For Sale By AI:
The Post-Mortem.

Data and AI are what I do for a living. So when it came time to sell the house, I couldn't help myself: I wanted to know whether the technology I spend all day on could actually run the most consequential transaction most people ever make.

I don't mean a chatbot fielding questions about the house. I mean the actual sale — what to list it at, which homes to compare it against, how to stage it, which fixtures to pull and which to leave, who to hire, what the photographer should shoot, what the ads should say, even how this page should read, right down to the tone of every reply to every tire-kicker who emailed. The AI made the calls. I was the hands.

The commission mattered too — a 2.5% listing side on a home like this is real money, and I'm not going to pretend I didn't notice. But the commission was the excuse. The actual question was simpler and more interesting:

When an AI runs the sale instead of an agent, can the market tell the difference?

It couldn't. What follows is the proof — and everything it taught me along the way.

---

01 The judgment layer is solved

This is the headline, so I'll say it plainly: nearly everything a traditional agent charges 3% for is judgment, and AI does that judgment now — quickly, cheaply, and, in my experience, well.

Think about what you're actually paying an agent for: the call on what to list at, the instinct to stage the basement so it reads as a bedroom and not an office, the eye to swap the dated brass and keep the hardwoods, the read on how to answer a lowball without getting defensive, the sense of which contingency to fight and which to wave through. My favorite one was the smallest. The AI told me to leave the sump pump out of the listing entirely — it's a good thing to own, but to a nervous buyer the word just whispers "this basement floods." None of that is manual labor. It's pattern-matching against thousands of past deals and a feel for how buyers actually think, which happens to be the thing a good model does better than almost anyone.

The pricing was the one place AI was honestly humble, and I want to be honest about it too. I had it backtest its own method against actual sales in the neighborhood, and on a house as unusual as this one, the error bars were wide. So it didn't hand me a number. It handed me a range — and the reasoning behind the range — and I made the call to list at the low end of it to buy velocity. The decision to undershoot was mine. The range was the AI's, and the market settled the argument: both offers I took seriously came in at list or above, and the house closed at full asking price. The range was right.

That distinction held up across the whole sale. The model was a phenomenal analyst and advisor. It was never the one who had to decide to sign.

02 AI does the analysis a human never would

If the judgment layer is the headline, this is the part that actually surprised me.

No agent on earth was going to pull the county's GIS records for every single-family parcel in Aurora Highlands, classify each one as corner or interior — by testing which lots touched two streets at once — filter for size, and prove with a map that my house is one of only 22 non-corner homes in the entire neighborhood on a 9,000-plus square foot triple lot. It's 458 parcels. Doing that by hand is economically insane for a human being. No commission justifies the hours.

For the AI it was an afternoon — and the result changed how I think about the whole job, because the analysis became the marketing. The scarcity study wasn't a back-office spreadsheet nobody sees. It turned into the map on the homepage, the "1 of 22" line in the ads, the single most persuasive argument for the price. The work that was too expensive for a human to bother with became the centerpiece of the pitch.

It turned up something I didn't know about my own house, too. Every qualifying lot in the neighborhood — all 22 — is three of the original 1920s building lots stitched together, because the place was platted in 25-by-120-foot strips and the only way to reach 9,000 square feet is to assemble three of them. Mine is Lots 5, 6, and 7 of Block 20. I'd owned the house for six years and never knew. And the analysis had one limit it wouldn't cross: when Zillow threw up a "prove you're human" check, the AI refused to click it — solving a CAPTCHA was a line it wouldn't step over — and handed the task back to me. Relentless, but bounded.

Then it did the same thing with data no agent could ever touch: mine. An agent can list your house, but they can't see your Gmail, your Costco order history, or the closing file from the day you bought the place. The AI could. The "$80,000 in receipted improvements" on my listing wasn't a round number I estimated — the AI went back through years of my own email and store records and rebuilt the actual invoice trail, line by line: the Trane HVAC system, the refinished floors, the egress window that turned the basement into a legal bedroom, down to the receipt for the garbage disposal. The story of the lot came out of my own closing documents — it read the House Location Survey the title company had sent me six years earlier, a flat image scan with no text in it at all, and pulled the exact platting off the page.

