AI in Mortgage Lending Statistics 2026
By Axis Intelligence Research
Co-authors: Sarah Davis & Sarah Mitchell | Last updated: October 13, 2026 | License: CC BY 4.0
It now costs a U.S. mortgage lender $10,936 to originate one loan, according to the Mortgage Bankers Association’s Q2 2026 data. Axis Intelligence Research calculates that this sits 37.6% above the 18-year norm, about $17.7 billion of excess cost across 2026 volume. That gap, not the chatbot, is the business case for AI in mortgage lending.
Quick Answer
AI in mortgage lending is long on motive and short on deployment. Only 7% of lenders told Fannie Mae in 2023 that they had fully deployed AI/ML, down from 14% in 2018, while 73% named operational efficiency as their reason to adopt. Meanwhile, origination costs average $10,936 per loan (MBA, Q2 2026), and Axis Intelligence Research estimates $17.7 billion of 2026 production spend sits above the long-run average.
Key Findings
- According to Axis Intelligence Research, the Mortgage Origination Cost Excess (MOCE™) reads 37.6% for Q2 2026: lenders spent $10,936 per loan against an 18-year MBA average of $7,945.
- Axis Intelligence Research estimates U.S. lenders will spend about $64.7 billion to originate mortgages in 2026, of which roughly $17.7 billion is cost above the historical per-loan norm.
- Fannie Mae’s 2023 lender survey found only 7% of mortgage lenders had fully deployed AI/ML, half the 14% share reported in 2018.
- According to Axis Intelligence Research, the Lender Intent-to-Deployment Ratio (LIDR™) rose from 3.0 in 2018 to 10.4 in 2023: ten lenders want AI for efficiency for every one that has it in production.
- As of October 13, 2026, mortgage AI sits under three live rulebooks: the six-agency AVM quality control rule (effective October 1, 2025), Freddie Mac Bulletin 2025-16 (effective March 3, 2026) and Fannie Mae Lender Letter LL-2026-04 (April 8, 2026).
How Many Mortgage Lenders Actually Use AI?
Fewer than the conference agenda suggests. The best primary measurement of lender adoption comes from Fannie Mae’s Economic & Strategic Research Group, which asked senior mortgage executives the same AI questions in 2018 and again in 2023, a rare like-for-like comparison in a field crowded with vendor surveys.
The 2023 results published by Fannie Mae show familiarity barely moved, from 63% to 65%. Full deployment went backwards: 7% of lenders said they had deployed AI/ML in 2023, against 14% in 2018. Limited or trial deployment rose from 13% to 22%. Axis Intelligence Research adds the two: 29% of lenders were running AI in production or pilot in 2023, compared with 27% five years earlier, essentially flat.
Expectations cooled as well. Just 29% of lenders expected to roll out AI/ML tools more broadly within two years, down from 38% in 2018. In the original 2018 study, 58% of lenders said they expected to adopt some AI solution within two years.
Mortgage Lender AI Adoption, 2018 vs 2023
| Metric | 2018 | 2023 | Change | Source |
|---|---|---|---|---|
| Familiar with AI/ML | 63% | 65% | +2 pp | Fannie Mae ESR |
| Fully deployed AI/ML | 14% | 7% | −7 pp | Fannie Mae ESR |
| Limited or trial deployment | 13% | 22% | +9 pp | Fannie Mae ESR |
| Deployed or in trial (Axis sum) | 27% | 29% | +2 pp | Axis Intelligence Research, from Fannie Mae ESR |
| Expect broader rollout within two years | 38% | 29% | −9 pp | Fannie Mae ESR |
| Primary motivation: operational efficiency | 42% | 73% | +31 pp | Fannie Mae ESR |
| Primary motivation: borrower experience | 41% | 7% | −34 pp | Fannie Mae ESR |
The lenders that had not adopted gave the same three reasons in both waves: integration complexity with existing infrastructure, no proven track record, and high cost. Fannie Mae reports that mortgage banks cite integration complexity more often than depository institutions, and that data security and privacy concerns grew significantly since 2018.
