Industries / Finance and business services / Finance and banking / Mortgage lending
Underwriting
This is deciding whether a home loan meets the lender's rules and those of whoever will later buy it, by weighing the borrower's credit record, income against debts and the loan's size against the home's value, and doing so fairly.
7
Milestones achieved, from 1938 to 2021.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1938
Redlining written into rules
Between 1935 and 1940 the Home Owners' Loan Corporation (HOLC), a federal agency, drew "residential security" maps for 239 US cities that graded neighbourhoods from A, the safest place to lend, to D, shaded red and labelled "hazardous", with race an explicit and often decisive factor. The Federal Housing Administration's 1938 Underwriting Manual warned against the "infiltration of inharmonious racial groups" and recommended deed covenants limiting who could live in a home, and the Fair Housing Act of 1968 outlawed racially motivated redlining.
LessonAreas on the lower-graded side of map borders had lower home ownership, house values and rents and more racial segregation for decades, so an underwriting rule can shape a neighbourhood long after the rule is gone.
Credit
Home Owners' Loan Corporation; Federal Housing Administration; later study by Daniel Aaronson, Daniel Hartley and Bhashkar Mazumder (Federal Reserve Bank of Chicago)
1974
Credit law bans discrimination
The US Equal Credit Opportunity Act (ECOA) of 1974 barred lenders from discriminating by sex or marital status, and its 1976 reenactment added race, colour, religion, national origin, age and income from public assistance. Its rules, Regulation B, require a lender that refuses credit to give the "specific" principal reasons; saying only that the applicant missed the lender's internal standards or its scoring cutoff is not enough. In a rule in force from 21 July 2026 the Consumer Financial Protection Bureau (CFPB) states that ECOA "does not authorize disparate-impact liability", that is, liability for a neutral practice that falls harder on a protected group; intentional discrimination is still banned, and the Fair Housing Act is a separate law.
Credit
US Congress; CFPB
1992
Public data tests fairness
The Home Mortgage Disclosure Act (HMDA) of 1975 made lenders report their home lending, and 1989 amendments added applications and the applicant's race, sex and income; the data showed Black and Hispanic applicants denied two to three times as often as white applicants. For a Federal Reserve Bank of Boston study, lenders supplied the financial, employment and property details behind their 1990 Boston-area decisions, and adding them cut the minority-to-white denial ratio from 2.7 to 1 to about 1.6 to 1: minority applicants with the same characteristics as white applicants would still have been denied 17% of the time, against 11% for white applicants.
LessonRaw denial gaps settle little until the lender's own decision factors are added, and here a gap remained after they were.
Credit
Alicia Munnell, Lynn Browne, James McEneaney and Geoffrey Tootell, Federal Reserve Bank of Boston
1995
Software and scores decide
In July 1995 Freddie Mac asked lenders to add credit bureau scores to their review of borrowers' credit histories, and that year it released Loan Prospector, automated underwriting software built on statistical study of how past loans had performed; Fannie Mae's Desktop Underwriter followed on a similar timeline, its early versions mostly automating Fannie Mae's existing manual rules. Lenders reported Loan Prospector accepting debt-to-income ratios, monthly debt payments divided by gross monthly income, as high as 72%, when manual guidelines usually capped them below 36%. In 2023 about two-thirds of US mortgages were underwritten through the two companies' systems, by the count of Stephanie Johnson (Rice University) and Nitzan Tzur-Ilan (Federal Reserve Bank of Dallas).
LessonJohnson and Tzur-Ilan find that the statistical rules let households borrow more and pushed up house prices, so the rules inside underwriting software change what people can borrow, not only how fast loans are processed.
Credit
Freddie Mac; Fannie Mae
2014
Limits on borrowing to income
In June 2014 the Bank of England's Financial Policy Committee recommended that UK lenders make no more than 15% of their new mortgages at 4.5 times the borrower's income or more, and test whether borrowers could still pay if Bank Rate rose 3 percentage points within five years. Ireland's central bank brought in its own mortgage measures in 2015; today they let first-time buyers borrow up to 4 times gross income and other buyers 3.5 times, with a 10% minimum deposit. From 28 October 2026 UK rules drop the 15% cap for each lender and keep it only for the market as a whole.
Credit
Bank of England Financial Policy Committee and Prudential Regulation Authority; Central Bank of Ireland
2018
Canada tests a higher rate
From 1 January 2018 Canada's bank regulator, the Office of the Superintendent of Financial Institutions (OSFI), expected federally regulated lenders to qualify uninsured borrowers at the greater of their contract rate plus 2 percentage points or the Bank of Canada's five-year benchmark rate. OSFI now calls this minimum qualifying rate "a mortgage stress test" and sets it at the greater of the contract rate plus 2% or 5.25%, for most newly underwritten mortgages.
Credit
OSFI, Guideline B-20
2021
Price test replaces debt cap
Under the CFPB's ability-to-repay rule of 10 January 2014 (see Applications), a "General Qualified Mortgage", the class of loan that gives a lender legal protection, had to keep the debt-to-income ratio at or below 43%, with income and debts counted by the rule's Appendix Q. A rule in force from 1 March 2021 replaced the cap with a price test: the loan's annual percentage rate must be less than 2.25 percentage points above the average prime offer rate, a published benchmark for similar loans.
Credit
CFPB
Watch for
Underwriting in two minutes
Better says its Tinman Credit Decision Engine, offered inside ChatGPT Enterprise, takes a median of 2 minutes 24 seconds to underwrite a typical loan and as little as 47 seconds for conventional loans to salaried borrowers, with licensed underwriters reviewing flagged differences. It also reports a defect rate of 0.06% against an industry average of 1.8%.
