Mortgage AI & Automation Software

The rules arrived from an unexpected direction

Anyone waiting for a federal regulator to set the terms for AI in mortgage lending has been watching the wrong door. Through 2025 and into 2026 the CFPB went the other way, withdrawing 67 guidance documents in May 2025 including both circulars covering adverse action from algorithmic and complex-model decisions, then finalising a rule that eliminates disparate-impact liability under ECOA effective 21 July 2026.

The binding requirements came from the GSEs instead, and they are now live. Freddie Mac issued Bulletin 2025-16 in December 2025, effective 3 March 2026, and it is prescriptive: enterprise-wide controls to map, measure and manage AI risk, documented roles and escalation paths, internal and external audit capability, alignment to a recognised framework, monitoring of model performance and security events, senior-management approval, and an express indemnification obligation. Fannie Mae followed with Lender Letter LL-2026-04, issued in April and effective 6 August 2026. It is principles-based rather than prescriptive, requiring written policies covering the full lifecycle of any AI or machine learning system, reviewed at least annually, with a designated owner.

One clause in the Fannie letter changes how you should read every vendor conversation. Vendors and subcontractors must be held to the same governance standards required of the seller or servicer. You cannot buy your way out of this obligation, and a vendor's SOC 2 report does not discharge it. Practitioner consensus is to build to Freddie's stricter standard, which then satisfies Fannie.

So the first question in any AI evaluation is not what the product does. It is whether buying it leaves you able to answer, with documentation, which AI tools touched a specific loan file, who used them and when, what data went into the prompts including borrower personal information, and what safeguards were active during that interaction. If the vendor cannot help you produce those answers, the product is unbuyable regardless of how well it performs.

The fair lending trap in this year's deregulation

This one is worth putting in front of your general counsel before it comes up in an examination.

The CFPB's Regulation B rule removes disparate-impact liability under ECOA. It does not touch the Fair Housing Act, where disparate impact survived the Supreme Court's 2015 decision in Inclusive Communities and where enforcement sits with HUD. Residential mortgage lending is squarely covered by the Fair Housing Act. A lender who reads the ECOA change as clearing disparate-impact exposure on an underwriting or pricing model has drawn exactly the wrong conclusion, and it is a conclusion several vendors will be happy to let you draw.

Two other things survived the withdrawal. ECOA still requires specific principal reasons for adverse action, because withdrawing a circular does not repeal a statute. And proxy discrimination remains illegal.

Meanwhile the states are moving in the opposite direction from the federal government. California's automated decision-making rules under the CCPA took effect on 1 January 2026. New York's FAIR Business Practices Act took effect on 17 February 2026, adding unfair and abusive standards to state law with attorney-general enforcement, and the New York AG has named mortgage lenders as an enforcement priority while the statute explicitly references emerging technologies. Colorado's AI Act, which treats mortgage lending as a consequential decision and imposes annual impact assessments plus an obligation to explain the degree to which an AI system contributed to an adverse decision, has been delayed once and is worth checking the current status of before you assume a date.

What the technology demonstrably does, and where the evidence stops

The adoption picture is a wide gap. A trade survey fielded in June 2026 found 83% of lenders evaluating AI and 17% running it in live production workflows. The rest are piloting or researching. Trust was the most-cited barrier, ahead of cost and internal expertise.

The best-evidenced use case in the industry right now is not underwriting. It is servicing voice, and the evidence is unusually good because it appears in a public company's disclosures rather than a press release. Rocket reported handling more than a million inbound servicing calls with an AI voice agent within three months of launch, with more than half of those calls otherwise requiring a servicing team member, task resolution around 25% faster than the previous IVR, and customer satisfaction of 4.5 out of 5. Better has reported roughly 100,000 calls a month with 35.5% of borrower inquiries handled end to end.

Origination productivity claims are also lender-disclosed rather than vendor-supplied, which makes them worth more. Rocket has reported closings per production team member up roughly 74% between March 2024 and March 2026, and around $300 billion of origination capacity with several hundred fewer production staff than in 2024. Read that carefully though: it is operating leverage harvested into a volume recovery rather than cost taken out, and it only pays if the volume shows up.

Document classification is the most widely deployed use case, with adoption more than doubling between 2023 and 2024. It is also the one with the weakest published results. No lender has published field-level accuracy or exception rates. Every accuracy figure in this category traces back to a vendor.

For condition clearing, the picture is thinner still. One vendor put more than 25,000 loans through an agentic pilot and then explicitly declined to disclose the performance metrics. A vendor that has run tens of thousands of production loans and will not quantify the outcome has told you something.

STRATMOR's summary of the failure mode deserves repeating in full, because it predicts most disappointing deployments: AI does not fix broken processes, it quickly exposes them.

Accuracy, and the one benchmark that exists

There is no independent test of mortgage document extraction accuracy in the public record. Vendors publish figures of 98%, 99% and higher without disclosing document mix, whether the measurement is field-level or document-level, whether confidence thresholding was applied, or whether the number is before or after human review. Those four omissions can move a headline accuracy figure by twenty points.

The nearest thing to a benchmark is a study published in March 2026 comparing a mortgage-specific system against a general-purpose frontier model across 90 questions on ten synthetic borrower scenarios involving payroll mismatches, undisclosed liabilities and suspicious deposits. It was commissioned by one of the vendors, so discount it accordingly. What makes it worth reading anyway is that the sponsoring vendor's own system scored 84%.