Sarah Davis, Axis Intelligence Research: The number that matters in that table is not 7%. It is the swing in motive. In 2018 lenders wanted AI to make borrowers happier; by 2023 they wanted it to make the file cheaper. That is what a lender says after two years of negative production margins. Nobody buys a better point-of-sale experience while losing money on every closed loan. They buy something that pulls hours out of processing and underwriting, and they want to see the cost line move before they sign a second contract.
Why Did Full AI Deployment Fall Between 2018 and 2023?
Fannie Mae does not publish a causal answer, and Axis Intelligence Research will not invent one. What the survey does show is a definitional shift: more lenders moved into the “limited or trial” bucket while fewer claimed full deployment. A plausible reading, consistent with the barriers lenders cited, is that 2023 respondents applied a stricter test to what counts as “deployed” once generative AI raised the bar. The data supports the narrower claim only: production-grade AI remained a minority practice among mortgage lenders.
What Is the Lender Intent-to-Deployment Ratio (LIDR™)?
The Lender Intent-to-Deployment Ratio (LIDR™) is an Axis Intelligence Research metric that measures how far lender demand for AI-driven efficiency runs ahead of actual production deployment.
Formula: LIDR = share of lenders naming operational efficiency as their primary AI motivation ÷ share of lenders reporting full AI/ML deployment
Readings:
| Survey year | Efficiency motivation | Full deployment | LIDR™ | Source of inputs |
|---|---|---|---|---|
| 2018 | 42% | 14% | 3.0 | Fannie Mae ESR |
| 2023 | 73% | 7% | 10.4 | Fannie Mae ESR |
According to Axis Intelligence Research, the LIDR rose 3.5-fold in five years. Both inputs come from the same survey instrument, asked of the same population of senior mortgage executives, so the ratio compares like with like. It does not say how many loans touch AI; it says how much unmet appetite sits on the lender side of the table, which is the number a mortgage technology vendor’s pipeline, and a GSE’s product roadmap, are built on.
How Much Does It Cost to Originate a Mortgage in 2026?
Independent mortgage banks and bank mortgage subsidiaries spent $10,936 to originate each loan in Q2 2026, or 308 basis points of loan balance, according to the MBA Quarterly Mortgage Bankers Performance Report released August 18, 2026. That expense line covers commissions, compensation, occupancy, equipment and corporate allocations. The long-run average from Q2 2008 through Q2 2026 is $7,945 per loan.
Revenue barely covers it. Production revenue averaged $11,909 per loan, leaving a pretax net production profit of $973 per loan, up from $727 in Q1 2026. The margin of 25 basis points remains below the 39 basis point average MBA has recorded since 2008. The average reporting company originated 1,958 loans in the quarter.
Axis Intelligence Research calculates that every dollar of Q2 2026 production profit required $11.2 of production expense, and that profit was 8.2% of production revenue. A lender that removes $1,000 of cost per file roughly doubles its per-loan profit at Q2 2026 economics.
Mortgage Production Economics, Q2 2026
| Metric | Value | As of | Source |
|---|---|---|---|
| Production expense per loan | $10,936 | Q2 2026 | MBA |
| Production expense, basis points | 308 bps | Q2 2026 | MBA |
| Long-run average expense per loan | $7,945 | Q2 2008–Q2 2026 | MBA |
| Production revenue per loan | $11,909 | Q2 2026 | MBA |
| Net production profit per loan | $973 | Q2 2026 | MBA |
| Net production profit per loan, prior quarter | $727 | Q1 2026 | MBA |
| Net production margin | 25 bps | Q2 2026 | MBA |
| Long-run average margin | 39 bps | Since 2008 | MBA |
| Expense-to-profit ratio | 11.2× | Q2 2026 | Axis Intelligence Research |
| Profit as share of revenue | 8.2% | Q2 2026 | Axis Intelligence Research |
Which Regions Have the Highest Mortgage Origination Costs?
MBA’s Q1 2026 regional cost breakdown put California at $12,864 per loan and the Northeast at $12,791, the two most expensive markets, with the Midwest lowest. High-cost regions are where per-file automation savings land on the largest base, which makes coastal lenders the logical first buyers of document AI and automated income verification.
What Is the Mortgage Origination Cost Excess (MOCE™)?
The Mortgage Origination Cost Excess (MOCE™) is an Axis Intelligence Research metric that measures how far the current cost to originate a mortgage sits above its long-run average, expressed as a percentage. It quantifies the cost problem AI is being bought to solve.
