> AI in mortgage, borrower side: 21.9% of 2025 purchase applications ended with no denial and no loan. Where borrowers drop, what a model can do, the rules.

Canonical: https://www.sellence.com/guides/ai-in-mortgage

Title: AI in Mortgage Lending: The Part After the Rate Quote

Lending Guide

# AI in Mortgage Lending: The Part After the Rate Quote, from Application to Close

By [Guy Frum](https://www.linkedin.com/in/guy-frum), Head of Growth Updated 11 October 2026 12 min read

Short Answer

In 2025, 21.9% of US home purchase applications, about 1.17 million, ended with no denial and no loan: the borrower withdrew, the file was closed as incomplete, or the borrower was approved and never took the loan ( [our analysis of 2025 HMDA data](https://ffiec.cfpb.gov/data-browser/), released 2026-06-23). AI in mortgage lending means software models doing work in the loan process, from reading documents and underwriting files to talking with borrowers between the rate quote and closing.

Captions are on screen. Tap Sound On to hear it, or read the transcript below.

**Read the Transcript**

1 in 5 home purchase applications ended with no loan last year. Not denied. The borrower just stopped. They stop in 4 places. After the rate quote. Halfway through the application. Waiting on documents. Approved, but never closed. Most AI in mortgage reads documents. Very little of it follows up with the borrower. At each stop, a model can answer the question, ask for what's missing, and bring the borrower back to the loan officer. Find where yours stop. Read the guide on sellence.com.

Most lenders point AI at documents and underwriting. This guide covers the borrower: where borrowers drop, what AI does in mortgage today, what a model can do at each drop point, and the rules for every borrower message.

## Where Borrowers Drop Between the Rate Quote and Closing

Pull-through here means MBA's origination measure: loan closings divided by full applications in the same period, with leads and preapprovals left out. It has fallen since 2021. In the MBA and STRATMOR Peer Group Roundtables, retail pull-through reached 55% at depositories in the first half of 2025, the lowest since 2000, and 69% at independent mortgage banks, the lowest since 2012 ( [MBA Newslink](https://newslink.mba.org/?p=264486), 2025-11-03). MBA adds it would be lower still with leads and preapprovals counted.

HMDA filings show where the rest went. Our query of the 2025 national data:

| 2025 home purchase applications | Share | What happened |
| --- | --- | --- |
| Originated | 66.8% | Loan closed |
| Withdrawn by the borrower | 15.2% (816,198 of 5,374,348) | Borrower told the lender to stop |
| Closed for incompleteness | 3.7% | Borrower never sent what was asked for, after a written notice |
| Approved, not accepted | 3.0% | Lender said yes, borrower walked away |
| Borrower-side exits (the 3 rows above) | 21.9% (about 1.17 million) | No denial and no loan |

Refinance runs higher: 27.9% of refinance and 29.4% of cash-out refinance applications ended these 3 ways. HMDA covers reporting institutions only and counts a borrower who applies at 2 lenders twice.

The margin on each closed loan is thin: $10,936 of production expense against $973 of pretax profit per loan at independent mortgage banks in Q2 2026 (MBA data via [HousingWire](https://housingwire.com/articles/imb-mortgage-profits-q2-2026), 2026-08-18).

Sources: our analysis of 2025 HMDA data, ICE Mortgage Monitor May 2026.

There are 4 drop points. Withdrawals, the largest exit at 15.2%, can happen at any step after the application.

### 1. Rate Shopped

The borrower asked for a rate, left a phone number or email, and has not applied. No public data measures how many of these become applications. What exists says the lender that gets the application usually keeps it: in the CFPB's survey of 2013 purchase borrowers, almost half seriously considered only 1 lender, and about 77% applied to only 1 ( [CFPB, January 2015](https://files.consumerfinance.gov/f/201501_cfpb_consumers-mortgage-shopping-experience.pdf) ).

Early contact also lifts satisfaction: J.D. Power's 2025 origination study scored it 32 points higher when the lender engages before the borrower starts shopping, and 64 points lower when first contact comes at the application ( [J.D. Power](https://www.jdpower.com/business/press-releases/2025-us-mortgage-origination-satisfaction-study), 2025-11-12, 10,067 borrowers).

### 2. Application Started

The borrower opened the application and left before submitting. We found no primary, recent abandonment rate for mortgage, and the figures on page 1 of Google trace back to vendor marketing. MBA lenders list confusion over the application process among their fallout reasons. Measure 2 rates yourself: applications started versus submitted, and the share of started applications with a phone number. Tools for this step are compared in [best AI for abandoned loan signup recovery](https://www.sellence.com/best-ai-abandoned-loan-signup-recovery).