Roughly one answer in six was wrong or partly wrong, on synthetic scenarios, in a test the vendor paid for and published. The general-purpose model also beat the specialised one on account verification. Anyone selling you 99% should be asked to reproduce that methodology.

Then there is drift, which is the accuracy you lose after you buy. Academic work found 91% of deployed machine learning models exhibited some degree of drift, and more recent research put generative model drift between 76% and 89%. The buying implication is a contract term rather than a feature: written advance notice of model changes with release notes, and inference-level logging retained in a machine-readable format.

Where the money actually goes

Two structural points before any pricing conversation.

The first is that a fully automated loan is legally blocked, not merely difficult. The SAFE Act defines loan originators as individuals, so an AI cannot hold an MLO licence, and TILA requires a human originator's name and NMLS ID on the application and the loan documents. Naming an originator with no actual involvement risks a deception finding. Whatever automation rate a vendor quotes, a licensed human remains in the file.

The second is that the residual is adversely selected. A system that handles 70% of files and routes 30% to people has not left you with a representative 30%. It has left you the hard tail: the self-employed borrower with a business structure that changed between tax years, the file with the deposit nobody can source. Staffing, training and quality control for that population do not scale down in proportion to the automation rate. Ask any vendor to model cost per exception after deployment rather than automation rate, and treat an inability to do so as a finding.

Fannie's own income calculator illustrates the boundary precisely. It offers representation and warranty relief on the calculation, and the exception population is defined by policy rather than by model confidence: fewer than twelve months of earnings, an unknown employment start date, a business structure that changed between years, returns more than three years old. All of those fall out to a human. And the relief covers the arithmetic only. Lenders remain responsible for the integrity of the data provided, which means the risk that the system read the wrong number off the document stays with you.

On price, be honest with yourself about what is knowable. No mortgage AI vendor publishes list pricing, and there is no verifiable per-loan contract price from any lender or vendor in the public record. Minimum commitments, floor and ceiling structures, escalator rates and implementation fees are not public. Peer lenders are your only real source.

Be careful what you find when you search. At least one authoritative-looking pricing page circulating in 2026 cites source documents that do not exist, including a segment disclosure and an industry forecast dated in the future. Its numbers appear to be fabricated.

What is knowable is the cost base underneath consumption pricing. One vendor has published that running lending document workflows on a managed general-purpose model costs around $0.30 per document against roughly $0.03 on a purpose-built model. A vendor that has not optimised its model stack is passing a tenfold cost difference through to you somewhere.

And watch the adjacent per-transaction costs that AI can quietly multiply. Employment verification runs about $66 per pull at the dominant provider, and a two-borrower file verified at underwriting and again before closing has been reported reaching $280. Any automation that increases verification frequency multiplies a cost that dwarfs most per-document AI pricing. Against an MBA-reported production expense of $11,898 per loan and pre-tax production profit of $727, that arithmetic decides the business case faster than the software licence does.

Will the vendor still be there

This category is younger and thinner-capitalised than anything else you buy, and the contract you are signing outlasts the runway of a good number of the companies selling into it.

Consider the scale mismatch. Recent funding rounds in mortgage AI include a $4.1 million seed, a $6 million raise, and a $30 million Series A. A mid-size independent mortgage bank in the MBA's first-quarter 2026 sample averaged 1,729 loans and $621 million of production per quarter. A vendor with $4 to $6 million raised has perhaps eighteen to twenty-four months of runway against a system you expect to depend on for five years.

Even survival is not the same as stability. One well-known mortgage AI vendor raised a $12.5 million Series A, conducted layoffs about eighteen months later, and is shipping product today with fresh GSE integrations. That story ends well. The instructive part is that a vendor can halve its staff mid-contract and you will feel it in support responsiveness and roadmap velocity long before you would ever see a bankruptcy filing.

Gartner's estimate that only around 130 vendors of the thousands claiming agentic AI are genuine, and its prediction that more than 40% of agentic AI projects will be cancelled by the end of 2027 on cost, unclear value and inadequate risk controls, is a base rate worth carrying into every one of these evaluations.

The gravitational reality underneath all of it: roughly 90% of US mortgages touch ICE's network. Sell-side research this year concluded AI strengthens large incumbents rather than disrupting the industry, and that smaller institutions generally lack the budgets, talent and structured proprietary data to compete on it. A point solution that survives is reasonably likely to survive inside a larger platform, which makes your change-of-control and data-return clauses more important than your feature checklist.

Two tests worth more than a demo

Everything above collapses into a pair of things you can actually do in an evaluation.

Run the same file twice, and ask to see both outputs side by side. Non-determinism is the defining property of a language-model agent and the defining problem for a representation and warranty framework. A rules engine returns identical output every time. A genuine agent may not. Either answer tells you something useful. What matters is whether the vendor knows which one their product is, and whether they can bound the variance. A vendor who has never run this test on their own product is not ready to be in your loan file.

Ask for the decision trace on one specific condition, in business language. Not model confidence scores. Which policy rule, which fact from which document, which workflow signal produced this outcome. If the answer amounts to the model deciding, you cannot generate a specific principal reason for adverse action, you cannot defend the decision to an examiner, and you cannot satisfy the governance framework your GSE now requires you to operate under.

A process distributed across four or five vendors whose outputs cannot be traced is not defensible under a rep and warrant framework. That is the sentence to keep in mind when a stack of individually reasonable point solutions starts assembling itself across your origination flow.