Formula: MOCE = (current production expense per loan − long-run average expense per loan) ÷ long-run average expense per loan
Inputs (MBA): current expense $10,936 (Q2 2026); long-run average $7,945 (Q2 2008–Q2 2026).
Computation: (10,936 − 7,945) ÷ 7,945 = 2,991 ÷ 7,945 = 37.6%
| Quarter | Expense per loan | Long-run average at the time | MOCE™ | Inputs |
|---|---|---|---|---|
| Q2 2025 | $10,965 | $7,750 | 41.5% | MBA |
| Q2 2026 | $10,936 | $7,945 | 37.6% | MBA |
According to Axis Intelligence Research, the MOCE narrowed by 3.8 points year over year, but most of that improvement came from the denominator: the long-run average itself rose from $7,750 to $7,945 as expensive recent quarters entered the series. Per-loan cost fell by only $29. In dollar terms, the excess stands at $2,991 per loan, or 27.4% of today’s cost.
How Much Could AI Save Mortgage Lenders in 2026?
Axis Intelligence Research does not publish a savings forecast, because no primary source measures AI’s per-loan impact across the industry. What can be sized is the cost pool AI is competing for.
MBA projects $2.1 trillion in 2026 originations, falling 5% to just under $2 trillion in 2027. Applying the Q2 2026 production expense ratio:
- Estimated 2026 origination cost pool: $2.1 trillion × 308 bps = $64.7 billion
- Estimated excess above the long-run norm: $64.7 billion × 27.4% = $17.7 billion
These are Axis Intelligence Research estimates, not observations. The expense ratio comes from MBA’s sample of independent mortgage banks and bank subsidiaries; applying it to all originations assumes depository lenders run similar cost structures, which MBA data does not confirm. The figures size the prize. They do not claim any lender will capture it.
Sarah Davis, Axis Intelligence Research: Read the $17.7 billion as a ceiling on what a mortgage AI vendor can credibly promise the industry, not a floor. Much of the excess is loan officer comp and fixed overhead spread across thin volume, and no document classifier touches either. The honest pitch is narrower: processing, underwriting conditions, income calculation, QC sampling. Those lines are where Rocket reports its hours came back, and they are a fraction of the $10,936.
What Does AI Do Inside a Mortgage Lender Today?
Public, quantified disclosures are scarce, and most come from the lenders themselves. The most detailed remains Rocket Companies’ April 9, 2024 Rocket Logic announcement, all figures company-reported for February 2024 unless noted. These are older data: Axis Intelligence Research located no newer quantified operating disclosure from a top-tier lender in a primary document.
| Rocket Mortgage AI operating metric (company-reported) | Value | Period |
|---|---|---|
| Documents received per month | more than 1.5 million | Feb 2024 |
| Share automatically identified | nearly 70% | Feb 2024 |
| Underwriter hours saved | more than 5,000 | Feb 2024 |
| Data points extracted from documents | 4.3 million | Feb 2024 |
| Share of data points processed automatically | nearly 90% | Feb 2024 |
| Team member interactions per loan | down nearly 25% year over year | Feb 2024 |
| Client turn times | down 25% | Aug 2022–Feb 2024 |
Source: Rocket Companies press release, April 9, 2024.
The use cases match what Fannie Mae’s 2023 respondents asked the GSEs to build: appraisal automation, borrower income and employment verification, data and documentation reconciliation, and compliance management. Among the seven ideas Fannie Mae tested, AI-based compliance review ranked first and anomaly detection for fraud and defects ranked second. Borrower-facing virtual assistants ranked lower, consistent with the collapse of the borrower-experience motive from 41% to 7%.
What Rules Govern AI in Mortgage Lending in 2026?
Three primary rulebooks now apply directly to mortgage AI, and a lender selling to both GSEs answers to all of them.
The AVM Quality Control Rule
Six agencies (OCC, Federal Reserve, FDIC, NCUA, CFPB and FHFA) jointly issued Quality Control Standards for Automated Valuation Models, 89 FR 64538, effective October 1, 2025. Institutions using AVMs in credit decisions must adopt policies covering five factors: high confidence in estimates, protection against data manipulation, avoidance of conflicts of interest, random sample testing and reviews, and compliance with nondiscrimination laws. As of October 13, 2026, the rule has been in force for 377 days. On September 29, 2026, the industry standards body MISMO published AVM testing guidance and best practices to give lenders a common testing method.
