Industries / Finance and business services / Finance and banking / Mortgage lending
Servicing
This is running a mortgage after it is made: collecting the monthly payments, paying the home's property tax and insurance from an escrow account, answering borrowers, and working with borrowers who fall behind, up to foreclosure.
7
Milestones achieved, from 1933 to 2023.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1933
Refinancing a million defaults
In 1933 the US Congress created the Home Owners' Loan Corporation (HOLC), which bought home mortgages in default from lenders and rewrote them as 15-year loans paid down in equal monthly instalments at a fixed rate of no more than 5%. By June 1936 it had made just over one million loans worth $3.1 billion, and it "often counseled delinquent borrowers and readjusted payment schedules" before foreclosing; about 20% of its loans still ended in foreclosure or with the home handed over, by David Wheelock's 2008 paper for the Federal Reserve Bank of St. Louis.
LessonWheelock calls the HOLC "arguably" "highly successful", but it lent only to owners it judged able to pay, and nearly half its applications were rejected or withdrawn.
Credit
US Congress; Home Owners' Loan Corporation
2009
A modification program falls short
A loan modification permanently changes a loan's terms, for example by lowering the rate, so that the payment becomes affordable. The US Treasury's Home Affordable Modification Program (HAMP), announced in February 2009, paid servicers to modify loans and was projected to help 3 to 4 million borrowers; by November 2013 about 1.3 million had entered a permanent modification, which the Government Accountability Office (GAO) called "well below" that estimate. When HAMP closed to new applicants at the end of 2016, Treasury counted 1,735,141 permanent modifications started, of which 648,037 were later disqualified after three missed payments.
LessonThe size of the cut mattered most: after 24 months, by Treasury's count, 15% of modifications that cut the payment by more than half had been disqualified, against nearly 38% where the cut was 20% or less.
Credit
US Department of the Treasury; measured by GAO
2012
Robo-signing settled for $25 billion
On 9 February 2012 the US federal government and 49 state attorneys general announced a $25 billion agreement with the five largest servicers, Bank of America, JPMorgan Chase, Wells Fargo, Citigroup and Ally Financial, over violations that included "robo-signed" affidavits, sworn statements signed in volume without checking, in foreclosure cases. The servicers had to give at least $20 billion of relief to borrowers, assign each borrower a single point of contact and stop foreclosing while a homeowner was being considered for a modification. An independent monitor, Joseph A. Smith Jr., tested them on 33 measures using samples of loans and in his last tests, for the third quarter of 2015, found no failures at Bank of America, Chase, Citi, Ditech or Wells Fargo.
LessonForeclosing in volume without checking the paperwork cost the five servicers $25 billion and years of outside testing.
archives.hud.govmortgage-settlements-dept-nmsarchive.cloudapps.unc.edu
Credit
US Department of Justice; Department of Housing and Urban Development (HUD); 49 state attorneys general
2013
Spain's evictions meet EU law
In Spain a borrower whose home was being sold to repay a loan could not raise an unfair term in the contract in those proceedings. Mohamed Aziz's home passed to Catalunyacaixa at 50% of its value when no one bid at auction, and he was evicted on 20 January 2011; on 14 March 2013 the Court of Justice of the European Union ruled that the Spanish procedure breached the EU's Unfair Terms Directive. Spain's Law 1/2013 of 14 May 2013 let borrowers object to a sale because of unfair terms, capped default interest on loans for a main home at three times the legal interest rate, and suspended evictions of especially vulnerable families, a suspension later extended to fifteen years from the law's start.
LessonWithout the power to stop a sale, the Court said, "it would thus be sufficient for sellers or suppliers to initiate mortgage enforcement proceedings in order to deprive consumers of the protection intended by the directive.".
Credit
Court of Justice of the EU, on a question from Commercial Court No 3 of Barcelona; Spain's parliament
2014
Federal rules for servicers
The Consumer Financial Protection Bureau's (CFPB) servicing rules under Regulation X took effect on 10 January 2014. A servicer must try in good faith to speak with a delinquent borrower by the 36th day and tell them about loss mitigation, the alternatives to foreclosure such as a repayment plan or a modification; it must send a written notice by the 45th day; it may not start foreclosure until the loan is more than 120 days behind; and if it receives a complete application more than 37 days before a sale it must decide within 30 days and may not move toward a sale meanwhile, which ended "dual tracking", pursuing foreclosure while reviewing a modification.