That's the part I'd underline for anyone wondering where this goes. The edge wasn't the AI alone, and it wasn't me alone. It was the AI fused with my own private record — the receipts, the emails, the documents only I had — turned into a story a buyer could actually verify. A brokerage can hire a thousand agents. It cannot get into your Gmail. The seller is the only one who holds this data, and the AI is the first thing that's ever been patient enough to assemble all of it into something persuasive. That combination is the real unlock — and it's the one piece of this that doesn't belong to the industry.

Both halves together are the asymmetric edge, and it's the one I'd tell other sellers to chase first. Not "AI replies to emails faster." AI does the rigorous, boring, data-heavy work no rational human would do unpaid — on the public record and on your own — and that work is often the most convincing thing you can put in front of a buyer.

03 The execution layer is still entirely human

Now the counterweight, because a post-mortem that only lists the wins is a brochure.

What no model could do was climb a ladder. Crimp a wire. Drive to Home Depot at 9 p.m. because the thing I needed wasn't the thing I bought. Reposition a hose when the sprinkler timer misfired at 6 a.m. Meet the stager, the photographer, the egress crew, the floor guys, and let them in, and walk the punch list, and do it again the next day. I did hundreds of those things. Over a few weeks. With a newborn in the house.

So, honestly: I don't think most homeowners should do this. Not yet. The interesting finding isn't that AI replaces realtors. It's that the judgment layer is now solved and the execution layer is still 100% human. A robot that can actually do the physical work in your house changes that equation completely. We're not there. But we're closer than I expected when I started — and the gap is narrowing from the judgment side, fast.

04 The best contractor is almost never the best-reviewed one

This is the most generalizable thing I learned, and it has nothing to do with real estate.

Yelp, Angi, Thumbtack, Houzz — none of them rank by quality. They rank by who pays to advertise and who has the time to chase reviews. The genuinely outstanding trades are heads-down, booked solid through word of mouth, and essentially invisible online. They don't need you to find them, which is exactly why they're hard to find.

So I had the AI build a tool to find them anyway — and it worked by doing the opposite of what you'd expect. It pulled Arlington County's entire building-permit history, about 18,000 records, and the Commonwealth of Virginia's full contractor-license file, roughly 50,000 more. Then it ranked contractors on the one signal a review can't fake: did their permits actually pass final inspection? A contractor whose permits are all closed out — pulled, inspected, finaled — is telling you something no five-star average can. And it inverted the usual logic on purpose: it up-ranked the trades with strong permit records and few or zero online reviews. A thin Yelp presence wasn't a red flag. It was the signal — the fingerprint of someone too busy doing the work to farm stars.

And then came the part that made the whole lesson click. The single best contractor the records turned up — a near-perfect inspection history, permit after permit closed clean over more than a decade — almost didn't surface at all. His master license, twenty-eight years old, was filed under his own name instead of his company's, so the matcher skipped right past it. The county had misspelled his business on half his permits, so his real track record looked like a fraction of itself. The tool's first pass ranked him mediocre. By the public record he was one of the best in the entire county — and he was nearly invisible to the very system I'd built to find invisible people.

That's the lesson, twice over. The best are hard to find because being good and being findable are different skills. If you want to find them, stop reading reviews and start reading the public record — permits pulled, licenses held, inspections passed. And if you don't want to build a database to hire a plumber: cast a wide net, ignore the stars, and weight the only two things that reliably predict quality — how clearly someone writes back, and how much thought is in the quote.

05 The agents did not love this

I'll tell this one straight because it's the funniest and the most revealing.

I held a brokers' open house. The AI built the guest list by pulling every agent who'd sold a home in my zip code, ranking them by volume, and writing each one a note that opened with their own number — "you've closed 118 homes in 22202; you're on a different planet." Around 250 agents, individually courted. The AI even modeled the likely turnout — three to ten percent — and reassured me that a quiet RSVP list was normal, not a bad sign. I catered for thirty. Two showed up. I bought Chick-fil-A for a crowd and ate the leftovers, more or less alone, for the better part of a month.