Freddie Mac Bulletin 2025-16
Freddie Mac’s Single-Family Seller/Servicer Guide Bulletin 2025-16, issued December 3, 2025, added an AI/ML governance framework to Guide Sections 1302.2 and 1302.8, effective March 3, 2026, a 90-day implementation window. Freddie Mac describes it as covering the responsible development, deployment and oversight of AI and ML systems, with requirements on transparency, accountability and ethical stewardship.
Fannie Mae Lender Letter LL-2026-04
On April 8, 2026, Fannie Mae published Lender Letter LL-2026-04, a governance framework for Seller/Servicers’ use of AI and ML in origination and servicing.
| Rule | Issuer | Key date | Days elapsed at Oct 13, 2026 |
|---|---|---|---|
| AVM Quality Control Standards (89 FR 64538) | OCC, Fed, FDIC, NCUA, CFPB, FHFA | Effective Oct 1, 2025 | 377 |
| Seller/Servicer Guide Bulletin 2025-16 (AI/ML governance) | Freddie Mac | Effective Mar 3, 2026 | 224 |
| Lender Letter LL-2026-04 (AI/ML governance) | Fannie Mae | Published Apr 8, 2026 | 188 |
Sources: Federal Register; Freddie Mac; Fannie Mae. Day counts calculated by Axis Intelligence Research.
For the federal fair-lending side, including the withdrawn CFPB adverse-action circulars on complex algorithms, see our AI credit scoring statistics, which own that dataset.
Sarah Davis, Axis Intelligence Research: The GSE rules change who carries the risk, and that matters more than the wording. Before 2026, an AI vendor’s document classifier was the vendor’s problem until something went wrong. Now a lender that sells to Fannie Mae or Freddie Mac is expected to govern the tool as its own. Read that next to a 7% full-deployment rate and the likely effect is clear: fewer tools, chosen more carefully, with the inventory and audit trail an investor can repurchase against.
How Big Is the Mortgage Data Pool AI Learns From?
The CFPB released 2025 HMDA data on March 31, 2026, filed by approximately 4,768 institutions. HMDA records include the automated underwriting system used on each application, which makes the dataset the public record of how machine-assisted underwriting actually runs across U.S. lending. Our US mortgage statistics and mortgage statistics by state track the volume, rate and delinquency side of the same market.
Methodology
Collection. Axis Intelligence Research assembled this dataset on October 12, 2026 from primary publishers: the Mortgage Bankers Association (Quarterly Mortgage Bankers Performance Reports for Q2 2025 and Q2 2026, Q1 2026 regional chart, October 11, 2026 forecast), Fannie Mae’s Economic & Strategic Research Group (2018 and 2023 Mortgage Lender Sentiment Survey special topics on AI), Fannie Mae Single-Family (LL-2026-04), Freddie Mac (Bulletin 2025-16), the Federal Register (89 FR 64538), the CFPB (2025 HMDA release) and Rocket Companies (company-reported operating metrics). Every figure carries its source URL and retrieval date in the CSV.
Axis metrics.
- MOCE™ = (current per-loan production expense − long-run average) ÷ long-run average. Inputs: MBA.
- LIDR™ = share of lenders naming operational efficiency as primary AI motivation ÷ share reporting full AI/ML deployment. Inputs: Fannie Mae ESR, same survey wave for numerator and denominator.
- Cost pool = MBA 2026 origination volume forecast × MBA Q2 2026 production expense ratio (308 bps). Excess pool = cost pool × (excess per loan ÷ current per-loan cost).
- All arithmetic verified twice in Python, once in floating point and once with exact rational fractions.
Scope. MBA performance data covers independent mortgage banks and mortgage subsidiaries of chartered banks, not all depositories, so the cost pool is an estimate. Fannie Mae’s survey covers senior executives of Fannie Mae lending customers and reports at the parent-company level; the most recent AI wave is 2023. Rocket figures are company-reported and date from 2024. Axis Intelligence Research excluded market-size projections for “AI in mortgage” software because none located discloses a reproducible method.