LessonIn 2016 CFPB examiners called servicing technology "outdated and deficient", and in exams completed in 2023 they still found servicers that failed to try to reach hundreds of delinquent borrowers and failed to send written notices to thousands.
law.cornell.edulaw.cornell.eduwebharvest.govfiles.consumerfinance.gov
Credit
CFPB, under the Dodd-Frank Act
2020
Forbearance on request
Forbearance is a pause or cut in payments for a set time. Under the CARES Act, borrowers with federally backed mortgages had only to ask to be granted it, and private lenders took part too. The Federal Reserve Bank of Philadelphia's RADAR Group, working from Black Knight's loan data, estimates that 8.5 million borrowers entered forbearance in the pandemic's first year; by early 2023, 22.5% had caught up with a lump sum, 58% had caught up with help from their servicer, such as a repayment plan, a deferral of the missed payments or a modification, 11.7% had paid off their loans, and about 0.1% were in default.
LessonThe group concludes that the program "mitigated a default wave like that experienced during the Great Recession of 2008-09", with servicers' workouts giving the longer-term help.
Credit
US Congress; mortgage servicers; measured by the RADAR Group
2023
A charter for UK borrowers
In June 2023 the UK government agreed a Mortgage Charter with lenders covering about 90% of the mortgage market. Borrowers who are up to date can switch to interest-only payments for six months, or lengthen their term and switch back within six months, without a new affordability check or harm to their credit score, and no borrower is to be made to leave home without consent within a year of the first missed payment, except in exceptional circumstances. The Financial Conduct Authority (FCA) counts about 354,000 mortgages whose monthly payments were lowered this way between July 2023 and June 2026, from firms' own reports.
LessonThe charter warns that the "overall costs over the life of the mortgage will be higher.".
Credit
UK government; signatory lenders; uptake counted by the FCA
Watch for
AI answers servicing calls
Rocket Companies says its AI voice platform handled 65% of inbound servicing calls in the third quarter of 2026 and resolved more than half of them without staff. Rocket closed its $14.2 billion purchase of Mr. Cooper in October 2025 and on 30 September 2026 serviced 9.1 million loans worth $2.0 trillion.
Achieved whenA regulator, auditor or independent study reports how often delinquent borrowers on AI-handled calls are given the loss-mitigation information Regulation X requires, and how their complaint rates compare with calls handled by staff.
Claimed by
Rocket Companies, October 2026
Servicing rules rewritten
On 6 November 2026 the CFPB sent a final rule, "Streamlining Mortgage Servicing for Borrowers Experiencing Payment Difficulties", to the White House Office of Management and Budget for review; its text is not public. The July 2024 proposal it follows would let servicers start helping borrowers before an application is complete, restrict some fees during review and require some notices in languages other than English.
Achieved whenThe final rule is published in the Federal Register and takes effect.
Claimed by
CFPB, November 2026, as reported by ABA Banking Journal
AI agents in servicing software
ICE Mortgage Technology unveiled borrower-facing AI voice and chat agents for its servicing platform, meant to answer borrowers' questions and carry out some loan actions. They were in beta testing, with no customers named and no performance figures.
Achieved whenNamed servicers use the agents with borrowers and an outside party publishes their accuracy or complaint rates.
Claimed by
ICE Mortgage Technology, March 2026
Open
01
Test AI on hardship calls
A public test set of at least 500 made-up calls to a servicer from invented borrowers who are behind or about to fall behind, each with an answer key listing what Regulation X requires the agent to do: tell a delinquent borrower about loss-mitigation options and how to apply, give housing counsellor contacts, not tell a borrower they are approved for an option before eligibility is decided, and respect the 120-day and 37-day foreclosure limits. AI voice and chat agents are scored against the key, and speech-to-text is checked on the recordings, because a misheard loan number, date or amount changes the answer. No real borrowers or loans.
