Lenders are writing loans for procedures they can never repossess, and software now approves some of them. Here is what the investors buying the risk actually get to see.
getty
In May 2025, investors bought $420 million of securities backed by loans used to finance elective medical procedures. The deal carried ratings from Kroll Bond Rating Agency and a familiar securitization structure. But it also came with a not so unusual feature: the underlying loans were unsecured. Unlike a mortgage or auto loan, there is no asset a lender can repossess if a borrower defaults. You cannot take back a set of veneers or a course of IVF. Once the patient leaves the clinic, the lender is left relying largely on the borrower’s promise to repay.
That deal was Cherry’s second of its kind, and bigger than the first. Together the two have moved more than $800 million of this paper into investors’ hands in under a year. The loans behind them sit in a corner of consumer lending with no settled name. I will call it elective-care credit, since that is what it is: borrowing for procedures people choose rather than need, sold at the moment of the choosing.
It is unsecured, frequently five figures, and these days the lending decision is usually made by a machine-learning model rather than a person reading a file. The software is the reason the business got big enough to securitize. Britain runs a year or two behind the U.S., with the same loans sitting on bank balance sheets instead of in bonds. The question worth chewing on is not how clever the models are. It is whether the investors holding the risk can see what those models are doing, and whether anyone has watched this market through a bad year. I do not think anybody has.
Who actually decides
A machine-learning scorecard is, at bottom, a pattern finder. An older credit model leaned on a few numbers: income, existing debts, a bureau score. A newer one runs through hundreds of signals together and hunts for combinations a human reviewer would never spot. Show it enough old loans and it learns which borrowers tend to pay, including some with thin credit files that the old method would have rejected on sight.
There is decent evidence the approach works. The economists Julapa Jagtiani and Catharine Lemieux found that drawing on data beyond the standard bureau file can improve credit predictions, particularly for thin-file borrowers the bureau struggles to read. The vendors that sell these models lean on that finding. Zest AI, one such firm, says its models are supervised and kept “human-in-the-loop”, trained and then locked rather than left to drift. Worth remembering that this is a sales pitch from a company that profits when lenders buy in.
One point matters for anyone tempted to wave the AI flag too hard. This is supervised statistical scoring, the workhorse kind, not a chatbot deciding your loan. The software ranks risk quickly and cheaply enough to clear an applicant while the consent form is still open. That speed is the whole reason the business scales. Slow it down to the pace of a human underwriter and the volumes collapse.
The flip side is that a model fast enough to price a lot of risk is also fast enough to create a lot of it, and the murkier the model, the harder it is to check. FinRegLab, a nonprofit research group working with the Stanford Graduate School of Business economists Laura Blattner and Jann Spiess, has made the case that machine-learning underwriting can complicate fair-lending supervision. Working out whether a tangled model is quietly treating similar borrowers differently is a harder job than it was with the plain scorecards it replaced. More accurate on average, less legible underneath. That trade-off runs through this entire market.
What Wall Street already bought
The surest sign this has become a real asset class is that it trades. Cherry’s first deal, the Securitization Trust 2024-1, issued $400 million of notes backed by roughly $432.5 million of elective-medical receivables. The follow-up in May 2025 was bigger, at $420 million of notes over about $452.8 million of receivables. Securitization sounds technical but the idea is plain. Thousands of small loans go into a pool, the pool throws off cash as borrowers repay, and investors buy slices of that cash flow ranked by how much risk each slice carries.
Who borrows? The collateral answers it. In the 2024 pool, dental work and medical-spa treatments each accounted for roughly 41% of the balance, 41.9% and 41.0% to be precise, with dermatology, optometry and plastic surgery making up the rest. The borrowers were not strapped. The average loan was about $1,600, the weighted-average Vantage score was 715, which is prime, and the loans charged around 14% a year. People with choices are deciding to borrow for their teeth.
Investors do not buy the loans themselves. They buy the slices, known as tranches, and what keeps the top slices safe is something the trade calls credit enhancement: a buffer of spare cash and lower-ranked slices that take the hit before senior holders lose a penny. On the 2025 deal, KBRA rated the four classes from A down to B, with the buffer running from roughly 23% at the top to under 4% at the bottom. Notice what travels and what stays put. The rating goes out into the market; the model that picked the borrowers stays inside the lender. An investor gets the grade and a loss estimate. The thing that did the choosing never leaves the building.
A detail in the 2025 deal deserves a second look. KBRA pointed out that, unlike the first deal, Cross River Bank was not originating fresh receivables for it. To someone holding the paper, who writes the loans, and on what model, is not housekeeping. It is the asset.
