AI data licensing for mortgage brokers

In 2022, 925 employer firms in U.S. mortgage and nonmortgage loan brokerage belonged to enterprises with at least 10 paid employees. They accounted for 75.9% of the industry's employment and 82.0% of its annual payroll.

A buyer must establish what work those firms documented and whether the seller can license the records.

Application histories may be useful when they connect a missing document or lender condition to a professional correction and a final outcome. A buyer would need to test that use against a sample with clear ownership and permitted uses.

Company size and payroll

DataPayouts calculated the figures from the Census Bureau's 2022 Statistics of U.S. Businesses. The 2022 release remained the latest SUSB data available on October 7, 2026.

Census reported 9,631 employer firms in the industry. The group with at least 10 enterprise employees included 479 firms in the 10-19 band, 293 in the 20-99 band, 90 in the 100-499 band and 63 in the 500+ band. Together, these 925 firms represented 9.6% of the total.

Their brokerage establishments employed 57,696 people and paid $5.73 billion in annual payroll. The 20-99 group accounted for 10,416 employees and $871 million in payroll.

In 2022 U.S. mortgage and nonmortgage loan brokerage data, the 10+ paid-employee enterprise group accounted for 9.6% of employer firms, 75.9% of paid employment and 82.0% of annual payroll. — We calculated these shares from 2022 Census SUSB data using 2017 NAICS code 522310. Each percentage uses the matching industry total. The 10+ group uses paid employment across the whole enterprise. Employee counts include part-time staff and exclude independent contractors. Businesses with no paid employees are excluded.
Sources: U.S. Census Bureau, 2022 SUSB

We calculated these shares from 2022 Census SUSB data using 2017 NAICS code 522310. Each percentage uses the matching industry total. The 10+ group uses paid employment across the whole enterprise. Employee counts include part-time staff and exclude independent contractors. Businesses with no paid employees are excluded.
Measure	Share in enterprises with 10+ paid employees (%)
Employer firms	9.6044
Paid employment	75.9128
Annual payroll	82.0241

Payroll gives an indication of business activity. Census does not measure the share of work that produced reusable records, the records retained or their licensing price.

The category, NAICS 522310, includes mortgage and nonmortgage loan brokers. A separate mortgage-only count cannot be derived from these figures.

Census also measures size across the entire enterprise, including operations outside brokerage. Employees at a brokerage location may therefore belong to a much larger enterprise. The size band should be read alongside the industry's employment count.

Employment and wage changes

The Bureau of Labor Statistics QCEW series provides more recent employment data. Annual-average private-sector employment fell from 128,336 in 2021 to 83,658 in 2025, a decline of 34.8%.

From 2024 to 2025, employment increased by 0.3%, while nominal annual wages increased by 11.3% to $10.65 billion. Commissions, bonuses and other compensation changes can affect that comparison. The series covers compensation across occupations, so it cannot establish the cost of one task.

A buyer assessing brokerage records would need to specify the task it wants to improve. That might be document extraction, income calculation, identification of missing evidence or review of a revised application. The required fields and examples would depend on that task.

Application outcomes

DataPayouts extracted 2025 application outcomes from the FFIEC/CFPB HMDA data API. We counted 11.77 million records in the five main application-outcome categories, of which 6.83 million resulted in originated loans.

The other 4.94 million records, or 42.0%, concerned applications that were denied, withdrawn, closed for incompleteness or approved but not accepted.

Of 11,766,671 HMDA application records in 2025, 42.0% did not become originated loans. The outcomes were originated loans (58.0%), approved but not accepted (3.6%), denied (18.0%), withdrawn (14.5%) and closed for incompleteness (5.9%). — We used the June 2, 2026 snapshot of 2025 HMDA data from all reporting financial institutions, covering all reported loan types, purposes and dwellings. The denominator includes action codes 1-5 and excludes purchased loans and separate preapproval records. A borrower may appear in several records.
Sources: FFIEC / CFPB, 2025 HMDA snapshot

We used the June 2, 2026 snapshot of 2025 HMDA data from all reporting financial institutions, covering all reported loan types, purposes and dwellings. The denominator includes action codes 1-5 and excludes purchased loans and separate preapproval records. A borrower may appear in several records.
Action code	Outcome	Application records	Share of codes 1-5 (%)
1	Loan originated	6827891	58.0274
2	Application approved but not accepted	417961	3.5521
3	Application denied	2115090	17.9753
4	Application withdrawn	1710723	14.5387
5	File closed for incompleteness	695006	5.9066

The calculation covers all reporting financial institutions and loan types. It cannot be used as a mortgage-broker conversion rate. A borrower may have made more than one application, and the calculation excludes purchased loans and the separate preapproval categories.

These outcomes provide a reason to examine records from unsuccessful applications as well as funded loans. An archive limited to funded loans excludes the work on cases that ended before origination.

A reviewer could link the missing document, the professional correction and the relevant lender condition to the final decision. A buyer could use reviewed examples to test whether a model identifies the same problem or checks the correction correctly.

This is a potential use that requires testing. Historical decisions also need review for errors and bias before they are used as training or evaluation labels.

Additional information beyond public data

The CFPB already publishes extensive loan-level HMDA data, with changes to protect privacy. A buyer considering another dataset would need to know what additional information it provides.

One possible collection would pair source documents with extracted fields, expert corrections and final resolutions. Each example would also identify the rule that applied at the time. Dates and rule versions are necessary when lending requirements have changed.

The number of files alone is a poor measure of this collection. Duplicate documents and missing links can reduce the number of complete examples available for a buyer's task.

There are existing products for some of these tasks. In a May 2026 ICE case study, wholesale lender Equity Prime Mortgage reported an increase from around two loans per underwriter per day to more than five. The lender attributed this to AI-assisted validation and changes in its working process. This vendor-published account concerns one customer, and other firms would need to assess their own results.

ICE also licenses mortgage and property data. That service confirms a commercial use for some mortgage data. Neither the service page nor the case study provides an offer for a brokerage's historical records.

Licensing terms and costs

A price assessment requires a defined use, a representative sample, established rights and agreed quality standards. It must also account for preparation and delivery costs. Public payroll figures cannot supply those inputs.

Mortgage records may contain credit reports, tax documents, account details and other sensitive borrower information. FTC Safeguards guidance covers mortgage brokers, and Regulation P restricts disclosure and reuse.

The seller needs to check whether its permission to process a loan also covers the proposed AI use. Removing names alone may leave contractual restrictions or information that identifies a borrower.

A proposed payment should cover legal review, preparation, quality checks, secure delivery and continuing obligations. The expected proceeds remain uncertain until a buyer has made an offer against a defined scope.

A later review can compare the periods covered by each CRM, loan-processing, document and messaging system. It can also check record counts, export options and gaps after software changes. The relevant question is whether messages, condition changes and outcomes can still be linked to the same case.

Start the DataPayouts fit check with your company, systems and record categories. The form does not request borrower files or system connections. A detailed inventory can follow if further review is appropriate.

Data sources and limitations

Census SUSB 2022 uses 2017 NAICS 522310. Enterprise size includes employment across the associated company. March paid headcount includes full-time and part-time workers without conversion to full-time equivalents.

We calculated the 10+ group by adding the 10-19, 20-99, 100-499 and 500+ bands. Payroll is nominal annual cash payroll. Census adjusted published employment and payroll values to protect confidentiality.

BLS QCEW uses annual-average covered employment, so we report it separately from Census. The HMDA calculation uses 2025 action codes 1-5 across all reporting institutions. These sources support comparisons of company size, employment and application outcomes. Dataset eligibility and price require a separate review.