Achieved whenAn outside party reports the times and the defect rate on a stated number of loans, for example through a Fannie Mae, Freddie Mac or auditor's quality review.
Claimed by
Vishal Garg, Better, March 2026
Income checks in minutes
TD Bank Group says an AI agent built by its Layer 6 unit scans client documents, calculates and verifies income, checks it against policy and writes a memo for underwriters, cutting work that averaged 15 hours per client to about three minutes in early results, and that "100% of new personal mortgages" go through it.
Achieved whenAn outside party, such as an auditor or OSFI, reports the time saved and how often the memos' income figures match underwriters' own on a stated number of files.
Claimed by
TD Bank Group, July 2026
A public loan-agent test
MortarBench gives AI agents 60 days of invented bank transactions and a standard loan application file and asks underwriting questions, such as which deposits need documenting and whether there are undisclosed debts; the best closed model got 77.1% exactly right, and the authors' own method raised that to 80.5%. Its code and data are published.
Achieved whenA team not involved in building it reruns the benchmark and reports similar scores.
Claimed by
Matthew Toles and colleagues at Columbia University and Tidalwave, a mortgage software firm, June 2026
Open
01
A debt-to-income answer key
A public set of made-up loan files with a worked answer key. Each file holds pay stubs, W-2s, tax returns with Schedule C (self-employment) and Schedule E (rental income) and a credit report listing several kinds of debt; the key works out qualifying monthly income, monthly debts and the debt-to-income ratio the way Fannie Mae's Selling Guide says. AI models are scored against it. Nothing is used on a real applicant. The files can come from the same generator as "Read the income papers" under Applications; this test scores the calculation, not the reading.
The briefCredit · unclaimed02
Check AI-written denial reasons
300 made-up denied applications, each decided by a simple scoring model whose weights are fixed and published, so the true main reasons for every denial are known exactly. AI models are asked to write the reasons for the adverse action notice, the letter telling an applicant why credit was refused, and are scored on whether the stated reasons match the true ones, whether they make up a reason, and whether they name a protected trait such as race or sex. No real decisions are involved.
The briefCredit · unclaimed
Open 01 · the brief
A debt-to-income answer key
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of made-up loan files with a worked answer key. Each file holds pay stubs, W-2s, tax returns with Schedule C (self-employment) and Schedule E (rental income) and a credit report listing several kinds of debt; the key works out qualifying monthly income, monthly debts and the debt-to-income ratio the way Fannie Mae's Selling Guide says. AI models are scored against it. Nothing is used on a real applicant. The files can come from the same generator as "Read the income papers" under Applications; this test scores the calculation, not the reading.
Why it is still open
Income and employment errors were the largest category of critical defects, errors serious enough to affect a loan's eligibility, in lenders' quality checks in the third quarter of 2025, at 27.24% of the total, by ACES Quality Management's count. A few points of error decide eligibility, because Fannie Mae caps the ratio at 50% for loans run through Desktop Underwriter and at 36% for manually underwritten loans, or 45% with strong credit and savings. None of MortarBench's question types asks for qualifying income, monthly debt or a ratio, and its baseline tells the model "Do not calculate totals yourself". Tidalwave's March 2026 test used 90 questions written by its own staff and did not publish them.
acesquality.comselling-guide.fanniemae.comarxiv.orghousingwire.com
Built with what exists
The public Fannie Mae Selling Guide; the IRS's public PDF forms, filled by a Python script with pypdf or ReportLab; MortarBench's open code as a model for the scoring harness; Claude, GPT and Gemini models and the open-weight Qwen3 as systems to test, with a small extraction model such as Meta's Muse Ember for reading the forms (see the Note "The jobs agents repeat").
First step
Write 20 files by hand covering salaried, hourly, self-employed and rental-income borrowers, work out their ratios from the guide, and run one model on them.
Done looks like
500 files covering at least 10 income types, among them salary, hourly pay with overtime, bonus, commission, self-employment, rental, pension and Social Security; two people work out each key separately and reconcile it; a published table gives, for at least three models, the share of files where the model's ratio is within 1 percentage point of the key and the share landing on the right side of the 36%, 45% and 50% limits.
Open 02 · the brief
Check AI-written denial reasons
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
300 made-up denied applications, each decided by a simple scoring model whose weights are fixed and published, so the true main reasons for every denial are known exactly. AI models are asked to write the reasons for the adverse action notice, the letter telling an applicant why credit was refused, and are scored on whether the stated reasons match the true ones, whether they make up a reason, and whether they name a protected trait such as race or sex. No real decisions are involved.
Why it is still open
Regulation B requires reasons that are "specific and indicate the principal reason(s)". On 12 May 2025 the CFPB withdrew its 2022 and 2023 circulars on giving reasons when complex algorithms are used, so lenders have less official guidance though the rule itself stands. Research by Miles Turpin and colleagues found that language-model explanations "can systematically misrepresent the true reason for a model's prediction". Searches found no published test of whether AI-written denial reasons match the reasons that drove a credit decision, only a proposal for one.
Built with what exists
scikit-learn for a logistic regression scoring model with eight factors; SHAP to rank each factor's contribution; Faker or the Synthetic Data Vault (SDV) to generate applicants; the models above as systems to test; grading by matching against reason codes, with rubric scoring by Gemini Decision 1 (see the Note "A thousand questions, one pass") and a person checking a sample.
First step
Build the scoring model and 30 denied files, and run one model on them.
Done looks like
300 files, each with its true top four reasons, and a published table giving, for at least three models, the share of notices whose reasons match at least three of the four, the share with a made-up reason and the share naming a protected trait; a person checks 50 graded notices and agrees with the automatic grade on at least 90%.