### 3. Documents Missing

The Loan Estimate is out and the file waits on pay stubs, statements or a letter of explanation. In 2025, 3.7% of purchase applications ended here. The window is short: ICE put average purchase time to close at 36.8 days in March 2026, the fastest since 2019, with a typical 11 days from application to rate lock and 26 from lock to close ( [ICE Mortgage Monitor](https://mortgagetech.ice.com/resources/data-reports/may-2026-mortgage-monitor), 2026-05-11).

Before closing a file as incomplete, Reg B requires a written notice naming what is missing and setting a reasonable deadline ( [12 CFR 1002.9(c)](https://www.ecfr.gov/current/title-12/section-1002.9) ). A text reminder can come first, and the written notice still has to go out.

### 4. Approved, Not Taken

The lender said yes and the borrower did not close: 3.0% of 2025 purchase applications. MBA lenders gave the reasons: the borrower applied with several lenders and chose another, the payment grew once taxes, insurance, HOA fees and closing costs were added, second thoughts about the economy, and the condition of the home.

## What AI Does in Mortgage Today

The MBA, AARMR and BCG survey released in September 2026 ( [report](https://apps.mba.org/pdf/BCGxMBAxAARMR%20Survey%20-%2018Sep26%20VShared.pdf), 2026-09-18) asked 31 lenders and servicers, covering about 40% of US origination volume, where they run AI. About 95% have at least 1 use case in production. The work sits mostly in the back office:

| Use case | Lenders with it in production |
| --- | --- |
| Data extraction from documents | 61% |
| Customer summaries and tailored outreach | 23% |
| Loan officer assistant and next-best-action | 23% |

For the typical lender, 39% of the 5 marketing and sales use cases are not in use at all, and only 6% of lenders report a significant customer experience benefit. Regulatory concern is the top barrier to scaling (59%), ahead of unclear ROI (45%), and 81% keep a human reviewing AI output.

In Fannie Mae's 2023 lender survey ( [report](https://fanniemae.com/media/49231/display), 219 institutions), among lenders using AI, 7% named borrower experience as the main goal, down from 41% in 2018, and an AI virtual assistant talking to consumers was the idea most often picked as least appealing (35%).

Borrowers are further along. In J.D. Power's 2025 study, 54% were completely comfortable with lenders using AI in origination and 31% partially, and 71% said it was very important to be told when AI is used. The borrower conversation between rate quote and closing is the least built part of mortgage AI today.

## What a Model Can Do at Each Drop Point

A model here means software that holds a 2-way conversation with the borrower by text, call or chat, in the lender's name, inside the rules below. Anything about rates or terms goes to a licensed loan officer.

| Drop point | What the borrower needs | What a model can do | What stays with the loan officer |
| --- | --- | --- | --- |
| Rate shopped | A fast answer and a reason to apply | Answer process questions, start the application, book a call | Any rate, payment or loan program |
| Application started | A way back in | Pick up where the borrower left off, explain why each item is needed, resend the link | Reviewing the file, issuing the Loan Estimate |
| Documents missing | To know exactly what is missing | Name the item, resend the upload link, remind at a time the borrower picked | The written incompleteness notice, conditions |
| Approved, not taken | Help with the doubt | Ask what changed, point to where each cost sits on the Loan Estimate, book the loan officer | Comparing offers, re-pricing, negotiating |

If you are comparing AI tools built for loan officers, [this comparison](https://www.sellence.com/compare-loanofficer-ai) sets out how they differ.

### An Example Exchange (Illustrative)

Illustrative only: the borrower, lender and loan officer are invented. Dana applied 4 days ago, received her Loan Estimate 2 days ago and gave written consent to texts. Her file waits on 2 documents.

> Day 4, 10:05 a.m.
> Lender: Hi Dana, it's Maya, the AI assistant on Chris Lee's loan team at [Lender]. Your file is waiting on 2 items: your 2 most recent pay stubs and your latest checking account statement. Want the upload link again? Reply STOP to opt out.

The message names the exact items and carries the opt-out, and Maya says she is AI.

> Dana: uploaded the stubs. not sure we're going ahead tbh, the closing costs scared us
> Lender: Thanks for telling me. Closing costs are itemized on page 2 of your Loan Estimate, and some of those services you can shop for yourself. Chris can go through each line with you. Does 9:00 or 12:30 tomorrow work for a 15-minute call?

The model points to where the numbers live, books the loan officer for the terms, and quotes no figure.

