment.
But production AI economics are more complicated.
There is the cost of the models themselves, API and inference usage, data infrastructure, engineering resource, integrations, orchestration, monitoring, security, governance and the ongoing maintenance required as models and workflows evolve.
Agentic models add another dimension. An agent completing one business task may make multiple model calls, search several systems, retrieve documents, interrogate databases and call external services before returning an answer or completing an action.
Multiply that across hundreds or thousands of users and workflows and the economics begin to matter.
There is also a genuine view across the AI market that the price of the models themselves will need to rise rather than fall, as demand for compute continues to outstrip supply. An organisation budgeting on today’s prices, at today’s usage, may therefore be underestimating the position twice over.
Not simply: “Can AI do this?” But:
What does it cost each time AI does this?
What human activity is it replacing or improving?
How much time is actually being saved?
Is the output more accurate and reliable?
Is the workflow measurably better?
And ultimately, what return are we generating against the total cost of deploying it?
The winners in property AI will not necessarily be the organisations deploying the greatest number of models. They may be the organisations that can identify the workflows where AI creates the greatest measurable return.
There is an important connection between these two issues: poor data does not just affect AI accuracy. It affects AI economics.
If an agent has to repeatedly search, reconcile, deduplicate and interpret conflicting records before it can complete a task, every workflow becomes less efficient.
More processing. More model calls. More engineering. More exceptions. More human intervention. More uncertainty. And ultimately, more cost.
A strong data foundation, therefore, is not simply a data-quality project. It is part of the commercial architecture of AI.
The cleaner and more structured the information being supplied to an agent, the more efficiently that agent operates.
“Poor data does not just affect AI accuracy. It affects AI economics.”
--
It is worth walking through what this looks like in practice, because without a golden record the job we are handing to the model becomes very much larger than most people assume.
Ask an AI model to draft a report on a commercial property. Before it writes a word, it has to understand the property. So it goes looking for data it may or may not find. It may find something, but on an unreliable web page, and then rely on it. It may fail to find it at all, because the information sits in a plan it has not been given. Or it has been given the plan, but the information is held in a map rather than a neat list, or under a naming convention that means nothing outside the organisation that created it.
So the model first has to work out what it needs to know, then find it from a source it can trust, then link it accurately to the right asset. That is three problems solved before the report has started.
Then the report needs context. Most reports set the building within the market it sits in and comment on its use, its planning potential and the planning policy that might apply to it, and what has transacted nearby and on what terms. Every one of those judgements bakes back to an accurate understanding of an asset, repeated across every comparable property, every transaction and every planning permission in that market.
The problem multiplies out. At that scale it becomes too large even for the most powerful models, which can then latch onto the wrong information, hit the limit of their context window, or answer confidently on an incomplete picture. Agentic architectures help with capacity, but even the best of them cannot retrieve what sits behind a gated API or inside a paid-for data file.
Which means the true number of tokens needed to properly answer the questions the property industry actually asks is far higher than the market understands, and the risk of an incorrect answer is very significant.
And it is happening quietly. Models are running down organisations’ credits in the background, and those credits are in effect cold, hard cash.
Very few organisations have connected the two. They are paying real money for compute that is being spent trying to work out which building is being discussed, and they usually discover it either when the bill arrives or when the answer turns out to be wrong.
There is a name for what is missing here - the industry calls it the Golden Record. The term came out of conversations with the largest agency businesses in the market, including CBRE, Knight Frank and Savills, and it has been used in other industries for years to describe the single trusted version of an entity.
The concept is that every asset has a trusted identity that becomes the anchor point for the information associated with it. That becomes increasingly powerful when organisations begin connecting their own data.
Where property differs from other industries is in how far down that record has to go. The market generally thinks about a golden record of ownership: resolve everything back to legal title and stop there — with all the gaps that leaves, and a quiet assumption that assets are bought, sold and let whole.
At Nimbus we resolve to title, and then go two stages further. We reconcile to the buildings that sit on that title, and then to the individual lettable demises within those buildings.
That is the level at which commercial property is actually let, valued, rated and transacted, and therefore the level at which an AI model has to operate if the answer is going to be right.
This thinking sits behind something we have been building at Nimbus for some time: Nimbus ID.
Our ambition was straightforward, even if the underlying challenge was not: to create a single source of truth for every UK asset.
That is not a roadmap item. It is built, and it is available today, either through our API or connected directly into an LLM.
Nimbus ID provides a persistent property and unit identity around which different pieces of information can be structured.
Rather than asking an AI model to determine repeatedly whether multiple records refer to the same property, the objective was to establish the asset first and connect the relevant information around it.
That creates a fundamentally different foundation for AI and agentic workflows.
Ownership data, planning information, transactions, leases, environmental information, proprietary datasets and an organisation’s own internal information are all connected back to the same underlying asset, or to the individual demise within it.
Coverage and match quality are not uniform across every dataset, and we document both openly rather than claim otherwise. But the principle holds: establish the asset first, then connect the information to it.
For our own agentic models, that foundation is critical. We do not want our AI simply to be capable of finding more information. We want it to understand which property that information belongs to.
A property company may have decades of internal information: inspections, valuations, transactions, leases, contacts, photographs, reports and institutional knowledge. The opportunity presented by AI is enormous.
But connecting an AI model to that information without first resolving the underlying asset structure risks creating another layer of complexity.
If instead that proprietary information can be linked to a persistent property identity, the organisation starts creating something much more valuable: a structured data ecosystem that both people and machines can understand.
That is where I believe some of the most interesting applications of agentic AI in commercial property will emerge.
There is enormous reason to be optimistic about AI in commercial property.
The technology is moving incredibly quickly, and I think we are going to see workflows transformed across agency, valuation, investment, development, asset management, lending and research.
But there is a danger that the industry becomes so focused on the intelligence sitting at the top of the technology stack that we overlook the foundations underneath it.
The most sophisticated AI model in the world cannot compensate indefinitely for fragmented, duplicated or poorly structured property data. Eventually, the quality of the foundation determines the quality of what can be built on top of it.
So perhaps the next phase of property AI needs to involve fewer conversations about models alone and more conversations about identity, structure, provenance, interoperability and economics.
Because AI is an amplifier. And before we focus entirely on making the amplifier more powerful, we should make sure we can trust the signal we are feeding into it.
For commercial property, I believe that starts with something deceptively simple: knowing exactly which asset we are talking about.
Dan Holland, Commercial & Operations Director, Nimbus
See how Nimbus connects trusted property data, asset identity and AI-ready workflows.