AI That Helps You Refinance Will Make Mortgages More Expensive
American mortgages are cheap partly because most people are too lazy to refinance when rates drop. Once AI agents remind everyone to refinance promptly, that option gets repriced, and mortgage rates rise.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI reminders to refinance make mortgages more expensive
American homeowners generally don't act when rates drop: Matt met someone on a plane who was bragging about his mortgage rate and planning to sell and buy a bigger house, and Matt talked him out of it. Morgan Stanley research suggests that if an AI agent reminds you to refinance when rates drop by 200 basis points, future refinancing will be much easier. Sounds like it saves you money, but more frequent refinancing means the prepayment option that people previously were too lazy to exercise gets exercised, and lenders reprice for that, so everyone's mortgage rate rises by 10 to 20 basis points.
— Matt LevineSmarter homeowners mean fewer people can buy homes
Fabrice Touré—former head of mortgage derivatives at Goldman Sachs, now a finance academic—writes bluntly in his paper: policies that make refinancing more frequent also raise equilibrium mortgage rates and reduce the likelihood that many borrowers can get credit. The reason is that the US 30-year fixed-rate, prepayable mortgage carries an expensive option, but most people exercise it poorly, which is why lenders are willing to lower rates. Once institutions or AI make everyone more rational, the option gets repriced, rates rise, and the people who were previously subsidized by low rates become fewer; more families can't afford homes.
— Matt LevineModels priced on past behavior will get blindsided
The shock hits investors first. Morgan Stanley warns that historical prepayment behavior will underestimate future prepayments; any model trained on past data misprices the fact that borrowers are more willing to exercise their prepayment option. As a result, some mortgage-related assets look cheap but are actually built on the assumption that borrowers stay lazy; when AI compresses refinancing friction from two weeks to ten minutes, these assets need a full repricing. Matt extends the logic: credit card points, high-yield savings accounts, and life insurance policies all work the same way—the financial system largely runs on customers not using every option to the fullest.
— Matt LevineAI makes the billable hour model obsolete
AI's squeeze on professional services hits law firms first: clients think AI can do the work of a squad of junior associates and demand lower fees. Firms bill by the hour on the surface, but the real profit comes from partners selling junior associates' hours at a high rate and paying them a lower salary; if AI does the grunt work, charging for junior associate hours no longer holds. Senior lawyers' judgment is still valuable, but it needs a different pricing model. The bigger problem: when juniors no longer spend hundreds of hours on basic tasks, where do the future partners develop their instincts?
— Matt LevineTop partners can go solo with AI
Since senior lawyers' value doesn't depend on headcount, the most talented will do the math first: rather than stay at a big platform to support a partner committee and train a pipeline of juniors, they'll leave and start a boutique with just themselves and six AI agents. Chris Kircher, a former Quinn Emanuel partner, has already resigned to start an AI-native firm built on that bet. If enough senior people leave, big firms can't retain them even if they know apprenticeships matter—the revenue-generating capacity leaves, the money-losing training system stays, and big firms struggle to survive.
— Matt LevineAmazon's auction hides a phantom bidder
The FTC and 22 state attorneys general are suing Amazon, saying its ad auction isn't the second-price auction it's advertised to be. In a normal second-price auction, the highest bidder pays the second-highest bid; Amazon adds a "soft reserve price"—the platform's own valuation of the ad slot, like a bidder hiding in the shadows. If a bid exceeds the soft reserve, the soft reserve becomes the second price; if a bid is below the soft reserve, the winner actually pays their own maximum bid. Amazon's response: ad auctions are a real-time black box anyway; no one can know the soft reserve in advance, and advertisers just look at return on ad spend.
— Matt LevineBidding against the chandelier is legal
The financial world forbids exchanges from secretly placing their own buy orders, but art auction houses have a legal performance. The seller sets an undisclosed reserve price, and when live bids don't reach it, the auctioneer calls out bids into empty space, pushing the price up to the reserve—called a "chandelier bid" because he often looks up at the chandelier. New York requires auction catalogs to disclose that the auctioneer may bid on behalf of the seller up to the reserve and won't indicate which bids come from the seller. Amazon's soft reserve is similarly the platform bidding for itself, but it doesn't explain the rules in such a footnote—that's what the FTC is pouncing on.
— Matt LevineSwitching ETFs dodges dividend taxes
Foreign funds (e.g., Cayman structures) pay withholding tax on US stock dividends, which drags on the thin margins of basis trades. Some traders exploit two nearly identical S&P 500 ETFs whose dividend dates differ by about a week: BlackRock's IVV and Vanguard's VOO. The tactic: sell IVV before its ex-dividend date and switch into VOO; when VOO is about to pay, switch back. After the ex-date, the fund price drops by roughly the dividend amount, so the investor avoids the taxed dividend and earns the same money back by selling high and buying low. This is an institutional tax maneuver; retail investors copying it would generate large capital gains taxes—not a TikTok money tip.
— Matt LevineIn their own words · checked verbatim
many policies that lead to more frequent refinancing also increase equilibrium mortgage rates and thus reduce residential mortgage credit access for a large number of borrowers.
Matt Levine6:09
The best, most crazy thing the U.S. mortgage market does is offer 30-year fixed rate mortgages that are prepayable.
Matt Levine8:12
historical prepay behavior will understate future prepays. So models trained on the past mispriced this option that borrowers will be more inclined to exercise.
Matt Levine10:13
if all the grunt work can be done by AI, then, like, that model doesn't work.
Matt Levine13:16
Nobody buys ads based on like simplified descriptions of the auction mechanics. They buy ads based on like return on ad spend, right?
Matt Levine21:29
This is not something that an influencer should recommend on TikTok.
Matt Levine33:39
Figures
| Morgan Stanley's estimated rise in mortgage rates after AI increases refinancing frequency | 10-20 basis points | 5:08 |
| Number of state attorneys general joining the FTC in suing Amazon | 22 states | 18:26 |
| Difference in dividend dates between IVV and VOO, two S&P 500 ETFs | about a week | 30:36 |
Glossary
- second-price auction
- An auction where the highest bidder wins but pays only the second-highest bid.
- soft reserve price
- A hidden minimum price set by the platform or seller in an auction, effectively adding a phantom bidder.
- chandelier bid
- An auctioneer's bid on behalf of the seller, calling out bids to empty space to push the price toward the reserve; named because the auctioneer often looks at the chandelier.
- basis trade
- A trade that buys the spot asset and sells the futures contract to capture the difference, essentially providing funding.
- apprenticeship model
- The professional services model where juniors do large amounts of low-level work and gradually develop into senior partners.
- withholding tax
- A tax deducted at source on payments like dividends; often levied on foreign investors.
How to listen
People in mortgages, asset management, law firm management, and ad auction operations, plus any founder concerned with how AI removing friction changes financial product pricing.
Skip the first ~4 minutes of pleasantries and introductions; from minute 5, the AI refinancing discussion begins and there are no filler segments.