> Dana: 12:30
> Lender: Booked: Chris calls you at 12:30 tomorrow. Here's the link for the bank statement in the meantime, so the file keeps moving while you decide.
> Call booked. 1 document left.

## The Rules a Borrower Message Must Follow

This is general information, not legal advice. Mortgage rules differ by state and changed in 2025 and 2026. Check with compliance counsel before a model sends a borrower any rate, payment or term.

### Never Quote a Rate, Payment or Term

Under Reg Z, any commercial message that promotes a credit transaction is an advertisement ( [12 CFR 1026.2(a)(2)](https://www.ecfr.gov/current/title-12/section-1026.2) ), and a templated text sent to many borrowers counts. In a mortgage ad, 4 trigger terms require full disclosure, including the APR: the down payment, the repayment period or number of payments, the payment amount and the finance charge ( [12 CFR 1026.24(d)](https://www.ecfr.gov/current/title-12/chapter-X/part-1026/subpart-C/section-1026.24) ).

A one-to-one reply in a live negotiation usually falls outside the ad rules, and 2 other rules still reach it. A written estimate of terms for one borrower before the Loan Estimate must carry a mandatory 12-point statement that the actual rate, payment and costs could be higher ( [12 CFR 1026.19(e)(2)(ii)](https://www.ecfr.gov/current/title-12/section-1026.19) ). Reg N bans any material misstatement of a rate, fee or other term in any commercial communication, including telemarketing scripts and messages over a cellular network ( [12 CFR 1014.3](https://www.ecfr.gov/current/title-12/section-1014.3) ).

The working rule: the model never states a rate, APR, payment, down payment or term in years. It points to the Loan Estimate and the loan officer.

### The 6 Items That Make an Application

A borrower has applied once the lender holds 6 items: name, income, Social Security number, property address, estimated property value and loan amount ( [12 CFR 1026.2(a)(3)](https://www.ecfr.gov/current/title-12/section-1026.2) ). The Loan Estimate is then due by the third business day (1026.19(e)(1)(iii)). A model that collects all 6 in a conversation has started that clock, so the transcript has to reach the loan origination system the same day. Before the Loan Estimate, a lender cannot require verification documents or charge any fee beyond a credit report fee (1026.19(e)(2)), so document chasing starts after it goes out.

Under Reg B, a model that reviews a borrower's information and says they won't qualify may have created an application, with an adverse action notice due (comment 2(f)-3 to [12 CFR 1002.2](https://www.ecfr.gov/current/title-12/section-1002.2) ).

### Preapproval Status

Reg N also bans misrepresenting whether a borrower has been preapproved (1014.3). The model states preapproval or approval status only as the loan origination system shows it.

### Consent to Text and Call

TCPA consent rules for texts and calls changed several times in 2025 and 2026, and our [quote-to-bind guide](https://www.sellence.com/guides/quote-to-bind) covers written consent, opt-outs and calling hours. A rate request on a comparison site gives consent to the companies named in that site's consent language, so check that yours is named.

### Licensing and a Human in the Loop

The SAFE Act licenses individuals who take applications and offer or negotiate terms, including presenting particular terms or recommending a lender ( [12 CFR 1008.103](https://www.ecfr.gov/current/title-12/section-1008.103) ). Clerical work is exempt, including talking with a borrower to obtain information for processing. The application, Loan Estimate and Closing Disclosure name a responsible loan originator and their NMLS ID ( [12 CFR 1026.36(g)](https://www.ecfr.gov/current/title-12/section-1026.36) ).

An MBA white paper prepared by Orrick (June 2026) concluded the SAFE Act does not require an AI tool to hold its own originator license, and recommended a licensed originator stays available to borrowers and oversees the process ( [HousingWire](https://www.housingwire.com/articles/mba-white-paper-ai/), 2026-06-10). It also flagged the risk of borrowers believing a human oversees a file that AI handles alone, one more reason for the model to say what it is. We found no state regulator guidance on this as of October 2026.

The working split: the model answers general questions, reminds, schedules, collects information and routes. The licensed loan officer presents terms, discusses the Loan Estimate, negotiates and decides.

### Fair Lending

A Reg B rule effective 2026-07-21 removed disparate impact liability under the Equal Credit Opportunity Act ( [91 FR 21620](https://www.federalregister.gov/documents/2026/04/22/2026-07804/equal-credit-opportunity-act-regulation-b) ), and for mortgage the Fair Housing Act and state laws can still impose it. The discouragement ban stands: a lender may make no statement that would lead a reasonable person to expect denial or worse terms because of a protected characteristic ( [12 CFR 1002.4(b)](https://www.ecfr.gov/current/title-12/section-1002.4) ). A model that answers "you probably won't qualify" after a borrower mentions retirement, public assistance income or marital status fits that pattern, so take that answer out of the scripts.

