As the real estate industry rapidly adopts artificial intelligence, calculating agent adoption of AI-powered platforms and proptech funding rounds is typically the easiest part. By now, it’s a given that most real estate agents use AI in some form, and virtually all proptech companies incorporate AI technology into their products.
Proving these AI tools pay for themselves is the part that’s harder to figure out.
Technology consulting firm Accenture noted this problem across the business world in a recent report that gives the discipline a name: “tokenomics.”
The cryptocurrency industry initially coined the term “tokenomics” to describe coin supply and distribution. The term has since migrated to the AI field to describe the shift from fixed software subscriptions to consumption-based token billing.
In AI parlance, it means treating the token, the basic unit an AI model processes, as a factor of production that has to be measured against the value it returns, not just tracked as an expense.
Accenture’s research found that only 23 percent of C-suite leaders surveyed report widespread, sustained business value from AI across their organization, even as AI-related spending is on pace to top $800 billion in 2026. The tech consulting firm found that fewer than 10 percent of users and workflows typically drive the vast majority of that spend.
The problem isn’t confined to real estate or proptech. Corporate leaders across many industries have broadly started questioning whether soaring AI spending is actually delivering returns.
Companies that rushed to adopt AI are now facing steep bills and murkier productivity gains. In one extreme case, an AI consultant told Axios that a client burned through $500 million in a single month. It resulted from never capping how many Claude licenses employees could use.
The problem hasn’t gone away. As of this summer, nearly 7 in 10 U.S. companies said at least some of their AI initiatives ran over budget in the past year, and a third said overruns happened mostly or always, according to a report from AI security firm WitnessAI.
Real estate technology has its own version of the problem, though it may be less pronounced than in other industries where AI adoption has been faster and more aggressive.
Inman spoke with proptech founders and tech leaders about how the industry is coming to grips with AI spending and what it’s doing to keep costs under control.
Real estate has been slower than other industries to adopt AI, Chen said, largely because brokerage work is relationship-driven rather than computational.
Joe Chen
Joe Chen
That’s starting to change, he argued, as AI tools move beyond coding applications and into sales support and back-office operations, areas he said weren’t viable for AI even two years ago.
Chen said Lofty is betting that token costs will keep falling as AI infrastructure investment continues, making sales-oriented AI tools cheaper to deploy at scale.
That bet comes with a tension he acknowledged directly. Lofty’s own AI sales agent has drawn strong customer reviews, but token costs have, in some cases, made the product unprofitable to run, he said.
“I believe research and new startups in the AI chip space will dramatically lower token costs, and each token will also become more powerful, delivering more value per token,” Chen said.
Lower AI token costs are already happening. CNBC recently reported that a key gauge of daily AI model prices hit its lowest point yet. It’s more evidence that the cost of running AI is falling as fast as its capabilities are rising.
The LLM Token Expenditure Index, a benchmark from intelligence firm Silicon Data that tracks daily pricing across leading models, fell to 97 cents on August 31. That’s the index’s lowest reading since it launched late last year.
A falling index points to lower costs to do the same work on ChatGPT, Claude, Gemini or similar tools.
Part of the drop traces to competition from open-source Chinese models, including Moonshot’s Kimi K3, which are underpricing offerings from leading U.S. frontier labs, Charles-Henry Monchau, chief investment officer at Syz Group, recently wrote.
The LLM Token Expenditure Index | Source: Silicon Data, CNBC
The LLM Token Expenditure Index | Source: Silicon Data, CNBC
An ‘amnesia problem’
As a unit of measurement, Shayan Hamidi compares AI tokens to kilowatt-hours for electricity.
“That’s how a factory pays to run its machines,” said Hamidi, the Rechat CEO and founder. “And when energy got a currency, the kilowatt-hour, then we could measure it and optimize it, and that’s when manufacturing scaled.”
With AI, Hamidi said we have a digital worker instead of a machine, and tokens are the currency for the work they do.
Shayan Hamidi
Shayan Hamidi
“And here’s what’s interesting, and here’s the part that you don’t really hear a lot of people talking about in our space,” Hamidi told Inman. “But real estate is about to have a very expensive token problem.”
It’s not because AI is expensive, Hamidi said. It’s because in the real estate industry, the data lives in “18 different places.”
“So we don’t have a token problem yet. Today we have an amnesia problem,” Hamidi said. “And when your data sits in 18 places, the system remembers nothing by default, obviously.”
That means a system has to relearn pretty much an entire business every time you ask it to do something. And companies pay for that lookup every time, so the “amnesia problem” turns into a token problem.
“And I think a very important point here, because there are not a lot of CRMs promising they integrate with all of your other tools, but integration just moves data,” Hamidi continued. “What they don’t do is create memory.”
A CRM that connects to your other tools can look things up, Hamidi said, but it doesn’t actually know a business by default. And every request starts from zero, and it goes fetching.
“And that’s fine with a human sitting and clicking, but it’s very expensive when a digital worker like AI is doing hundreds of steps,” Hamidi said.