About This Dataset
File: ai-in-mortgage-lending-statistics.csv, 67 rows, one observation per row, with value, unit, as-of date, geography, segment, source organization, source document, source URL, retrieval date, primary-source flag, Axis-calculated flag and formula note.
Creator and publisher: Axis Intelligence Research.
License: CC BY 4.0. Free to reuse, including commercially, with attribution.
Citation line: Axis Intelligence Research, AI in Mortgage Lending Statistics 2026, 2026.
How to Cite This Page
APA: Axis Intelligence Research, & Davis, S. (2026, October 13). AI in mortgage lending statistics 2026: Lender adoption, the $10,936 cost per loan and the new GSE rules [Dataset]. Axis Intelligence. https://axis-intelligence.com/ai-in-mortgage-lending-statistics/
MLA: Axis Intelligence Research, and Sarah Davis. “AI in Mortgage Lending Statistics 2026: Lender Adoption, the $10,936 Cost per Loan and the New GSE Rules.” Axis Intelligence, 13 Oct. 2026, axis-intelligence.com/ai-in-mortgage-lending-statistics/.
Chicago: Axis Intelligence Research and Sarah Davis. “AI in Mortgage Lending Statistics 2026: Lender Adoption, the $10,936 Cost per Loan and the New GSE Rules.” Axis Intelligence, October 13, 2026. https://axis-intelligence.com/ai-in-mortgage-lending-statistics/.
AI in Mortgage Lending: Questions Lenders and Borrowers Are Asking
Does a lender selling loans to Fannie Mae or Freddie Mac need an AI governance program?
Yes. Freddie Mac Bulletin 2025-16 added AI/ML governance requirements to Guide Sections 1302.2 and 1302.8, effective March 3, 2026, and Fannie Mae issued its AI/ML governance framework in Lender Letter LL-2026-04 on April 8, 2026. Both apply to Seller/Servicers using AI in origination or servicing, including vendor-supplied tools.
Is an automated valuation model an AI tool under federal mortgage rules?
The federal rule regulates AVMs by function, not by technique. The six-agency Quality Control Standards for Automated Valuation Models, effective October 1, 2025, require five quality control factors, including nondiscrimination compliance, whenever an AVM is used in a covered credit decision, whether or not the model uses machine learning.
How much does AI reduce the cost of originating a mortgage?
No primary source measures an industry-wide figure. MBA reports a production cost of $10,936 per loan in Q2 2026, and Axis Intelligence Research calculates $2,991 of that sits above the 2008–2026 average. Company-reported results exist: Rocket Mortgage disclosed more than 5,000 underwriter hours saved in February 2024.
What share of mortgage lenders have fully deployed AI?
7%, according to Fannie Mae’s 2023 survey of senior mortgage executives, down from 14% in 2018. Another 22% were deploying AI on a limited or trial basis, bringing production plus pilot use to 29%, an Axis Intelligence Research sum.
Which mortgage tasks are lenders most likely to automate with AI?
Back-office ones. Fannie Mae’s 2023 survey found AI-based compliance review was the most appealing application, followed by anomaly detection for fraud and defects. Lenders asked the GSEs to build tools for appraisal automation, income and employment verification, and document reconciliation.
Why are mortgage lenders slow to adopt AI?
Lenders that had not adopted AI cited integration complexity with existing systems, lack of a proven track record and high cost, the same three barriers in 2018 and 2023, according to Fannie Mae. Data security and privacy concerns grew significantly over the period.
Will AI lower mortgage closing costs for borrowers?
Not directly in the data available. Lender production cost is not the same as borrower closing cost, and no primary source links lender AI adoption to borrower fees. What MBA data shows is a thin margin: $973 of profit on $11,909 of revenue per loan in Q2 2026, which leaves lenders little room to pass savings on before they recover margin.
What is the Mortgage Origination Cost Excess (MOCE™)?
MOCE™ is an Axis Intelligence Research metric equal to the current per-loan production cost minus its long-run average, divided by that average. Using MBA data, it read 37.6% in Q2 2026 ($10,936 vs $7,945) and 41.5% in Q2 2025.
Related Axis Intelligence Research analysis: AI credit scoring statistics · US mortgage statistics · mortgage statistics by state · AI in finance statistics · AI in insurance statistics · US banking statistics