Every other corner of financial services now has an AI-native champion. But the single largest loan most Americans will ever sign is still manufactured on software older than the iPhone – and that gap is about to become the best trade in fintech.
Walk onto a sell-side floor today and a first-year analyst is building comps and drafting a CIM with an AI agent looking over their shoulder. Rogo, the model built for finance, just raised a $160 million Series D and says 35,000-plus professionals across 250-plus institutions – Lazard, Jefferies, Moelis, Nomura, Rothschild – now run on it. Private equity associates triage data rooms with Hebbia. Treasury, fraud, research: all of it has been colonized by remarkably accurate AI tooling.
Then there’s mortgage. A market that originates north of $2 trillion a year, the biggest piece of consumer credit in the country, is still built on a stack designed during the Clinton administration.
To be precise, “AI hasn’t touched mortgage” isn’t exactly accurate. AI has touched mortgage, but not at the core system layer. For instance, Rocket runs generative models over millions of documents a month; Better’s “Tinman,” Ocrolus, and Candor sell automation workflow tools; and Fannie Mae itself just stood up its own AI fraud unit with Palantir. But look at what every one of those tools has in common: it bolts onto a core it does not control. The system of record is still the loan origination system, and the credit decision is still rendered by automated underwriting engines shipped in the 1990s as rules-based black boxes.
So why did the most lucrative vertical in lending get stranded?
First, the 800 pound gorilla. One company sits in the middle of American mortgage: Intercontinental Exchange. ICE owns Encompass, the origination system that runs roughly half of all U.S. originations. It bought Encompass’s maker, Ellie Mae, in 2020, then paid $13.1 billion for Black Knight in 2023. The FTC sued to block that deal and forced ICE to divest the second-place origination system and the dominant pricing engine, warning that the combination would “drive up costs, reduce innovation, and limit lenders’ choices.” When the system of record belongs to a toll collector, outside AI doesn’t get to rebuild the road – it pays for access, enters through ICE’s on-ramp, and plays by their API rules. And a lender who hates the arrangement can’t simply leave. Ripping out an origination system is a multi-year root canal no risk-averse mortgage COO volunteers for.
Second, the rate shock froze the buyers at exactly the moment they should have been forced to modernize. The 30-year fixed bottomed at 2.65% in January 2021. It’s stuck in the mid-6s today, with no return to 5% in sight. Refinancing, the volume engine of the whole industry, is dead because nearly every homeowner is locked into a rate below where the market now sits. Origination cratered from a record $4.4 trillion in 2021 to under $1.8 trillion in 2024, with lending activity down almost 70% from the peak. When your top line collapses, you don’t sign six-figure software contracts. You cut headcount and freeze every tech purchase.
I believe the next 24 months will mark an inflection point for AI in the mortgage industry.
For starters, unit economics for mortgage originators remain broken; AI is now finally capable of saving costs in a way that finally makes the math impossible to ignore. It cost lenders an average of $11,230 to produce a single loan. Mortgage CFOs are left scrambling for ways to cut costs and generate efficiency amid an unpredictable macro rate environment. The CFO who waved off AI in 2023 is now the CFO with no other move.
On top of that, buyers have become solvent and tech-curious again. Production profitability in the second quarter of 2025 was the best since 2021. The Mortgage Bankers Association expects $2.2 trillion in single-family originations in 2026, up 8%, and its own analysts now say lenders are actively exploring how to cut origination costs through technology. By one industry survey, 83% of lenders plan to grow generative-AI budgets in 2026. Desperate and solvent at the same time is the precise condition that creates a software market.
Most importantly, the reasoning that underwriting once locked in a black box can now be generated on demand. Fannie Mae’s Desktop Underwriter has generated machine-rendered credit findings that examiners have treated as authoritative for decades; what’s new is that generative models can now produce the same structured rationale using more unstructured documents, like tax returns, bank statements, and agency guideline overlays. And the regulatory appetite for machine-rendered explainability is already established: the CFPB’s 2023 Circular on AI in adverse action notices confirmed that machine-generated explanations can satisfy ECOA and Regulation B requirements.
If the models are capable and the regulatory permission is emerging, why is the opportunity still unclaimed? Because every contender keeps attacking product surfaces that are walled. Try to own the credit decision and you run into the GSEs, who gatekeep every conforming loan through their own underwriting engines. Try to build a new system of record and you’re asking a risk-averse COO to put their job on the line (and to pay ICE’s integration tax for the privilege of attempting it). The obvious entry points are the most heavily defended ground in the industry.
The undefended surface is the work itself. If we break the $11k cost-to-originate into its parts, most of it isn’t software. It’s people. Humans control intake, document collection, income and asset verification, clearing conditions, quality control, post-closing: origination is, at its core, a labor business. And the industry has already conceded that this labor can sit on offshore processing floors in Manila and Bangalore. The only open question is whether the next generation of outsourcing is staffed by thousands of people costing millions per year, or by AI agents overseen by a handful of forward-deployed engineers.
The company that wins the mortgage AI race brings an operator, not a tool. It runs on whatever stack the lender already has. It doesn’t need agency approval because it prepares all the paperwork, but the approval/rejection decision stays with humans (which is what regulators want). And it gets paid the way a CFO actually wants to pay in 2026 – per closed loan, against outcomes, not in six-figure SaaS seats. This model is already working in other corners of financial services, like insurance and accounting. Mortgage has all the same ingredients – a high-cost, labor-heavy back office – and yet it remains the one corner of financial services where no one has shown up to do that work with AI.
Mortgage’s AI moment is finally here – the largest, slowest, most defended market in consumer lending, thawing at exactly the moment the technology is ready and the buyers are desperate enough to move. The window to do work from the bottom up is open, and it won’t stay that way for long.
