Most explanations of mortgage default read like a checklist of foreclosure timelines pulled from a state-law primer, and they miss the part that actually matters — by the time a borrower's payment is missed, the default has almost always already happened, and the only real question is whether anyone in the chain of custody saw it coming. I've spent a decade looking at this problem from both sides — first building Pagaya's consumer-loan origination platform at roughly $10 billion a year of flow, and now at Fundable, where my co-founder and I build the AI platform that fund and asset managers run their operation on: a single source of truth across everything the fund touches, plus autonomous agents that do the work of each function, from diligence and acquisition through monitoring and workout. This post is the version of "what happens when a borrower defaults" that I wish existed when I was trying to learn this for the first time, written from the perspective of the operators who live with the consequences.
The 30-60-90 sequence — what mechanically happens
The clock starts the day the payment is missed, and the servicer's playbook fires automatically in stages. The first fifteen days are essentially a grace period — a late fee is assessed, automated reminders go out, and no real action is taken against the loan. By day thirty the loan is officially delinquent, and the servicer's loss-mitigation team picks it up, opens a borrower-outreach loop, and starts trying to understand whether this is a one-off cash crunch or the start of something structural. Around day sixty the loan becomes seriously delinquent, credit-bureau reporting kicks in, and the servicer evaluates workout options — forbearance, a repayment plan, a modification, or some hybrid. Day ninety is the inflection point, because the loan is now in default under most note covenants, the servicer issues a formal demand letter or notice of intent to foreclose, and the file routes to a foreclosure attorney in judicial states.
What most people miss is that the loan was almost always already in trouble well before day one of the missed payment — the borrower's debt-to-income deteriorated, a property-tax delinquency was sitting unmonitored on the county recorder's website, an LLC ownership change happened quietly, or a homeowners-insurance lapse was filed and never propagated back to the servicer's tape. In the consumer-credit world I came out of, the dominant default pattern wasn't the cash-crunch borrower who paid for two years and then hit a life event — it was the never-really-was-paying borrower, the early-defaulter who fell behind in the first six months and exposed mis-underwriting at origination. That pattern shows up in mortgages too, particularly in non-QM and bridge, and it's the one that institutional buyers should be most paranoid about during diligence, because an early default is almost always a story about something the underwriting missed, not something that subsequently went wrong.
The biggest misconception about default
The biggest misconception is that default is a discrete event — a borrower was paying, then they stopped, and now you deal with it. In practice, default is the visible tail of a distribution of warning signals that almost always existed beforehand and went unmonitored. Property tax delinquency. Insurance lapse. A title cloud from a divorce or an LLC restructure. A property-condition deterioration visible in permit records. A debt-to-income shift that the originator's underwriting captured at funding and then never refreshed for the entire life of the loan.
One of my advisors — who scaled a real-estate credit book from zero into the billions at a major alternative-credit shop, and ran credit at a money-center bank before that — put this in language I now repeat constantly:
"My bank never asks me to verify my income ever again."
Once a loan funds, the lender's view of the borrower freezes in time, and five years later that same borrower could be making half their original income, levered up across two more properties, and the servicer wouldn't know until the payment stops. That's the misconception. The default didn't begin when the payment stopped. The default became visible when the payment stopped.
Two real catches — what the surprise actually looks like
The cleanest way to make this concrete is to walk through two recent examples, both real, both from the last sixty days of our diligence work.
The first happened during a tape-vs-documents validation we ran for a residential-mortgage servicer affiliated with a $60B-AUM institutional credit investor — the servicer had purchased a non-QM tape and was about to onboard it for special servicing. Our platform pulled public-record enrichment on every loan and flagged one as having an active pre-foreclosure filing already in the county recorder's office. The originator's tape said "current." The buyer's analysts had reviewed the file and signed off. Both missed it. The servicer's transaction lead put it bluntly afterward — "I'm curious how that's not something we were able to flag on our end." The cost of that miss isn't just the loan, it's the principle that the data in the tape and the reality on the ground had diverged, and nobody downstream had a way to see it.
The second was sharper. A non-QM originator was sending a tape to that same buyer, and we ran our AI analyst over the loan documents and cross-referenced the borrower's three bank-statement PDFs against the liquid-assets line of the tape. The tape claimed an ending bank balance that exceeded what the statements actually showed by $1.5 million — which is not a rounding error, that's the difference between a creditworthy buyer and a fraud risk. The originator hadn't caught it because their workflow doesn't reconcile documents to numbers, it just collects documents. The buyer's executive, watching us walk through it live, said the quiet part out loud — "see what we're missing standpoint."
Both of those loans were going to default eventually. They were already in trouble. The question wasn't whether the default would happen, it was whether anyone in the chain — the originator, the buyer, the servicer, the third-party-review firm — would see it before money changed hands.