The briefCredit · unclaimed02
Escrow statements with planted errors
A public set of at least 1,000 made-up loans, each with a year of escrow activity and the annual escrow statement that Regulation X (12 CFR 1024.17) requires, with the correct figures recomputed. At least 300 statements carry one planted error each, such as a cushion above one-sixth of the year's payments out, a surplus of $50 or more not refunded to a borrower who is up to date, a shortage of a month's payment or more demanded back in fewer than 12 months, a late tax or insurance payment, or a wrong new monthly payment. Open-source calculators and AI models that read the statements as PDFs are scored on errors caught and correct statements wrongly flagged.
The briefCredit · unclaimed
Open 01 · the brief
Test AI on hardship calls
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 at least 500 made-up calls to a servicer from invented borrowers who are behind or about to fall behind, each with an answer key listing what Regulation X requires the agent to do: tell a delinquent borrower about loss-mitigation options and how to apply, give housing counsellor contacts, not tell a borrower they are approved for an option before eligibility is decided, and respect the 120-day and 37-day foreclosure limits. AI voice and chat agents are scored against the key, and speech-to-text is checked on the recordings, because a misheard loan number, date or amount changes the answer. No real borrowers or loans.
Why it is still open
Rocket says AI handles 65% of its inbound servicing calls and ICE is testing voice agents, with no outside test of either. CFPB examiners found servicers' staff getting the same steps wrong, failing to reach hundreds of delinquent borrowers and telling borrowers they were approved before eligibility was decided. Public benchmarks cover neighbouring work only: MortarBench (June 2026) tests origination agents and INSURE-Dial (2026) tests insurance-benefit calls, and searches of GitHub and Hugging Face found no servicing call set.
Built with what exists
The rule text at 12 CFR 1024.39 and 1024.41; a general language model to write scripts and play the borrower; a text-to-speech service such as ElevenLabs to voice the borrower; open-source voice-agent frameworks such as Pipecat or LiveKit Agents to connect the simulated caller to the agent under test; ElevenLabs Scribe v3 (see the Note "Fewer word errors") and OpenAI's open-source Whisper for speech-to-text, compared on loan numbers, dates and amounts; Gemini Decision 1 to score each transcript against many yes-or-no questions at once (see the Note "A thousand questions, one pass"), checked by two people.
First step
Turn 12 CFR 1024.39 and 1024.41 into about 15 yes-or-no checks, write 20 scripts, and have two people answer the checks for each script without seeing each other's answers.
Done looks like
At least 500 calls across at least 20 scenario types, each answer cited to a paragraph of Regulation X; two reviewers label a 100-call sample and their agreement is published; scored runs for at least three AI agents; word error rates on numbers, dates and amounts for at least two speech-to-text systems on 100 recorded calls.
Open 02 · the brief
Escrow statements with planted errors
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 1,000 made-up loans, each with a year of escrow activity and the annual escrow statement that Regulation X (12 CFR 1024.17) requires, with the correct figures recomputed. At least 300 statements carry one planted error each, such as a cushion above one-sixth of the year's payments out, a surplus of $50 or more not refunded to a borrower who is up to date, a shortage of a month's payment or more demanded back in fewer than 12 months, a late tax or insurance payment, or a wrong new monthly payment. Open-source calculators and AI models that read the statements as PDFs are scored on errors caught and correct statements wrongly flagged.
Why it is still open
A reference calculator now exists: escrow-analysis-lab (MIT licence, September 2026) computes the statements under 1024.17 with one worked sample loan and 47 tests, and loan-servicer (November 2026) includes an escrow analysis whose rules it calls "illustrative". Neither publishes statements with known errors to score against, and a Hugging Face search found no such set. Errors happen: CFPB examiners found escrow payments that never reached the tax office or insurer and were not resent for months, leaving borrowers with penalties.
github.comgithub.comfiles.consumerfinance.govlaw.cornell.edu
Built with what exists
escrow-analysis-lab and loan-servicer as reference calculators; the worked examples in Regulation X's Appendix E as ground truth; Python with ReportLab to print the statements as PDFs; Meta's Muse Ember to pull figures from the PDFs (see the Note "The jobs agents repeat"); invented names, addresses and amounts only.
First step
Enter the Appendix E examples into escrow-analysis-lab and check whether it reproduces each one to the cent.
Done looks like
1,000 statements, at least 300 with planted errors of at least eight types; at least two calculators reproducing every Appendix E example to the cent; catch rates and false-flag rates published for each calculator and for at least two AI models.