The Buyer No AI Model Can See: Inside Washington’s All-Cash 32%
The settled position on AI in this business is that machines take the paperwork and people keep the relationships. Ryan Serhant made the cleanest version of that argument in Forbes last month, splitting the work into revenue-supporting and revenue-producing, then handing the first category to a fleet of AI agents running operations, finance, and sales support so his people can spend their hours in rooms rather than spreadsheets. It's a good argument. In Washington, as I'm sure in New York, the constraint has never really been the paperwork. The constraint is knowing whose name is about to go on it.
32% of luxury sales across the DC area closed all cash last quarter, against 17% of the market overall. That gap usually gets read as a story about wealth. I want you to read it as a story about visibility instead. Then, consider who sits inside it: government contracting principals holding an exit that won't surface in a public filing for two more quarters, the diplomatic tier buying through entities built specifically to make the purchaser unsearchable, private equity partners mid-raise who have very good reasons not to have an eight-figure acquisition attached to their name in a public database, chief executives relocating on a package negotiated in closed session. A meaningful share of these buyers can't be publicly identified before closing, and some of them can't be identified after it either. No model has visibility into who in that group is about to move, because the signal that would train it never becomes public. It lives in a phone call... or if you saw my video from the other day, in a text message.
Dismissing the technology would be lazy. A model will tell you that McLean's 22101 and Bethesda's 20817 led the Mid-Atlantic in luxury sales volume last quarter, that five of the region's top ten luxury ZIP codes sit inside the DC Metro with Georgetown's 20007 and Northwest's 20016 among them, and that more than half of everything sold in McLean last quarter was luxury inventory. All of it is real, useful, and 90 days late. Every one of those data points is a picture of a decision somebody already made, in a room the machine wasn't in.
Serhant's $50,000,000 penthouse is the clearest illustration anyone has offered. Buyer and seller each ran the price past ChatGPT at the eleventh hour, and it told the buyer he was overpaying while telling the seller he was underselling, which is roughly what you'd expect from a system optimized to agree with whoever is typing. (Both parties got the answer they walked in with, which is a service of a kind, I guess.) The deal nearly died there. What brought it back was off-market context, the sort no machine can scrape. Serhant's own summary of the limitation is the sharpest I've heard: large language models "know the history of the internet, they don't know the path forward."
That story circulates as an AI failure. The more useful diagnosis is that it was a comparable sales failure. The model wasn't wrong about the data it had; it was confident about a trophy asset with no honest comp set, and confidence without a comp set is a well-formatted guess. At the top of this market, that describes a good number of properties.
Part one of two. Continued in What a Machine Was Never Going to Do.
Marc Cashin is the Founder and CEO of FORWARD at Corcoran McEnearney, representing buyers and sellers in Washington DC, Maryland, and Northern Virginia. This article was first published on LinkedIn.
