From Pilot to Practice

Across the real estate industry, a familiar pattern is emerging. An investment analyst starts using a GenAI tool to summarise long market reports. The output is useful. She shares the trick with a colleague, who adapts it to flag due diligence risks. Six months later, both are still using their tools, but the tools remain personal habits rather than team practice. There is no agreed set of use cases, no common way to check the output and no wider application across the team.

That is where many real estate organisations are today. Early use cases and pilots have been built and individual users see real value. What has not yet happened is the move from individual experimentation to repeatable team workflows: from a useful shortcut to a process that runs consistently and from a few adopters to a capability that lives inside the organisation.

This is not a failure. It is a normal pattern in the early stages of major technology adoption. The more interesting question is what the teams moving beyond the pilot phase are doing differently.

GenAI adoption in real estate is now widespread among large investors, owners, managers and corporate occupiers. Many organisations are testing several use cases at the same time. Yet overall maturity remains low. The sector has moved quickly into experimentation, but more slowly into scaled operating practice.

The pattern we see most often is consistent: the intention to use AI is nearly universal, but only a small share of teams describe themselves as actively scaling it. Others are still relying on general tools such as ChatGPT, Claude or Copilot for routine tasks, usually without a structured link to workflow, data, governance or quality control.

The next barrier is rarely the tool itself. The recurring constraints are data quality and access, workflow design, governance and skills. These are the hard parts of adoption. They are also where the opportunity sits, because they are the parts an organisation can shape directly.

The organisations making the most progress have stopped asking only, "Where can we use AI?" They are asking, "What should this workflow look like in three to six months and where should AI sit inside it?"

Adding AI to an existing process can create individual productivity gains, which are useful but limited in scope. Redesigning the process around what AI systems can now do reliably tends to open up something larger. Some steps disappear. Some are reallocated. New steps, such as systematic verification and review of AI-generated outputs, become part of the workflow rather than an informal check at the end.

This distinction matters because review and verification need their own time budget. AI can shorten the production of a draft document, initial calculation, investment memo section or market summary. But verifying that output against source documents, underlying data and committee-grade standards still takes real work. Workflows that treat this as a small final overhead often fail under scrutiny. The time saved can be significant, but it is usually less simple than headline productivity figures suggest.

We saw this play out with a European core-fund manager. Three analysts were using AI to draft sections of investment memos, each in a different way. Output quality was uneven and the investment committee had started to push back. Rather than adopting a new tool, the team rebuilt the memo template around three explicit AI-supported sections. Each section had defined inputs, a defined check and a defined sign-off. Production time fell by about a third and confidence in the process improved. The redesign sat entirely in the workflow, with no change to the underlying tools.

This kind of redesign is harder in real estate than in some other sectors. Transactions are bespoke and processes are often customised by deal. But that is an argument for doing the work, not against it. Teams that map their core workflows, identify the elements that repeat and design AI into those elements deliberately are more likely to build something durable. Individuals who simply apply new tools to old processes may gain speed, but they rarely create scale.

Many real estate teams already have a few people who are capable with AI tools. They know which prompts work, which platforms suit which task and where the limits are. The rest of the team uses the tools occasionally, often without much structure.

That asymmetry is manageable in a pilot phase. It becomes a problem when the goal is repeatability. Repeatable workflows require shared standards and shared standards require shared skill. When only a few people know what good practice looks like, others either avoid the tools or use them in ways that create false confidence.

One mid-sized asset management team illustrates the point. A single analyst was producing market briefs faster and better than anyone else. Over six months, he had built a personal toolkit across three AI platforms. Nobody else knew how to reproduce it. When he took two weeks of leave, the team produced nothing comparable. The lesson was clear: the firm had relied on one capable individual rather than building capability across the team. The skill needed to move out of his head and into the team.

Internal expertise is now one of the most common barriers to adoption. Budgets for AI tools are being approved, but skill development is not always keeping pace.

The organisations that scale invest in raising the whole team to a practical working level. That does not mean turning every analyst into a prompt engineer. It means ensuring that everyone can use AI responsibly within their part of the workflow, understand when output needs to be checked and apply a common quality standard to what they produce.

Capability-building also needs to be grounded in real estate work. The relevant test is whether a memo, sensitivity analysis or market research brief still stands up when an investment committee asks detailed questions, not whether someone can write a clever prompt.

A predictable feature of GenAI tools is that output can look more confident than it deserves to be. The narrative may read well. The reasoning may appear internally consistent. The numbers may look plausible. And occasionally, a key fact will still be wrong.

This has organisational consequences. Investment committees should not trust analysis that the team cannot defend. Committees should be cautious about placing weight on recommendations they cannot trace back to source documents and underlying data.

In the pilot phase, individual users often develop their own checking habits. They return to source documents, test comparables and flag anything that looks too clean. These habits can work when the user is also the reviewer. In a scaled workflow, the user and reviewer are often different people. Output moves from one team member to another, from one team to another and sometimes to an external committee. Without explicit checking routines, errors travel further than they should.

The teams that scale treat verification as a designed part of the workflow, not a personal habit. They define which outputs require which checks. They train the team to test the work. They document common failure modes for each use case, so analysts know what to look for. This can feel bureaucratic, but it is what makes the tools usable at scale. It gives the organisation, not just the individual user, a basis for trusting the output.

Data protection, confidentiality and regulatory compliance are usually raised early in any AI conversation. Too often, they are then parked with legal or another control function and revisited only when a pilot runs into a problem.

A more effective approach is emerging in organisations that scale well: governance is treated as part of workflow design from the start. Which data can go into which tool? What should be anonymised and at what stage? Which use cases are permitted with confidential information and which are not? Who decides when the rules need to change?

Organisations that move ahead put governance frameworks in place before they scale their use cases, not after. Late governance is one of the common reasons a promising pilot stalls before it reaches live operation, especially when the use case touches sensitive transactions or proprietary data.

These are not the most exciting questions, but they determine whether AI can be used with confidence across a team. Getting them right early reduces friction later. Leaving them ambiguous is one reason many pilots remain stuck.

Many real estate market participants are now working through the same adoption curve. The next step is not simply more experimentation. It is building the operating conditions that allow useful experimentation to become everyday practice.

There is also a quieter point that often gets lost in the productivity debate. When organisations move from pilot to practice, the work itself can become more interesting. Analysts spend less time on mechanical extraction and more time on judgment. Investment committees receive better prepared cases. Investment and asset managers can work with more data, more sources and more scenarios than was previously practical.

The productivity case is real, especially in a sector with large volumes of unstructured documents, repeated analytical patterns across transactions and decisions that benefit from systematic comparison. But the longer-term advantage will belong to organisations that combine productivity with workflow discipline, team capability, verification routines and governance.

For organisations still finding their footing, the most useful next step is rarely another pilot. It is to step back and look at the pilots already running.

Which use cases have the clearest workflow logic? Where are techniques being shared across the team and where are people still working in parallel? Where would a modest investment in shared training and structured verification turn an individual skill into a team capability?

An honest answer to those questions usually points to the next concrete step. The organisations acting on that answer now may not look dramatically ahead this quarter. But two years from now, they are likely to be in a different conversation altogether.