Fifty Years of Mortgage Technology, and What Each Wave Left for the Reviewer
Mortgage technology has been through five distinct waves since the 1970s. Each was substantial, each delivered real gains, and each left the same residue: the loan file still has to be checked by someone, against rules that keep changing, and the check still has to be defensible years later.
The pattern is worth tracing, because it predicts which claims about the current wave will age well.
Wave 0: standardisation was the first technology
Before any system, there was a form.
When the secondary market developed, loans had to be comparable to be sold. That required agreement on what a loan file contained, which produced the standardised forms still in use: the uniform residential loan application, the uniform appraisal report, the uniform underwriting and transmittal summary.
This was not a computing advance and it was the most consequential change in the list. Every subsequent technology operates on the structure standardisation created. A system can extract a 1003 because there is a 1003.
What it automated: nothing. What it enabled: everything after it.
Wave 1: mainframe servicing, 1970s–80s
The first computerisation targeted servicing rather than origination, for a straightforward reason: servicing is arithmetic performed monthly at scale: payments, escrow, interest, delinquency. It suited the machines of the period.
Origination stayed on paper, because origination is not arithmetic. It is the assembly and assessment of evidence.
The pattern begins here: technology arrives first where the work is computable, and stops at the boundary of judgement.
Wave 2: automated underwriting, mid-1990s
Fannie Mae launched Desktop Underwriter in 1995, with Freddie Mac's Loan Prospector following in the same period. [1] These were rule engines wired to eligibility guidelines with statistical default models embedded.
This was the genuine turning point, and it is widely misremembered. Automated underwriting did not automate the underwriter's judgement. It automated eligibility determination given these values, does this loan meet these guidelines, and issued a recommendation together with conditions.
The conditions are the point. The system determined what still needed to be proven, and a human proved it. The judgement moved rather than disappeared: from "does this qualify" to "is this documentation adequate."
What it automated: rule evaluation against supplied values. What it did not: whether the values were right.
Wave 3: loan origination systems and imaging, 1990s–2000s
Origination moved into workflow platforms. Documents were scanned, indexed and routed. Paper files became electronic files.
The gain was logistical and large: no lost folders, parallel work, visible status. But an imaged document is a picture of a document. The information inside it remained unavailable to the system holding it, so the reading work was unchanged and the checking work was unchanged. Both were simply performed against a screen.
What it automated: movement and storage. What it did not: reading, or checking.
Wave 4: compliance as software, 2010s
The post-crisis regulatory build-out made compliance a systems problem rather than a procedural one.
The ATR/QM rule took effect on 10 January 2014, requiring a documented, verified determination of ability to repay. [2] TRID followed on 3 October 2015, delayed from August by the CFPB, replacing the Good Faith Estimate, HUD-1 and Truth in Lending disclosures with the Loan Estimate and Closing Disclosure, and imposing content and timing requirements. [3]
Timing requirements are what made this a software problem. A rule about when a disclosure was received cannot be satisfied by a careful reader looking at the document; it requires a system that records events and compares dates. Compliance engines, disclosure platforms and audit trails followed necessarily.
What it automated: verifying that defined, checkable events occurred in a required order. What it did not: whether the substance underlying them was correct.
Notably, this wave produced the industry's first real infrastructure for evidence records kept specifically to demonstrate compliance later. That is the thread the current wave should be picking up.
Wave 5: machine reading, 2015 onward
OCR became extraction; extraction became machine learning; machine learning became vision-language models that read documents in something close to the way a person does, including formats they were not specifically built for.
This is a genuine advance and it closes the gap Wave 3 left open. A document is no longer a picture. Its contents are available to the system.
But note precisely what has been solved. The system can now read the file. Whether the file is sound whether the four income documents agree, whether the disclosure arrived in time, whether the programme's requirements were met: is a different question, and reading is a prerequisite for asking it rather than an answer to it.
What it automated: reading. What it did not: deciding.
The pattern, and what it predicts
Five waves, one shape, with a qualification the neat version of this argument leaves out. It is not true that fifty years of technology automated everything except judgement. Automated underwriting embedded statistical default models and issued eligibility decisions; that is judgement of a kind, delegated deliberately. Compliance platforms do more than compare dates. The honest pattern is narrower: each wave automated a layer that could be specified, and each left behind the part that required assembling evidence and defending the assembly later.
Two things explain why, and neither is a limitation of the technology.
Mortgage decisions must be defensible against a standard that has since changed. A file reviewed today may be examined in three years, when the guideline has been revised. Defending it requires showing which standard was applied at the time, which means retaining the inputs, the rule version and the configuration, not merely naming the system that produced the answer.
The consequence of being wrong is asymmetric and delayed. A repurchase demand carries an estimated average cost of $32,288 and arrives long after the decision. [4] Feedback that slow and that expensive favours mechanisms that are reproducible over mechanisms that are merely accurate on average.
This is why every wave has automated the determination and left the judgement, and why the useful question about any new mortgage technology is not how capable it is but whether its output can be reconstructed later.
Where that leaves the current wave
Machine reading completes Wave 3's unfinished business and hands the result to Wave 2's rule engines, which never went away and were always the defensible part.
That framing predicts the durable version of current AI in mortgage: a model that reads documents, feeding a versioned rule set that decides. Not because models are unreliable, and not because a model cannot be versioned. It can; a model identifier, snapshot and settings are all recordable. The reason is narrower. Recording which model ran does not let you re-derive what it concluded, and a decision defended years later has to be re-derivable from a retained record.
The waves that lasted are the ones that made something reproducible. That is the test to apply to this one.
Sources
- Fannie Mae launched Desktop Underwriter in 1995; Freddie Mac's Loan Prospector (now Loan Product Advisor) followed in the same period. Published sources vary between 1995 and 1997 for the pair, and this review did not settle the discrepancy.
- CFPB, Ability-to-Repay and Qualified Mortgage Rule, generally effective 10 January 2014, Regulation Z, 12 C.F.R. § 1026.43. The ability-to-repay requirement applies to consumer credit secured by a dwelling; § 1026.43(a) excludes extensions of credit primarily for a business, commercial or agricultural purpose.
- CFPB final rule of 21 July 2015 moving the TRID effective date to 3 October 2015, applying to applications received on or after that date. Regulation Z, 12 C.F.R. § 1026.19. Delivery and timing requirements for the Loan Estimate and Closing Disclosure; the corrected-disclosure waiting-period triggers are at § 1026.19(f)(2)(ii), which include the addition of a prepayment penalty.
- STRATMOR Group, Unpacking the drivers and costs of GSE repurchase demands, reported by National Mortgage News. The $32,288 figure is a study estimate of cost per repurchase demand, not per completed repurchase; income and appraisal together account for 57% of demands in that study. A category share does not establish that those demands were preventable.