The CFPB withdrew its 2022 circular on adverse action reasons from algorithms on 2025-05-12 ( [CFPB](https://www.consumerfinance.gov/compliance/guidance/withdrawn-guidance/) ), and the rule behind it still applies: every adverse action notice gives specific reasons (1002.9).

## How to Use AI in Mortgage, in 5 Steps

1. Count Your Exits Withdrawn, closed for incompleteness, and approved but not accepted, from your own HMDA data, by channel and loan purpose.
2. Pick 1 Drop Point Start with the drop point that holds the most files.
3. Write the Rule Sheet First Write the rule sheet before the scripts: no rates, payments or terms, a same-day handoff once the 6 application items are in, preapproval status only as your system shows it, and opt-outs.
4. Set the Handoff Which questions go to a licensed loan officer, and how fast.
5. Measure Against a Holdout Pull-through and exits for borrowers the model talked to, against a matched group it did not.

## FAQ

### What is mortgage AI?

Mortgage AI is software that uses models to do work in the loan process: reading documents, supporting underwriting, helping loan officers and talking with borrowers. In a September 2026 MBA survey of 31 lenders, 61% had document data extraction in production.

### How do you use AI in mortgage?

Pick 1 measurable problem. On the borrower side, count withdrawn, incomplete and approved-but-not-accepted applications, start with the largest group, keep rates and terms with a licensed loan officer, and measure pull-through against a holdout group.

### Can AI quote a mortgage rate to a borrower?

It should not. A rate, payment, down payment or term in a message can trigger Reg Z disclosures or count as a written estimate before the Loan Estimate, and Reg N bans misstating any of them. Leave them to the Loan Estimate and the loan officer.

### Does an AI tool need a mortgage loan originator license?

A June 2026 MBA white paper prepared by Orrick concluded the SAFE Act does not require an AI tool to hold its own originator license. It recommends a licensed originator stays in the loop.

### What is mortgage pull-through?

Pull-through is loan closings divided by applications in the same period, full applications only. MBA and STRATMOR data put retail pull-through at 55% for depositories and 69% for independent mortgage banks in the first half of 2025.

### Do borrowers want to know when AI is used?

Yes. In J.D. Power's 2025 mortgage origination study, 71% of borrowers said it is very important that the lender tells them when AI is used, and 54% were completely comfortable with lenders using it.

Sellence built a persuasion model that follows up with borrowers by text and call in your company's name for weeks, in one thread that remembers what each borrower said, and answers the question or doubt behind an unfinished file while rates and terms stay with your licensed loan officers. Nations Lending is a customer. See [how it works for lending](https://www.sellence.com/lending), or [book a demo](https://www.sellence.com/demo) to see it on your borrower funnel.

