AI Is Warping Home Price Negotiations, Real Estate Executives Warn

Buyers and sellers are increasingly turning to general-purpose artificial intelligence tools to price properties and evaluate offers — and industry veterans say the results can derail deals that human agents spent months arranging.

In March, real estate executive Ryan Serhant, CEO of the company bearing his name, posted a video to Instagram describing how a $50 million transaction nearly collapsed after both parties independently consulted ChatGPT. The post has since drawn more than 3 million views.

"At the last minute the seller uses ChatGPT, asks it, 'Should I sell at this price?' And maybe because of how he asked, whatnot, ChatGPT basically told him no, you should not sell at that price, it's worth more," Serhant said.

The buyer then ran his own query, and the tool — made by OpenAI — told him he was overpaying. "It gave him comparables that showed why, without context and without actually understanding the property," Serhant said.

Serhant said he ultimately saved the deal by walking both parties through what AI cannot do. "It doesn't know the future, it can't predict the future. It doesn't know intentions, doesn't know emotions, doesn't know what buyers are circling, doesn't know off-market comparables, doesn't understand, fully, replacement costs, and doesn't actually optimize for the deal," he said. "AI can model a market. It can't model a deal."

Serhant has also launched his own AI-powered workflow automation platform for real estate agents, called S.MPLE, signaling that his concerns are about how the tools are used rather than the technology itself.

Kamini Lane, CEO of Coldwell Banker Realty, said her agents are observing a growing share of clients — on both sides of transactions — consulting large language models such as Anthropic's Claude and OpenAI's ChatGPT before setting prices or making offers.

Lane said these general-purpose tools miss critical local knowledge. "One of the most important things that agents can see, that ChatGPT, or any other AI tool is not going to know, is [what's] up and coming. So neighborhoods that are up and coming, design features that are up and coming," she said. "Anecdotal data that agents are aggregating through their conversations, that is something that no AI tool is ever going to be able to aggregate in the same way that a real estate professional can."

Lane also raised a structural concern about how generative AI models are trained. "Artificial intelligence is trained to be sycophantic, it's trained to give you the answers that you want, so that you will continue to engage, and so AI is more likely to give you the price that you want versus the price at which a home is going to sell for," she said.

Zillow, which launched its algorithmic Zestimate pricing feature in 2006, recently introduced an "AI mode" designed to guide homebuyers through their search by learning their specific preferences and enabling personalized conversations with its pricing tool.

Nicholas Stevens, Zillow's vice president of product and AI, said the platform's approach differs from generic AI tools because agents are required to upload detailed floor plans, 3D visual captures, and comprehensive property data before the system offers guidance. "It actually sees a remodeled kitchen. It actually sees upgrades in the house, and that's useful, both for buyers but also homeowners thinking about selling or remodeling as well," Stevens said.

Stevens framed Zillow's approach as context-dependent rather than generalized. "AI guidance for consumers needs to be connected to real context, real data, real ability to take action," he said. "Then that AI guidance needs to be deeply connected to what a real estate agent is attempting to do."

Zillow's AI mode is currently oriented toward buyers; Stevens said a seller-facing tool is in development.

The concerns raised by Lane and Serhant reflect a broader tension taking shape across industries where AI is being applied to high-stakes, data-rich decisions — one in which the tools are capable enough to generate convincing outputs yet not calibrated to the judgment, local knowledge, or emotional complexity that human professionals bring to individual transactions.

Disclaimer