Britain takes the slower road
Britain runs a similar engine down a different road. Rather than pool the loans and sell them on, UK providers send borrowers through credit brokers to a bank that keeps the loan on its own books. Go into a clinic for full-mouth implant work and the finance on offer usually runs through a chain. The clinic introduces you to a broker, the broker to a finance partner, and the loan itself is written by a deposit-funded bank.
21D, a Warrington provider of full-mouth implants, sits at the start of one such chain. It is worth being exact about what it is, because there are two very different kinds of technology in this story and they should not be confused. 21D is, in part, an autonomous clinical-AI system: according to a recent account of its architecture, it takes a patient’s scans, runs them through an AI pipeline that plans the surgery end to end without a human planning technician, and prints a surgical guide, with roughly 98% of that workflow automated and the surgeon stepping in only at the operating table. That is the clinical side. On the money side, 21D is simply a credit broker regulated by the Financial Conduct Authority. It does not lend, and the software that plans the implants has nothing to do with the software that approves the loan.
The loan itself is written by Secure Trust Bank, reached through the broker V12 Retail Finance (FCA reference 679653) and, in dental finance, the specialist broker Medenta (FCA reference 715523). The cash does not come from capital markets. It comes from a bank’s deposits. The Sunday Times ranked 21D 24th on its 2025 list of fastest-growing private companies.
That balance-sheet model changes the incentives. A bank that keeps the loan keeps the loss, which tends to concentrate the mind on who gets a yes. Dr. Rajesh Vijay, a co-founder of 21D, says borrowing is the exception among his patients. “Consistently around the 15% mark take finance in the UK,” he told me. “Most patients either do it through savings or via their own bank.” He puts the company’s low default rate down to screening before any treatment begins: a “strict triage pathway… establishing affordability early before we even start,” so that patients “understand that it is a service that provides them with great value and we expect them to make payments in return.”
One clinic is not a market, and a broker sees only its own corner of one. Vijay is frank that the view upstream is cloudy. Getting “a clear answer” out of a finance partner on its approval and decline criteria, he says, “has always been a struggle.” Take that for what it is worth: the person feeding loans into the chain telling you he cannot always see how the lending call gets made.
Whether Britain follows America into the bond market is the live question. The plumbing is regulated and the demand is plainly there. What is missing so far is the will to pool these loans and sell them on. It is a space to watch, because the day a UK lender decides to do it, the same gap between who holds the model and who holds the risk crosses the Atlantic.
The risk nobody can repossess
Take away the deal structures and you are left with a plain unease. A growing stack of unsecured loans, picked by models that outsiders cannot easily inspect, lent to borrowers who have never lived through a downturn together. Any one of those facts is manageable. Put them in a pile and supervisors start shifting in their seats.
Regulators are already circling the wider point-of-sale credit market, even though these particular loans are regulated. In Britain, the FCA has confirmed that buy-now-pay-later lending falls under its rules from 15 July 2026, pulling it under the Consumer Duty and forcing affordability checks before money moves. Its own survey put buy-now-pay-later use at nearly 11 million UK adults in 2024. A longer-term medical loan is a different product from an interest-free instalment plan, so the two should not be lumped together, but the direction is one way. Credit sold on a promise to pay later is getting more scrutiny, not less.
This is where the optimists and the worriers split, and both have a case. The optimists, with the Jagtiani and Lemieux work behind them, argue that sharper models read near-prime and thin-file borrowers more fairly and let people who would once have been turned away get treatment they can afford. The worriers, with FinRegLab’s research behind them, argue that the same opacity that makes the models effective also makes their risks hard for anyone outside the lender to audit. The argument is not settled, for a simple reason: the thing that would settle it has not happened yet.
What the buyer is holding
So think about the investor at the far end of the chain, holding a rated piece of elective-care paper. The lender’s model has read the borrower in close detail. The rating agency has seen a pool and a loss assumption. The investor has a letter grade and a yield. And the one thing that might have tied all of that back to something physical, the collateral, was gone the instant the money paid for the procedure. The gap between the people who can see the borrower and the people who own the loan is the real story here, more than any clever algorithm.
It has felt comfortable because credit has been calm the whole time this market has existed. Good weather only. The honest test of whether the models hold, and whether the buyers grasp what they own, is a bad year that has not arrived.
The bottom line
If you are weighing this paper, or any rated consumer-loan deal sold on a software-driven approval process, treat the credit rating as the start of the work, not the end. Ask who originates the loans now, not who did at launch, what the model actually keys on, and how the pool would behave if defaults rose from today’s calm levels. The yield is visible. The thing generating it is not, and on these loans there is no collateral to fall back on.