Industries / Finance and business services / Finance and banking / Mortgage lending
Applications
This is taking a borrower's request for a home loan and checking the papers behind it, such as the application form and the proof of income, savings, employment and identity, usually by a loan officer, the licensed person who takes the application.
7
Milestones achieved, from 2008 to 2021.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2008
Loan officers licensed and listed
State mortgage regulators launched the Nationwide Mortgage Licensing System and Registry (NMLS) in January 2008, and the SAFE Act, signed on 30 July 2008, requires every mortgage loan originator, the person who takes a borrower's application, to be licensed by a state or registered through NMLS. NMLS must show the public, free of charge, each originator's work history and any public disciplinary actions. The regulators counted 165,116 state-licensed originators holding 569,190 state licenses in 2019, and another 415,978 registered at 8,878 banks and other federally regulated deposit takers.
mortgage.nationwidelicensingsystem.orgmortgage.nationwidelicensingsystem.org
Credit
state mortgage regulators and the Conference of State Bank Supervisors (CSBS), which runs NMLS; US Congress
2014
US ends unchecked income
In the mid-2000s many US lenders made "low-doc" and "no-doc" loans, on which the applicant stated an income and nobody checked it; the economists McKinley Blackburn and Todd Vermilyea found income overstated by "almost 20 percent" on the average home-purchase application of 2005 and 2006. From 10 January 2014 the Consumer Financial Protection Bureau's (CFPB) ability-to-repay rule requires lenders to check what applicants state against "reasonably reliable third-party records", and a no-doc loan cannot be a qualified mortgage, the class of loan that meets the rule's standards.
LessonWhen nobody checks the number on a form, the number goes up; a JPMorgan Chase memo quoted in the paper told loan officers whose stated-income applications the software rejected to "Inch it up $500".
Credit
CFPB; Blackburn (University of South Carolina) and Vermilyea (Federal Reserve Bank of Philadelphia), who measured the problem
2014
UK ends self-certified mortgages
In a self-certified mortgage the borrower declared an income and the lender did not check it. The Financial Services Authority (FSA), then the UK regulator, found that around half of all UK mortgage applications in 2007 and 2008 were processed without checking income. Rules from its Mortgage Market Review took effect on 26 April 2014, and in May 2016 the Financial Conduct Authority (FCA) reviewed firms making about 75% of lending since then and found that income had been checked in every lending decision it looked at.
LessonThe FSA found that "some consumers were willing to overstate their income in order to obtain a mortgage".
Credit
FSA and FCA
2015
Applications move online
Quicken Loans introduced Rocket Mortgage in 2015; it collects a borrower's income, asset and credit records electronically and can approve a loan from an online application "in as little as eight minutes", in the words of Federal Reserve Bank of New York economists. Andreas Fuster, Matthew Plosser, Philipp Schnabl and James Vickery found that lenders with fully online applications grew from 2% to 8% of US mortgage lending between 2010 and 2016, and processed applications about 20% faster than others without higher defaults.
LessonA form that is easy to finish is also easy to leave half answered: online lenders had more applications with race or sex missing, because borrowers could finish without giving them.
Credit
Quicken Loans; measured by Fuster and colleagues (2018)
2016
Dutch payroll records replace letters
Dutch mortgage applicants used to bring a werkgeversverklaring, a salary statement signed by the employer, and figures from the broker De Hypotheker and the lender Florius showed it was handed in 2.4 times on average because details were missing or wrong. In a 2016 pilot called Handig!, applicants at 25 De Hypotheker branches instead downloaded their pay and job history from UWV, the agency that runs employee insurance, and ABN AMRO, ING and Rabobank had joined by March 2017. The method, now called Inkomensbepaling Loondienst (IBL), works out the income a lender tests from that record under published rules that lenders can also reach by an API, a link between programs.
Credit
IG&H Consulting & Interim, which began the project in 2014 by Consultancy.nl's account; HDN (the Dutch mortgage data network), NHG (the national mortgage guarantee), UWV and Vereniging Eigen Huis, by De Hypotheker's account
2016
Income checked from data
Fannie Mae opened a service to all lenders in December 2016, after a 2015 pilot, in which approved vendors supply data from banks or the Internal Revenue Service, with the borrower's consent, and Fannie Mae's software checks income, employment and assets against it; the lender is then spared some demands to buy the loan back over those items. FHFA's inspector general found that use grew from 3.8% of Fannie Mae's purchases in early 2017 to 13.8% in early 2019, and Freddie Mac started its own program in December 2018.