And a handful of the agents who did respond didn't respond warmly. One — at a national brokerage that markets itself as the AI-forward future of the business — wrote, unprompted, to inform me that what I was doing was illegal. She opened by noting kindly that she could see I was on paternity leave, then explained that "only a licensed real estate broker can assist a buyer or seller in the purchase or sale of a home for compensation." She was describing me selling my own house. In the very same email she conceded the exception that covers exactly that — "a seller can sell their own property on their own" — and then predicted I'd fail and hire an agent anyway. Another lectured me that my 2.5% buyer-side offer was below "standard," and that I'd hurt my own marketing by leaving the photos out of the invitation. The replies had a tell in common: not one of them engaged with whether it was working. They went straight to why I wasn't allowed to do it.

Which, once I got over the sting of it, I found genuinely clarifying. People don't get hostile about things that don't threaten them. The reaction wasn't about me being unqualified. It was about the model working.

I don't hold it against them. If someone ran a public experiment designed to show that the core of my job could be automated, I'd be prickly too. But it's worth saying out loud: the resistance to this isn't coming from the data. The data was fine. It's coming from the incentives. And incentives are a much harder thing to change than software.

06 The offer nobody teaches you to read

This is where it got financially interesting, and where I learned something I genuinely didn't know going in.

I received multiple offers, both at list or above. Two of them are worth comparing, because they taught the whole lesson. One came in about $10,000 over asking — and carried a 3% buyer-agent commission. The other came in right at list and carried no buyer-side commission at all. On paper the first was higher. In my bank account the second was higher, and not by a little — because three percent of $1.8 million is a lot more than ten thousand dollars. I took the second. The headline number, the one a seller's eye goes to first, was the worse deal.

Gross price is a vanity metric. Net proceeds are the truth. And the gap between them is, almost always, a commission — frequently one the seller is paying on behalf of a buyer they never met, to an agent on the other side of the table.

Which leads to the part I think actually matters for where this whole market is going.

07 The commission that quietly changes hands

When an unrepresented buyer shows up in a normal sale — someone who walks in without their own agent — here's what usually happens to the buyer-side commission that was budgeted into the deal: the listing agent keeps it, or steers it to an agent friend who "represents" the buyer in name only. The money was always going to be paid. The only question is who catches it. And the unrepresented buyer, the one person who could have captured it, almost never does.

In my sale, there was no listing agent to catch it. So when the buyer came to the table without an agent of their own, that money didn't leak sideways to a broker. The house closed at full list price with no buyer-side commission paid to anyone — the 3% that normally evaporates into the transaction simply didn't exist. It stayed split between the two people who actually had skin in the deal: the buyer, who didn't finance someone else's fee into their mortgage, and me, who kept my full asking price. Every dollar went to a principal instead of an intermediary.

That is a structural transfer that happens millions of times a year, mostly invisibly, and I only saw it clearly because I'd removed the intermediary who normally collects it.

For whatever a guy who sold one house is worth, here's where I think it goes. The listing-side commission is already under pressure — flat-fee MLS services like the one I used are eating it. The next thing to unbundle is the buyer side. As buyers get AI tools to evaluate homes, read inspections, and sanity-check contracts themselves, more of them will go unrepresented — and the 2.5-3% that currently flows almost automatically to a buyer's agent becomes contestable. It can go three places: to the buyer as a lower price, to the seller as higher net, or to whichever intermediary is still standing when the music stops. For most of the last forty years it's gone to the intermediary by default. I don't think that default survives the decade.

The brokerage industry is sitting on something like a hundred billion dollars a year in commissions. The fight over who keeps that money has barely started. This sale was a tiny, one-house data point in it. But the data point pointed somewhere.

08 What the industry actually owns

If judgment is the part the AI ate, the fair question is what's left — what the business still controls that a homeowner can't just walk around. I hit exactly two things, and they're worth naming, because they are the real moats, and one of them honestly gave me pause.