How default plays out across loan types
Residential prime defaults are mostly macro-driven — unemployment shocks, divorce, medical, in roughly that order — and the servicer playbook is well-defined because the regulatory frame around CFPB, RESPA, and state foreclosure law has been hammered out over decades. Recovery is slow but predictable, and modifications work because the borrower usually wants to keep the home.
Non-QM is where the surprises live, because these loans are underwritten on bank statements, DSCR, or asset-depletion methodologies, meaning the original underwriting was already non-standard and the variability of borrower outcomes is wider. A DSCR loan on a short-term-rental property where the underlying rental income evaporates can go from current to seriously delinquent in two months, and the originator's bid in the secondary market is also tighter because the buyer is pricing in higher loss severity from day one.
Residential bridge and fix-and-flip is its own animal entirely — short duration, six to eighteen months, interest-only, balloon at maturity, with the borrower operating the property rather than living in it. Default in this segment usually means the rehab stalled, the borrower went over budget on construction, or the exit through refinance or sale didn't materialize, and recovery is almost always operational in the sense that the lender has to either step in and finish the project, sell the property as-is, or restructure into a term loan. Our long-running data partner — who's run distressed residential bridge loans for a decade and built a longitudinal dataset of roughly 11,500 loans with 1,500 funded outcomes — shows that the worst recoveries in this segment correlate with late-stage rehab abandonment, meaning the borrower has drawn most of the construction budget, the walls are open, and the asset is worth less than the loan.
And then there's the small-balance commercial non-QM market — the under-five-million-dollar, mom-and-pop-investor, regional-lender, credit-union segment that institutional buyers under-cover the most. The same advisor I quoted earlier called it "a market I think is super untapped," and the reason it's untapped is precisely because the diligence cost per loan is high relative to the loan size, which is exactly the kind of friction that AI-native operations are built to remove.
What recovery actually looks like at the loan level
Recovery on a non-performing loan is the bleeding wound of the credit-fund world. The CEO of our data partner — someone who's run distressed bridge loans for a decade and counts as one of the most experienced operators in this segment in the country — was unfiltered when we sat down to talk about it. "I believe my problem is the biggest problem of the industry." And "We're all eating the same shit." Recovery isn't elegant, it's a multi-month, multi-vendor, multi-conversation slog where the lender is simultaneously running borrower outreach, property-condition assessments, title work, broker price opinions, foreclosure-attorney coordination, and in the bridge world sometimes finishing construction themselves.
The economics are brutal. A non-performing loan costs roughly three to four times the human labor of a performing loan, recovery vendors charge accordingly, and they do it manually and miss critical signals because the data lives across team WhatsApps, customer email threads, broker phone calls, and county-clerk PDFs — none of it normalized, none of it queryable, none of it visible to the analyst making the decision. The same operator told us — "Recovery companies do it manually, so they cost a fortune, and they miss a lot."
The ground truth, from that operator's ten-year ledger of funded outcomes, is that the difference between a 30% recovery and an 80%+ recovery on the same kind of loan is almost entirely the quality of the operation, not the underlying collateral. Same property type, same geography, same loss event — the operator's playbook moves the recovery curve more than any other variable in the model.
What separates a great recovery operation from a bad one
Three things, in order of impact.
They know where the loan actually is on day one, not day sixty. Bad operations find out about default when the payment is missed, while great operations have continuous monitoring on the loan, the property, the borrower, and the title chain, and they see deterioration months before the missed payment. Property-tax delinquency posts to county records weeks before the borrower starts skipping mortgage payments, insurance-lapse notifications are public record, permit filings show whether a rehab is progressing or has stalled, and none of this is hidden — it's just unmonitored. The cheapest dollar in recovery is the dollar you spend before the borrower defaults.
They unify their data instead of letting it fragment across vendors. A bad recovery shop has the borrower-outreach log in one CRM, the title work in a vendor portal, the broker price opinions in a Dropbox, the foreclosure attorney's filings in email, and the property-condition photos on someone's phone — and the loss-mitigation analyst making the decision is looking at the seven percent of the picture that happens to be in their tab. A great shop has all of it in one operating model with one chain of custody, so when an analyst pulls a loan they see the full picture, the borrower's full history of contact attempts, the full title chain, the full set of valuations, and the full set of public-record events that have hit the asset.
They rank severity instead of treating everything as equal. A common failure mode I see in legacy systems is that every loan with a missing document or a stale field gets flagged yellow, and the result is that the entire screen is yellow and nothing is actionable. The transaction lead at the institutional servicer I described earlier said it out loud — "I need a little bit more guidance on the specific recommendations, critical versus medium." Great recovery operations triage, bad ones drown, and the difference is whether the system was designed to surface signal or just to surface volume.
What the numbers actually look like
The honest answer is that recovery numbers vary wildly by segment, geography, and operator quality, and most of the public benchmarks are stale, vendor-published, or both. Here's what I can speak to from primary data.
