Mortgage AI got the document work. The judgment is where it's stuck.

A mid-market lender we talked to this spring had an AI tool that classified incoming loan documents with real accuracy. It sorted the W-2s from the bank statements from the 1003s, fast, at a fraction of the old cost. Then the file hit an underwriter’s desk and a human did every piece of judgment by hand, exactly like 2019.

That is the shape of mortgage AI right now. Per STRATMOR’s 2024 data, 63% of AI-using lenders run document classification and 54% run document reading. Underwriting decisions sit at 21%. AI got the easy work and stopped at the door of the hard work.

The reason it stopped is not that the models can’t read a loan file. It’s that nobody engineered the trust to let one near a decision that costs $32,288 when it’s wrong.

1. Where mortgage AI actually is

Mortgage AI lives in the document layer. Document classification and indexing at 63% of AI-using lenders, document reading at 54%, intranet help at 29%, and underwriting decisions at 21%, per STRATMOR’s 2024 Technology Insight data.

Adoption itself is real and fast. STRATMOR puts lender AI/ML use at 38% in 2024, up from 15% in 2023, a jump of more than two times in two years. The question stopped being whether shops adopt AI. They have. The question is where it landed, and it landed on the high-volume, low-judgment tasks first.

63% of AI-using lenders run a third-party vendor; only 17% use AI built into their LOS. STRATMOR Technology Insight data, 2024. That split explains the ceiling. The vendors ship the document layer because it’s the safe, sellable 60%, and the judgment work stays on the underwriter’s desk because no vendor product reaches it.

2. Why it stopped at the easy work

The 21% ceiling on underwriting AI is a trust limit, not a capability limit. The models can read the file. The cost of them reading it wrong is what holds them back.

Document classification is forgiving. Misfile a bank statement and someone re-sorts it in ten seconds. Underwriting judgment is not forgiving. A misread income figure or a wrong property-type call becomes a condition, an exception, a defect finding, or a repurchase. Per the MBA, the average cost to produce a loan ran about $10,965 in Q2 2025, and the average repurchase cost is $32,288. When a wrong answer is a five-figure event, the bar to let a probabilistic model make the call is high, and most vendors have not cleared it.

The demand is there. Fannie Mae’s lender survey shows 73% of lenders now name operational efficiency as their primary motivation for AI, up from 42% in 2018, while the share citing consumer experience collapsed from 41% to 7%. Lenders want AI in the workflow. They want it where a mistake is cheap, and they have not been given a reason to trust it where a mistake is not.

3. The part most vendors won’t say: AI shouldn’t make the call

Here is the uncomfortable position, and it’s the one we build to: the model probably should not make the underwriting decision at all. The frontier is not automating the judgment. It’s making the work around the judgment trustworthy enough that a human clears more files.

We shipped an eligibility engine for a multi-program home equity platform with zero LLMs in the decision path. The model never makes the call. A versioned rules table does, every decision writes one audit row, and a human reviews the edges. That is not AI-lite caution, the pattern where a vendor keeps AI safely away from anything that matters. It’s the opposite. It’s putting AI exactly where it’s reliable and keeping it out of where it isn’t.

So the splashy “AI underwrites your loans” demo is aimed at the wrong target. The boring, valuable target is “AI does the prep and the extraction so the underwriter clears 40% more files at the same defect rate.” One of those makes a good conference slide. The other one funds loans.

4. What it takes to ship into the judgment gap

Closing the judgment gap is a trust-engineering job, not a model-selection one. Four things move AI from the demo to the file.

First, confidence labeling with abstention. The model returns how sure it is, and below a set threshold it hands off to a human instead of guessing. A field that comes back at 0.62 confidence is a flag, not an answer.

Second, anti-fabrication rules. The model never invents a value it cannot find. A missing income figure returns as “not found,” never as a plausible-looking number. A confident wrong answer is more dangerous than a blank, because nobody double-checks it.

Third, a deterministic core. The decision logic, the eligibility rules, the pricing, the conditions, lives in versioned code a human can read and audit. The model feeds that core. It does not replace it.

Fourth, drift detection and structured review. Every prompt and model change runs a review pass before it ships, and accuracy is watched so a quiet regression gets caught before an underwriter does. We track false positives, the fields the model marked high-confidence and got wrong, and hold them to a hard ceiling. On production enrichment work that has run around 0.4% over the trailing window. That number is the whole game. It is the line between an underwriter who trusts the prep and one who re-checks every field by hand.

What broke

On an early enrichment build, the agent returned a borrower’s property type as single-family with high confidence. The document was a condo rider it had skimmed past. High confidence, wrong answer, which is the worst combination, because a confident wrong answer is the one a human waves through.

It surfaced in the false-positive audit two weeks later. We rebuilt the extractor to cite the source span for every field it returned, to abstain when the source was ambiguous, and to never raise its own confidence without a citation behind it. Confidence you can’t trace to a document is not confidence. It’s a guess wearing a number.

Where this goes

The lenders who win the next two years will not be the ones who put a model on the underwriting decision. They’ll be the ones who made the judgment-adjacent AI trustworthy enough that the underwriter leans on it, with the decision itself still deterministic and still auditable.

I’d bet a dinner that the first mid-market shop to quietly double its underwriter throughput does it with no AI anywhere near the credit decision. The model will read every file. A human and a rules engine will still make every call. That isn’t a limitation. That’s the design.

If you’re a mortgage originator and your AI stops at the document layer while the judgment still runs by hand, that’s the teardown. Book the origination teardown → (/book/teardown)