A shift in how proptech gets priced
As Lucy, Rechat’s AI assistant, takes on more autonomous, multi-step work, Hamidi said total token spend is rising sharply while the cost per completed task is falling. Rechat is absorbing that cost rather than passing it to brokerages. Lucy’s pricing hasn’t changed even as her workload has expanded, Hamidi said.
“The cost per completed job is going down,” Hamidi said. “That’s really the number that matters when we’re talking about costs.”
He pointed to Rechat’s in-development agentic website builder as an example. He said engineering optimizations alone have cut production costs for a single site by more than 80 percent since the project started, and that’s separate from any industrywide drop in token prices.
Hamidi expects the shift to reshape how proptech gets priced industrywide, moving away from per-seat software fees and toward payment for completed work. He compared it to Salesforce CEO Marc Benioff’s framing of AI agents as a “digital labor” market.
Measuring return on that spend remains harder than measuring the cost, Hamidi acknowledged.
Rechat can attribute token costs down to individual completed tasks, he said, but tying that spend to closed transactions — the metric he expects brokerages to eventually care about most — isn’t something anyone in the industry has solved.
“The cost per token used today is a vendor’s problem,” Hamidi said. “But I think the moment brokerages start building their own tools, and by the way, a lot of them are starting to do that, this is going to be a brokerage problem as well. Because they’re going to figure out quickly that these things cost a ton.”
‘Firm AI token control’
AI spend and compute costs raise another question for proptech firms and real estate companies, as Hamidi suggested. Traditional real estate software pricing was built around something predictable: seats.
The old way was to buy a CRM license per agent, and the bill scales with headcount. AI breaks that model because token consumption scales with usage intensity, like how many times an agent asks for a listing description, how many documents get processed, and how deep an AI agent reasons through a task, not with how many people are logged in.
Irena Cara
Irena Cara
“Enterprise software models that depend on fixed seat licenses completely collapse when power users cause a 300 percent increase in generative queries, month over month,” said Irena Cara, co-founder of Nova Wrap, a Dubai-based interior design company using proptech while “struggling to tame the burden of increasing operational costs.”
Cara said real estate technology businesses have to absorb huge monthly infrastructure costs if they don’t deploy model routing architecture. This directs property-related queries to simpler models and reserves advanced models for complex queries.
Cara said Nova Wrap implemented AI request throttling and aggressive prompt caching, reducing backend AI costs by 45 percent while improving overall output.
“Nova Wrap is committed to firm AI token control and will reduce costs from runaway API bills,” she said.
Own the model, own the economics
George Zheng, founder and CEO of Edensign, has taken that logic further by not renting frontier models at all. Edensign uses generative AI and proprietary data to reinvent how spaces are visualized, designed and transacted. The company trained its own AI model rather than building on general-purpose systems like Claude, ChatGPT or Gemini.
“Compared to companies that rely on the big models, the expense really depends on how much API volume you’re making. Once you have more customers, usage increases dramatically,” Zheng told Inman. “We spend more upfront, but once the model is trained, the per-token cost is pretty low compared to calling those big models.”
George Zheng
George Zheng
Zheng said Edensign now sells access to its model to other companies, including what he described as a partnership with Bright MLS.
He argued that a specialized model built for real estate tasks has better unit economics at scale than a system stitched together from general-purpose APIs, even though it costs more to build in the first place.
Edensign’s pricing has shifted away from charging by raw token or credit volume toward charging for outcomes, such as a completed room analysis, a marketable staging image, a usable listing recommendation. That’s a distinction Zheng said matters more than most cost conversations in AI acknowledge.
“Don’t optimize the cost per token. Optimize the cost per useful outcome,” Zheng said. “The cheapest generation is not the one with the lowest compute cost. It’s the one that produces a usable result with the fewest retries and the least human correction in the loop.”
Savings don’t stay savings for long
None of this is necessarily unique to real estate and proptech.
Accenture’s report frames it as a Jevons paradox playing out across every industry adopting AI. Cheaper tokens don’t lead to lower AI bills. They lead to more AI use.
The firm’s own simulation found that if token prices fell 25 percent today, only 15 percent of organizations would bank the savings and the rest would reinvest it into doing more.
Accenture’s prescription is a single view of cost, usage and return that CFOs and CIOs jointly own, model access gated by role or task complexity, and a recurring loop of diagnosing and adjusting spend. This is likely closer to how large enterprises manage capital allocation than how most brokerages or proptech startups currently think about a software bill.
The vendors furthest ahead on this aren’t necessarily the ones spending the least on AI. They’re the ones who can say, with a number attached, what that spending is buying.
“We think about the cost per usable property outcome,” Zheng said. “That’s something we can redefine token economics around, in a real estate context, compared to a general-purpose business.”
This October, Inman turns its focus to AI and its rapid rise in real estate. During Artificial Intelligence Month, we’re digging into the startups shaking things up and the established players folding AI into their offerings.
