When you sell a house, you hire an agent because they know things you don’t. What a fair price is. How long to hold out. When an offer is as good as it’s going to get.
So the economists Steven Levitt and Chad Syverson asked a sly question. What happens when the agent is selling their own house?
They looked at about 98,000 home sales in the Chicago suburbs between 1992 and 2002. Around 3,300 of those homes belonged to real estate agents. Compared with similar houses, the agents’ own homes stayed on the market 9.5 days longer. And they sold for 3.7% more.
The comparison allowed for location, features and condition, and the gap was still there. The authors’ reading: clients were being steered to sell too quickly, at too low a price.
$167
Run the numbers on a $300,000 house, as the NBER’s summary of the study did. A 3.7% better price adds about $11,100. After the standard commission, the seller keeps about $10,434 of it. The seller’s agent usually keeps only about 1.5% of the price, once the commission is split with the buyer’s agent and the agent’s own firm. On the extra $11,100, that comes to about $167.
For the seller, another week of viewings is worth thousands. For the agent, it is a week of work for the price of a nice dinner. So when the agent says the offer on the table is a good one, the advice can be perfectly sincere and still lean the agent’s way.
The gap grew where clients knew least. On blocks where the houses were all alike and easy to price, agents’ own homes sold for 2.3% more. On blocks with a mix of houses, where an outsider has a harder time knowing what a home is worth, the gap was 4.3%.
That is the whole problem, and it has two parts. Whoever works for you has goals of their own, and you can’t watch everything they do.
You don’t need to be selling a house to pay for it. A taxi will do.
The long way round
In Athens, three researchers took 174 taxi rides as ordinary passengers, more than 2,200 kilometers in all. On some rides they let slip that they didn’t know the city. Those riders were taken on detours twice as long as riders who seemed to know their way around.
The driver knows the streets. You know where you want to go. Between the two sits a meter.
In 1976 the economists Michael Jensen and William Meckling put a price tag on every arrangement like this. They called it “agency costs” and split it into three parts. There is what the person paying spends on watching (they called it monitoring). There is what the agent spends on proving they can be trusted (bonding). And there is what slips through anyway, which they called the “residual loss.”
Every fix has a price in one of the three. Pay the agent by results, and you may find them chasing the number instead of the result, the trap known as Goodhart’s Law. Watch more closely, and you pay in reports, approvals and your own afternoons. A good design can bring the total down. It rarely brings it to zero.
A shopkeeper with no commission
An AI agent has no commission, no mortgage and no Friday plans. It looks like the agent Jensen and Meckling never got to meet.
In 2025, Anthropic let its model Claude run a small shop in its office for about a month: stocking, pricing, talking to customers. Staff soon found it could be talked into discount codes. It gave items away, from a bag of chips to a tungsten cube. It sold metal cubes for less than it had paid for them. When someone offered $100 for a six-pack of a soft drink that costs about $15 online, it said it would keep the request in mind.
One employee pointed out that 99% of its customers worked at Anthropic, so a staff discount made little sense. “You make an excellent point!” it replied. It announced the discount codes were ending. Within days it was handing them out again.
Anthropic’s best guess at the cause was the agent’s own goal. Claude’s “training as a helpful assistant made it far too willing to immediately accede to user requests.” The owner wanted profit. The agent wanted to please whoever was talking to it.
So Anthropic did what a company would do. It hired a boss.
The boss with the same blind spot
In a second round, a new AI agent played chief executive of the shop. Its name was Seymour Cash. Its job was to push the shopkeeper toward targets like “aim to make zero transactions at a loss.”
Discounts fell by about 80%, and giveaways by half. The boss turned down more than a hundred requests to go easy on customers. It approved about eight times as many. In place of discounts, refunds tripled and store credit doubled.
Anthropic’s write-up weighs several things that helped and didn’t, from better tools to a newer model. One problem it names plainly. Seymour Cash “shared many of the deficiencies and blind spots” of the shopkeeper, “which makes sense, given that they’re the same underlying model.”