The first is the MLS. The database that every buyer's agent and every listing site pulls from is gated to licensed members, and for decades that gate was the whole moat: list with an agent or be invisible. That part is already broken. A flat-fee service will put your home on the MLS for a few hundred dollars with no agent and no commission attached — I used Beycome; Houzeo and a handful of others do the same thing. That one step is what makes everything else possible. Because the portals syndicate straight from the MLS, my house showed up on Zillow, Redfin, and Realtor.com, and in every agent's search, looking exactly like an agent-listed home. The gate is still standing. The toll just dropped from two and a half percent to about three hundred bucks.

The second one I didn't see coming, and it's the part nobody tells you about: the paperwork is copyrighted. The standard purchase contract — the offer form the whole industry uses without thinking — belongs to the National Association of Realtors and its state and local affiliates, and it's licensed only to members. As a for-sale-by-owner seller, I wasn't entitled to use it. So I couldn't just hand a buyer "the standard contract" everyone expects. I had to commission a neutral purchase agreement through a title company and run the deal on that instead, which cost time, cost money, and quietly added real risk to the biggest transaction of my life — a one-off contract is exactly the kind of place a deal springs a leak. That isn't a problem AI makes disappear. It's a moat built out of copyright and habit, and it was the one piece of going around the industry that genuinely made me stop and think.

So that's the honest shape of the thing. The industry's defenses were never the judgment; that's the part being commoditized while we watch. They're the MLS gateway and the legal scaffolding around the contract. One of those has already fallen to a few-hundred-dollar service. The other is a form behind a membership wall — and I wouldn't bet on it holding forever either.

09 The part nobody warns you about

I want to end on the thing I was least prepared for, because the rest of this essay makes me sound more clinical than I actually felt.

The AI removed the decision burden. It did not remove the emotional one.

When a buyer walked through and passed, I felt it. When a showing was quiet, when the feedback was lukewarm, when a day went by with nothing, I felt that too — a low hum of what if I'm wrong, what if nobody wants it, what if the confident data guy mispriced his own house in front of everyone he knows. I had the analysis. I had the comps. I had a model telling me, correctly, that the price was right and the traffic was normal and that passing is what most buyers do most of the time. And I still had to fight the spiral.

That surprised me more than anything technical did. I assumed that if you took the decisions off a person's plate, you'd take the stress with them. You don't. The judgment was outsourced. The worry was mine, start to finish, and no model was going to carry it for me.

Which is maybe the truest summary of the whole experiment. AI ran the sale. I felt the sale. Those turned out to be two completely different jobs.

---

What I actually believe now

The judgment layer is solved. An AI priced, positioned, staged, and marketed a $1.785 million home as well as any agent I've worked with, did analysis no agent ever would, and the house closed at full list price in June — with not one dollar of buyer-side commission paid to anyone. I wasn't the first to try this; a homeowner in Florida and a reporter in upstate New York beat me to it. But their sales topped out under a million dollars, and as far as I can tell, this is the most expensive home anyone has sold this way — one owner, no listing agent, an AI in the decision seat the whole way. That part of the job — the part that's been worth 3% for half a century — is being commoditized in real time, and the people whose income depends on it know it, which is why some of them are angry.

The execution layer and the emotional layer are still entirely human. Until a robot can crimp a wire and a person can stop caring whether the neighbors think they botched it, you cannot fully automate selling a house, and you probably shouldn't try.

And the real money — the thing this experiment accidentally put a spotlight on — isn't the listing commission everyone argues about. It's the buyer-side commission that changes hands quietly, and increasingly, doesn't have to. When I ran this sale myself, that money came home to my family instead of leaking to an intermediary. As the tools that made that possible get into more hands, that's going to happen more often, to more people, for more of the largest transaction of their lives.

I'm not telling you to do what I did. It was a lot of work, and I had a newborn, and I ate a frankly upsetting amount of cold Chick-fil-A. But I'd do it again. The market couldn't tell the difference — and the difference went to me.

— Stuart