Brokerages Are Racing To Adopt AI. Costs And Headaches Are On The Rise

Artificial intelligence is the No. 1 buzzword in business, and it's no different in commercial real estate, where transaction specialists are being pushed to reinvent how they work.

As firms race to weave AI into their operations, some have integrated new software into the day-to-day work of thousands of employees. But many have found that developing tools and teaching a relationship-driven industry to use highly advanced technology have delivered a disappointing return on investment so far.

“Having to build it with costs going up, when you have to factor in ongoing support, it does change the calculus,” Dealpath Managing Director Gary Kao said. “A big area of focus for a lot of firms now is not only thinking about how to deploy AI, but you need to think about how to be efficient with AI.”

The proportion of corporate real estate firms running AI pilots has exploded from 5% to 92% over the past three years, according to a JLL survey of more than 1,000 industry professionals. Budgets have been reallocated to prioritize AI integration into workflows, with the top categories being strategic advisory, increasing cyber and data security, and upgrading digital infrastructure. 

Yet just 5% of respondents said they have achieved most of their program goals. At the same time, the gaps between the “cans” and “cannots” are widening as technology becomes increasingly advanced and expensive. 

In its 2025 annual report, CBRE reported spending nearly $1.7B on computer hardware and software, up $300M from the year before. But on the firm's fourth-quarter earnings call, CBRE CEO Robert Sulentic said AI integration is projected to drive a 25% reduction in research costs.

In the first quarter, Cushman & Wakefield reported spending roughly $337M in operating and administrative expenses — a $30.9M, or 10%, year-over-year increase that was attributed to higher salaries and technology costs. 

The company has released a variety of tools to its brokers in waves, from Microsoft Copilot to Jasper and Unframe AI. Employees can request more advanced access, according to a senior broker who requested anonymity to speak candidly about his experience.

The brokerage declined to make anyone available for an interview for this story.

“Cushman kept on pushing out announcements, like, ‘Make sure you don't use unsanctioned AI,’ and I think everybody and their cousin was using it at this point,” the broker said. “I know several other people at other firms and in our office who were like, ‘I'm using Claude. It's too good not to use.’”

Companies from across the industry have reported struggling with “shadow AI,” a term for employees using AI without official approval and parameters.

Cushman & Wakefield has since bought an enterprise Claude account, which the broker estimated approximately 600 people have access to. A Slack channel has been established in which brokers share agents that they have created and other tips.

Claude can be fed data to put together comps or offering memos, which can then be manually improved. In cases where brokers aren’t tied to nondisclosure agreements, AI can analyze and summarize development models and other spreadsheets. 

“Now I have enough data that in the middle of the day I can call my client back,” the broker said. “I used to have to do an hour or two of work to where I felt like I was coherent enough to have a meaningful conversation with the client.”

Some colleagues use AI to handle more mundane tasks, like creating client lists or summarizing and writing emails — something that the broker fears may be subject to the “dangers of hallucinations.” 

“I'm much more confident because our enterprise version is sandboxed,” the broker said. “Our clients give us information in confidence, and we would never mishandle it intentionally, but you worry about how the AI might.”

Those concerns aren’t unfounded.

Although two-thirds of companies allow employees to independently develop or deploy AI agents, only 60% provide formal, organizationwide policies and frameworks about use, according to a survey by EY. Half said they don’t have a high level of visibility into employee use of AI agents.

Virtually all — 99% — of the 975 C-suite executives surveyed by EY reported financial losses resulting from inadequate controls. Nearly two-thirds suffered losses of more than $1M. On average, companies that have experienced risks have lost a conservative estimate of $4.4M.

Part of the issue is a lack of fundamental infrastructure and talent. Roughly half of CRE firms say workforce gaps, more stringent return-on-investment expectations and budget constraints are making technology procurement more difficult than three years ago, according to JLL. 

But driving ROI can be a bit of a paradox, according to Grotto AI CEO Nick Deveau, who has spent nearly a decade building AI products. Getting employees on board requires demonstrating value. Doing so often requires employees using the tools. 

“If it's only a top-down mandate, it's not going to work,” Deveau said. 

He added that large organizations tend to be better at encouraging employees to use new technology than smaller companies. 

“You see it in the way they buy software. The firms that want to innovate know that they can't just do it themselves. They have to procure software, they run RFPs, they have a process,” Deveau said. “In the smaller orgs that are [not] truly early adopters, it can be really, really slow-moving.”

JLL has invested more than $445M in nearly 60 startups since 2017 through its JLL Spark Global Ventures arm, many of which are AI-native. Those technologies are then scaled across its network, a practice that has made adoption easier, executives said.

“We have an in-the-trenches view over what is being developed, what is successful, what is being disruptive,” JLL Spark growth principal Daniel Correa said. “To be able to have a stake in things that eventually start to change the industry is really valuable.”

Within the brokerage, groups of beta testers develop and try their own tools and agents before they are rolled out to larger teams. JLL Global Chief Data Officer Rick Ferrino calls this a “citizen-driven” approach. 

About a quarter of JLL's employees use the company’s internal AI platforms daily, and users estimate saving an average of two hours per week as a result. The brokerage’s proprietary generative AI model has more than 100,000 total users and has processed more than 35 million prompts.

“We're in this interesting inflection point where people are pretty confident that there is a clear ROI, but they can't quite prove it. They see teams are moving way faster, but also expenses are going up,” Ferrino said of the industry. “Is the juice worth the squeeze? I hear that analogy way too often, and ultimately the answer is yes.”

Falling behind means that brokerages may need to do more to catch up, especially as costs rise. Kao said he has seen more brokerages opting to outsource AI needs rather than build bespoke tools in-house.

“Build versus buy is very clean-cut when tokens cost nothing,” Kao said. 

Software-makers like OpenAI and Anthropic have started using token models to charge users instead of flat-rate subscriptions. The more complex the task, like writing code for a new feature, the greater the cost.

Even tech companies are running up against a wall. Meta and Amazon previously graded AI utilization on “tokenmaxxing,” putting up employee leaderboards to encourage use. Now, after blowing through budgets, Meta, Uber and Walmart have instructed workers to limit their token spending. 

“We had to quickly shift away from ‘token usage equals good’ to ‘token usage equals potential value,’” Ferrino said. “And value [is determined by] working with people to actually showcase what they've developed and why it's better than the status quo.”

The Cushman broker said that before receiving an enterprise account, he gave his team access to an account he personally paid for. One employee immediately overused it, running up a significant bill. 

Now, the firm has a token limit. Those who want to surpass it can make a business case to management.

Additionally, Cushman has created a prompt library that instructs users on how to word requests. Employees are also encouraged to use more basic AI models for simple searches so that they aren’t using a “blowtorch to light a cigarette,” the broker said. 

Walker & Dunlop Senior Managing Director Aaron Appel said the Maryland-based firm has spent “substantial money” on enterprise accounts while building out its own proprietary systems. At the same time, executives have tested those tools to ensure that they are being used efficiently. 

In one case, two analysts were put head-to-head to identify every industrial owner within a certain size and geography without an announced institutional capital partner. One did the task manually, using applications like CoStar, Reonomy and Real Capital Analytics. The other used Claude. 

Claude completed the task within 10 minutes, while the manual analyst took about a week. The only issue: The AI chatbot missed several results. 

“It is a tool,” Walker & Dunlop Senior Managing Director Mo Beler said. “It is not an entire Swiss Army knife at the moment.”