The Equal Credit Opportunity Act and its implementing regulation, Reg B, place the compliance obligation on the creditor. Not on the creditor's software vendors. Not on the LOS platform. Not on the document extraction tool. The lender who makes a credit decision or takes any adverse action carries the fair-lending obligation, full stop.
This is not a technicality to hide behind. It is a structural fact that every lender's compliance team should understand clearly before evaluating any AI tooling for their origination process, because the analysis starts in different places depending on what stage of the process the tool operates in and what it produces as output.
What follows is a practical breakdown of how ECOA applies to AI-assisted pre-underwriting, what documentation lenders should maintain, and where Maestro sits relative to the risk profile that CFPB guidance has been focused on.
What ECOA actually requires of creditors
ECOA prohibits discrimination against credit applicants on the basis of race, color, religion, national origin, sex, marital status, age, or receipt of income from a public assistance program. Reg B implements ECOA and requires creditors to, among other things, provide written notices of adverse action explaining the principal reasons for denial, respond to applications within specified timeframes, and retain records of applications and actions.
Disparate impact is a central concern in ECOA enforcement. A credit policy that is facially neutral but produces materially different outcomes for applicants of different protected classes can constitute an ECOA violation even without discriminatory intent. This is the doctrine that makes AI-based credit decisioning particularly sensitive: if a model produces differential approval or pricing outcomes by protected class, the facially neutral algorithmic output does not insulate the lender from liability.
The CFPB's 2022 and 2023 guidance on AI in lending has focused precisely on this scenario: models that use proxy variables or non-traditional data inputs that correlate with protected characteristics, and whose decision logic may not be explainable enough to support the adverse action notification requirements of Reg B. This is the right focal area for regulators. It describes a real and significant risk in credit decision AI.
The pre-underwriting preparation stage is different
Document intelligence tools that operate at the pre-underwriting preparation stage do not make credit decisions. They do not score applicants, rank files by approval likelihood, or generate any output that recommends approval, denial, or specific pricing. Their output is a structured summary of what the submitted documents say, along with completeness flags for missing or inconsistent information.
The distinction matters because the ECOA risk that regulators and compliance teams are focused on lives in the credit decision itself, not in the data preparation step that precedes it. An adverse action notice is required when a creditor takes adverse action against a credit applicant. Pre-underwriting document preparation is not a credit decision and does not trigger adverse action obligations.
This does not mean that pre-underwriting AI exists outside compliance considerations entirely. There are still questions a lender's compliance team should ask. But those questions are different from the questions applicable to a credit decisioning model, and the compliance documentation required is correspondingly different.
The questions a lender's compliance team should ask
For a pre-underwriting document extraction tool, the compliance questions center on three areas:
What data does the tool read? A document extraction tool that reads only financial document data (income, assets, employment, property information) does not ingest protected characteristic information as an input. If the tool reads the borrower name or address and uses those as inputs to its extraction model, that raises different questions. Maestro reads financial document fields. It does not use borrower name or address as extraction inputs.
Could the tool's output differentially affect how files are prepared or prioritized for protected-class applicants? This is the subtler question. If a document extraction tool's completeness flags are systematically less thorough for certain document types that correlate with borrower demographics (for example, if it performs worse on 1099 income documentation that self-employed borrowers in certain industries are more likely to use), that differential performance could affect which files reach underwriting with clean preparation and which arrive with unresolved gaps. Lenders should understand their extraction tool's performance across the document types their borrower population actually uses.
What is the documented workflow between the tool's output and the credit decision? The key compliance question is whether there is a clear, documented human decision point between the tool's output and any credit action. For Maestro, the workflow is: extraction and completeness check output goes to the loan officer and processor for review, then the reviewed and verified file goes to the underwriter who makes the credit decision. The underwriter is not receiving a recommendation. They are receiving an organized file. That workflow should be documented by the lender.
What compliance documentation should cover
For lenders using Maestro, we recommend that compliance documentation address the following:
Tool scope and function. Document that Maestro is a document extraction and pre-underwriting preparation tool, not a credit decisioning system. Describe the inputs (document types), outputs (structured field summaries and completeness flags), and the workflow position (pre-underwriting preparation, prior to underwriter review).
Data inputs and exclusions. Document that Maestro's extraction inputs are financial document fields and that borrower demographic information (name, address, race, national origin, sex) is not used as an input to the extraction model.
Human review step. Document the specific workflow step at which a human loan officer or processor reviews Maestro's output before the file proceeds to underwriting. The underwriter who makes the credit decision should be working from a file that has been reviewed and verified by a human at the preparation stage, not directly from Maestro's raw output alone.
Performance monitoring approach. If your lender's compliance program includes ongoing monitoring of AI tools, document the scope of that monitoring for Maestro. At the pre-underwriting stage, relevant monitoring might include reviewing whether completeness flags are being applied consistently across different borrower document profiles.
What Maestro does not do, and why that matters
Maestro does not make credit decisions. It does not score applications. It does not issue or recommend approval, denial, or adverse action. It does not set rates. It does not replace the licensed underwriter who carries the credit decision authority under the lender's regulatory framework.
We designed the tool this way deliberately. Mortgage origination is a highly regulated activity and the most consequential decision in the process, the credit decision, needs to remain with the licensed professionals who carry the responsibility for it. Our job is to make the file preparation step faster and more accurate so that when the underwriter reviews the file, they are working with complete, verified information rather than spending their review time chasing down what a document actually says.
This is not a limitation or a hedge. It is the product design. The loan officer and underwriter remain fully in the loop at every stage where a judgment call or credit decision is being made. Maestro handles the information organization work so they can focus on the judgment work.
A note on the evolving regulatory landscape
CFPB guidance on AI in lending has been evolving, and lenders should expect continued regulatory attention to AI tools throughout the mortgage origination process. The 2023 CFPB circular on the applicability of FCRA adverse action requirements to algorithmic credit scoring and the subsequent interagency statement on AI and banking risk management reflect ongoing regulatory engagement with this area.
As that guidance develops, the compliance analysis for pre-underwriting preparation tools may be further specified. We monitor this guidance and update our documentation recommendations accordingly. What we are confident about is that the structural position of pre-underwriting document extraction, upstream of credit decisions with a documented human review step between extraction output and any credit action, is the right design posture for a tool operating in this regulatory environment.
If your compliance team has specific questions about how Maestro fits within your fair-lending compliance program, we are glad to have that conversation directly. We have documentation on the tool's input scope, the workflow position, and the design choices that keep it upstream of credit decisions, and we can work through how that maps to your compliance documentation requirements.