LessonThe checking moves to the data vendors; FHFA sees them as the main risk, and the inspector general had not tested how accurate their checks were.
Credit
Fannie Mae, which calls the set of tools Day 1 Certainty; measured by FHFA's Office of Inspector General
2021
A new standard application form
The Uniform Residential Loan Application (URLA) is the standard form for loans sold to Fannie Mae and Freddie Mac, their Form 1003 and Form 65. On 23 August 2016 the two companies published "the first substantial revisions made to the form in over 20 years", with plainer wording, revised questions on race and ethnicity, and a shared data standard, the Uniform Loan Application Dataset (ULAD), so that every lender sends the same fields. After several delays it became mandatory for all new loans sold to the two companies on 1 March 2021.
Credit
Fannie Mae and Freddie Mac
Watch for
A voice agent takes calls
Better.com's AI loan assistant, Betsy, runs on ElevenLabs' voice agent platform and does eligibility checks, pricing and rate locks over the phone, with licensed advisers stepping in "where required". ElevenLabs says it handles "nearly 100,000 monthly calls", automates 35.5% of mortgage inquiries end to end, and that Better's cost to make a loan fell 41% in 2025.
Achieved whenA regulator, auditor or independent study measures how often Betsy's answers, prices and locks are correct, and how much of an application it completes before a licensed person takes over.
Claimed by
ElevenLabs and Better, February 2026
Free fraud checks on papers
Ocrolus made its document-fraud tool, Ocrolus Detect, free for all its mortgage customers; it screens bank statements, pay stubs and tax forms for tampering. Ocrolus says 12% of the loan applications on its platform contain suspicious activity, out of 750,000 credit applications it analyzes a month.
Achieved whenAn independent test on documents with known edits reports Detect's catch rate and false-alarm rate, or a lender or regulator reports how many flagged files turned out to be fraud.
Claimed by
Ocrolus, October 2026
AI-assisted loan officers
Rocket Companies says its loan officers who use AI tools handle nearly 40% more clients than a year earlier.
Achieved whenAn outside count, such as an analyst's study of public loan records, shows loans per Rocket loan officer rising by about that much.
Claimed by
Rocket Companies, second-quarter 2026 results, August 2026, as reported by Pulse 2.0
Open
01
Read the income papers
A public test of AI reading a borrower's income papers as scans and photos. Make 500 invented borrower files, every page marked SPECIMEN, each with two pay stubs, two years of W-2s (the yearly US wage statement), a federal tax return with Schedule C business profit for the self-employed, two months of bank statements and a filled-in URLA, rendered both as clean PDFs and as photographed or scanned pages. The answer key gives every field, such as employer, gross pay, year-to-date pay, business profit and large deposits, and about ten checks across documents per file, such as whether the employer's name matches on the stub, the W-2 and the URLA.
The briefCredit · unclaimed02
Spot the altered pay stub
A public test set of invented US income and asset papers, 1,000 genuine and 1,000 altered. The edits include changed gross pay, a changed employer name, an inserted bank deposit, digits copied from another document and a page made wholesale by an AI model. Each altered page comes with a mask of the changed area and a note of what changed, as a native PDF and as a scan. Detectors are scored on catch rate and false alarms.
The briefCredit · unclaimed03
Score the phone application
A public test of AI voice agents that take home-loan applications by phone. Write 300 invented borrower profiles, each with a complete URLA answer key of at least 40 fields, and record calls with synthetic voices, some with accents, noise and callers correcting themselves. Score each agent on how many fields it records correctly, on wrong values it records without doubt, and on whether it hands over to a licensed person when the caller asks. No real application is taken.