On residential bridge and fix-and-flip, our data partner's funded-outcome dataset — 1,500 labeled outcomes across roughly 11,500 loans over ten years — shows recovery curves that are heavily bimodal. Loans where the operator catches deterioration early and intervenes recover something in the 80-to-95-cent range on the dollar net of costs, while loans where the deterioration was missed until the borrower had already abandoned the rehab recover in the 40-to-60-cent range. The difference, again, isn't the asset. It's the operation.
On the cost side, non-performing loans cost roughly three to four times the operational labor of performing loans, which lines up with the consumer-credit dynamics I saw in a prior life. On the diligence-cost side, institutional buyers running third-party-review work today are paying two-to-five-hundred dollars per loan for what is essentially a labor-intensive document review with twenty-to-thirty-percent margins for the vendor, and block-trade diligence on a $500M tape can be a $15M exercise before the buyer even closes.
Timeline on judicial foreclosure varies state by state — Florida and New York can stretch eighteen-to-twenty-four months, while Texas and Georgia can close in four-to-six — and the longer the timeline runs, the bigger the recovery gap grows between great and bad operations, because time is the enemy of recovery in distressed assets.
The three diligence mistakes that turn defaults into surprises
The first is trusting the tape over the documents. Buyers receive a loan tape from the originator, do statistical sampling against the underlying file, and call it diligence — but the tape and the documents diverge constantly, FICO scores are stale, LTVs are computed against original appraisals from two years ago, vesting is listed wrong because the title changed and nobody updated the tape, and liquidity numbers are pulled from a stale page of the borrower's bank statement. The $1.5M liquidity discrepancy from earlier in this post is a textbook example. You cannot validate the tape from the tape. You have to validate the tape against the underlying chain of evidence.
The second is treating diligence as a one-time gate instead of a continuous practice. A buyer runs diligence at acquisition, signs off, and then never re-runs it, and six months later the borrower's circumstances have shifted, a tax lien has posted, an insurance policy has lapsed, and the buyer's risk model is now stale. Best practice in any large book is to re-diligence on a rolling basis — at minimum, refresh public-record enrichment quarterly on the entire portfolio.
The third is leaning on third-party-review firms as the safety net when those firms themselves miss things constantly. One of the institutional credit investors we work with put this directly: if you can show that an asset was incorrectly underwritten by a TPR firm, you're weeding out risk those firms charged hundreds of dollars per loan to miss. The TPR isn't a backstop. It's another input, and it should be validated like any other input.
What AI changes about all of this
The thing AI changes is not the analysis — humans were already capable of finding the discrepancies, the pre-foreclosures, the lapses, and the title clouds. The thing AI changes is the unit economics of doing the work at scale. A trained credit analyst can review fifty loans a day carefully, while an AI analyst with the right tooling can review fifty thousand, with the human focused on the few percent of loans that actually need judgment — and that's an order-of-magnitude shift in how much of a portfolio you can actually keep an eye on, which is the single largest change to credit operations in my career.
The second shift — and this is where I think most commentary gets it wrong — is that the moat in AI-driven credit is not the model. The model is a commoditized input that gets cheaper and better every quarter. The moat is data gravity — having a cross-counterparty view that a single institution can't replicate. A bank that builds its own AI on its own data will always be limited to what its own book has shown, while a platform that sees loans across many originators, many buyers, and many servicers, with years of labeled outcome data behind it, sees patterns that no single institution can see. You can build the parser. You can't build the dataset.
The third — controversial, but I'll say it anyway — is that the entire operation collapses onto a single layer, not a stack of disconnected vendors. Today you have one set of vendors doing diligence, a separate set doing valuation, a separate set doing servicing, and a separate set doing recovery, and each one looks at one slice of the loan and hands off to the next — so the context dies at every handoff. The next decade is one operating model across the entire credit lifecycle: a single source of truth with chain of custody intact end-to-end, and autonomous agents running each function on top of it. The funds that run on that layer, instead of stitching together a vendor per function, are the ones that compound.
The takeaway
When a borrower defaults on a mortgage note, the formal process — the demand letter, the notice of intent, the foreclosure timeline — is the visible part, and it's also the part that legal counsel and the state statutes will walk you through perfectly well without any help from me. The part that actually determines whether you make money on the loan or lose money on the loan happens before the formal process starts, and it's about whether the people holding the asset have an operating model that can see deterioration in time to intervene. The buyers and servicers who win in the next decade are the ones who treat default not as a discrete event to react to but as a continuous probability that they monitor in real time across every loan in the book, with one chain of custody, one view of the borrower and the property, and the analyst's attention focused on the few percent of loans where judgment actually matters.
That's the operating layer we build at Fundable, and if you're running a credit book where any of the patterns in this post sound familiar — the missed pre-foreclosure, the stale tape, the fragmented recovery data, the analyst at 10pm — that's the conversation worth having.