## Sources

1. FFIEC, HMDA Data Browser, 2025 Snapshot National Loan-Level data (filings as of 2026-06-01, released 2026-06-23), our query by action taken, 2026-10-11. [https://ffiec.cfpb.gov/data-browser/](https://ffiec.cfpb.gov/data-browser/) and release [https://www.ffiec.gov/news/press-releases/2026/an-06-23](https://www.ffiec.gov/news/press-releases/2026/an-06-23)
2. MBA Newslink, Chart of the Week: Retail Channel Mortgage Pull-Through, Marina Walsh, 2025-11-03 (MBA and STRATMOR Peer Group Roundtables, H1 2025 data). [https://newslink.mba.org/?p=264486](https://newslink.mba.org/?p=264486)
3. MBA Quarterly Mortgage Bankers Performance Report, Q2 2026, reported by HousingWire, Sarah Wolak, 2026-08-18. [https://housingwire.com/articles/imb-mortgage-profits-q2-2026](https://housingwire.com/articles/imb-mortgage-profits-q2-2026)
4. ICE Mortgage Monitor, May 2026 edition (March 2026 data), released 2026-05-11. [https://mortgagetech.ice.com/resources/data-reports/may-2026-mortgage-monitor](https://mortgagetech.ice.com/resources/data-reports/may-2026-mortgage-monitor)
5. CFPB, Consumers' mortgage shopping experience: A first look at results from the National Survey of Mortgage Borrowers, January 2015. [https://files.consumerfinance.gov/f/201501_cfpb_consumers-mortgage-shopping-experience.pdf](https://files.consumerfinance.gov/f/201501_cfpb_consumers-mortgage-shopping-experience.pdf)
6. J.D. Power, 2025 U.S. Mortgage Origination Satisfaction Study, released 2025-11-12. [https://www.jdpower.com/business/press-releases/2025-us-mortgage-origination-satisfaction-study](https://www.jdpower.com/business/press-releases/2025-us-mortgage-origination-satisfaction-study)
7. MBA, AARMR and BCG, The state of AI among mortgage lenders and servicers, report dated 2026-09-18 (fielded April to July 2026, 31 respondents). [https://apps.mba.org/pdf/BCGxMBAxAARMR%20Survey%20-%2018Sep26%20VShared.pdf](https://apps.mba.org/pdf/BCGxMBAxAARMR%20Survey%20-%2018Sep26%20VShared.pdf)
8. Fannie Mae, Mortgage Lender Sentiment Survey special topic, Artificial Intelligence and Mortgage Lending, October 2023 (fielded August 2023). [https://fanniemae.com/media/49231/display](https://fanniemae.com/media/49231/display)
9. eCFR, 12 CFR 1026.2 (definitions, application). [https://www.ecfr.gov/current/title-12/section-1026.2](https://www.ecfr.gov/current/title-12/section-1026.2)
10. eCFR, 12 CFR 1026.19 (Loan Estimate timing, pre-Loan Estimate limits). [https://www.ecfr.gov/current/title-12/section-1026.19](https://www.ecfr.gov/current/title-12/section-1026.19)
11. eCFR, 12 CFR 1026.24 (advertising, trigger terms). [https://www.ecfr.gov/current/title-12/chapter-X/part-1026/subpart-C/section-1026.24](https://www.ecfr.gov/current/title-12/chapter-X/part-1026/subpart-C/section-1026.24)
12. eCFR, 12 CFR 1026.36 (loan originator rules). [https://www.ecfr.gov/current/title-12/section-1026.36](https://www.ecfr.gov/current/title-12/section-1026.36)
13. eCFR, 12 CFR 1014.3 (Reg N, prohibited representations). [https://www.ecfr.gov/current/title-12/section-1014.3](https://www.ecfr.gov/current/title-12/section-1014.3)
14. eCFR, 12 CFR 1008.103 (SAFE Act state licensing). [https://www.ecfr.gov/current/title-12/section-1008.103](https://www.ecfr.gov/current/title-12/section-1008.103)
15. eCFR, 12 CFR 1002.2, 1002.4 and 1002.9 (Reg B definitions, discouragement, notifications). [https://www.ecfr.gov/current/title-12/section-1002.2](https://www.ecfr.gov/current/title-12/section-1002.2), [https://www.ecfr.gov/current/title-12/section-1002.4](https://www.ecfr.gov/current/title-12/section-1002.4), [https://www.ecfr.gov/current/title-12/section-1002.9](https://www.ecfr.gov/current/title-12/section-1002.9)
16. Federal Register, Equal Credit Opportunity Act (Regulation B) final rule, 91 FR 21620, published 2026-04-22, effective 2026-07-21. [https://www.federalregister.gov/documents/2026/04/22/2026-07804/equal-credit-opportunity-act-regulation-b](https://www.federalregister.gov/documents/2026/04/22/2026-07804/equal-credit-opportunity-act-regulation-b)
17. CFPB, Withdrawn guidance (Circular 2022-03 withdrawn 2025-05-12), page last modified 2026-07-22. [https://www.consumerfinance.gov/compliance/guidance/withdrawn-guidance/](https://www.consumerfinance.gov/compliance/guidance/withdrawn-guidance/)
18. HousingWire, MBA white paper on AI-powered mortgage and federal law (prepared by Orrick), 2026-06-10. [https://www.housingwire.com/articles/mba-white-paper-ai/](https://www.housingwire.com/articles/mba-white-paper-ai/) (Orrick summary, 2026-06-15: https://www.orrick.com/en/Insights/2026/06/Examining-AI-Powered-Mortgage-Through-the-Lens-of-Federal-Law)

Related

- [Sellence vs LoanOfficer.ai](https://www.sellence.com/compare-loanofficer-ai)
- [AI for Abandoned Loan Signup Recovery](https://www.sellence.com/best-ai-abandoned-loan-signup-recovery)
- [Quote-to-Bind Ratio Guide](https://www.sellence.com/guides/quote-to-bind)
- [Glossary](https://www.sellence.com/glossary)