The briefCredit · unclaimed
Open 01 · the brief
Read the income papers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI reading a borrower's income papers as scans and photos. Make 500 invented borrower files, every page marked SPECIMEN, each with two pay stubs, two years of W-2s (the yearly US wage statement), a federal tax return with Schedule C business profit for the self-employed, two months of bank statements and a filled-in URLA, rendered both as clean PDFs and as photographed or scanned pages. The answer key gives every field, such as employer, gross pay, year-to-date pay, business profit and large deposits, and about ten checks across documents per file, such as whether the employer's name matches on the stub, the W-2 and the URLA.
Why it is still open
MortarBench, the public benchmark for AI loan agents posted in June 2026, says firms use such agents "despite a lack of any public benchmark"; its 90 questions over 9 invented loan files use structured data only, with bank statements as data records and applications as ULAD files, and no pay stubs, W-2s or tax returns. RealDoc-Bench has a mortgage section, but "each item is a single-document question".
Built with what exists
Python with Faker for invented people and employers and ReportLab to draw pay stubs and bank statements; the IRS's fillable W-2 and 1040 PDFs, filled in with pypdf; Augraphy to make pages look printed, scanned or photographed; the ULAD schema from Freddie Mac's site. Systems to test: Claude, Gemini, and a small extraction model such as Meta's Muse Ember (see the Note "The jobs agents repeat"), with Gemini Decision 1 answering each file's yes-or-no cross-document checks in one pass (see the Note "A thousand questions, one pass").
First step
Write a generator for one pay-stub layout and the W-2, make 20 files with a planted mismatch in five of them, and score one model on them.
Done looks like
500 files with at least 10,000 labeled fields and 5,000 cross-document checks, under an open licence with a scoring script; scored runs of at least three systems, with field accuracy (amounts within $1) and check accuracy reported separately for clean PDFs and photos; a second person re-checks 50 files and agrees with at least 99% of the key.
Open 02 · the brief
Spot the altered pay stub
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of invented US income and asset papers, 1,000 genuine and 1,000 altered. The edits include changed gross pay, a changed employer name, an inserted bank deposit, digits copied from another document and a page made wholesale by an AI model. Each altered page comes with a mask of the changed area and a note of what changed, as a native PDF and as a scan. Detectors are scored on catch rate and false alarms.
Why it is still open
MortarBench's records carry no tampering labels. The best-known payslip forgery set, from La Rochelle University, has 477 altered French payslips and is available only by emailing the authors; the docverify project trains on synthetic French payslips, publishes no dataset, and warns that "every number is measured on synthetic payslips". Ocrolus reports 12% of applications as suspicious with no public test behind the figure.
Built with what exists
The generator from the item above; PyMuPDF to edit PDF text, and Pillow and OpenCV for image edits; docverify's Apache-2.0 tamper model as a baseline; a long-document reviewer model such as Claude for checking arithmetic and names (see the Note "Seven more flagged").
First step
Make 50 pay stubs with the generator, change the gross pay in a copy of each, and run docverify on all 100.
Done looks like
2,000 documents across three document types and at least five kinds of edit, each with a mask, and a public table for at least three detectors giving the share of edits caught at a 1% false-alarm rate, by kind of edit and by PDF against scan.
Open 03 · the brief
Score the phone application
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI voice agents that take home-loan applications by phone. Write 300 invented borrower profiles, each with a complete URLA answer key of at least 40 fields, and record calls with synthetic voices, some with accents, noise and callers correcting themselves. Score each agent on how many fields it records correctly, on wrong values it records without doubt, and on whether it hands over to a licensed person when the caller asks. No real application is taken.
Why it is still open
Better and ElevenLabs publish call volumes and automation shares, not how accurately fields are captured; MortarBench tests text and structured files, not speech; and the CFPB found in 2023 that "chatbots sometimes get the answer wrong". No public test of voice intake was found.
Built with what exists
Faker profiles; text-to-speech voices to play the callers; ElevenLabs Scribe v3 to transcribe calls for scoring (see the Note "Fewer word errors"); an agent built on ElevenLabs Agents, the platform Better uses, as one system to test; ULAD field names for the key; a scoring script in Python.
First step
Write 20 profiles with full URLA answers, voice them with two synthetic callers, run them through one agent and score the fields.
Done looks like
300 calls and their keys published, at least two agents scored on field accuracy, confident errors and hand-over rate, and a second person scoring 10% of calls with their agreement reported.