How We Test AI Model Outputs
CrossCheck evaluates AI-assisted underwriting through ongoing monitoring and statistical testing designed to identify where fair lending risk may arise in the credit decision process. Our approach does not stop at the model score. We assess AI model outputs, lender-defined configuration thresholds, manual reviews, and final credit decisions to understand how automated underwriting performs in practice and where disparities may emerge. The goal is to help institutions move beyond high-level AI governance discussions and into a clearer, evidence-based view of how AI credit underwriting affects real applicant outcomes.
CrossCheck AI Fair Lending Funnel Analysis℠
At the center of this work is CrossCheck’s AI Fair Lending Funnel Analysis℠, which evaluates applicants as they move from initial model scoring to final credit decision. The funnel-based approach helps determine whether fair lending risk arises from the AI model itself, lender-defined business rules, human discretion, or traditional credit standards. It also emphasizes challenger model benchmarking, using traditional, or benchmark, models as a baseline, enabling institutions to distinguish existing underwriting disparities from those introduced by AI-enabled processes. This methodology supports a more practical fair lending risk assessment by helping institutions see not just whether disparities exist, but where they are created and what may be driving them.
Our AI Fair Lending Testing Framework
- AI Model Output Testing
We analyze AI model outputs to identify whether score patterns may create fair lending risk within an automated underwriting process. - AI Configuration Review
We test score cutoffs, thresholds, and other business logic to determine whether policy design contributes to disparities beyond the model output itself. - Manual Review Analysis
We evaluate the effect of underwriter discretion, overrides, and review-stage decisioning to assess whether human intervention reduces or increases fair lending risk. - Final Decision Testing
We assess final approval and denial outcomes to understand cumulative disparity across the full AI-enabled credit decision funnel. - Challenger Model Benchmarking
We compare AI-enabled outcomes to a benchmark or legacy framework to help isolate whether observed disparities reflect traditional underwriting patterns or changes introduced by AI. - Attribution of Disparity
We provide a structured view of where disparities emerge across funnel layers, supporting documentation, transparency, and targeted file review planning.
Meet Your AI Fair Lending Experts
Rebecca Escario
Ryan Addabbo
Hamza Abourifq
Jonathon Neil
Representative Engagements
AI Fair Lending FAQs
What is AI model output testing?
Fair lending AI model output testing is the ongoing review of how AI-assisted underwriting performs in practice. It looks beyond the model itself to evaluate AI model outputs, configuration thresholds, manual reviews, and final credit decisions to assess fair lending compliance risk.
Why is AI model output testing important for fair lending?
Fair lending risk can arise at different points in an AI-enabled credit process, not just from the model score. Output testing helps identify whether disparities may be linked to the AI model, lender-defined business rules, human discretion, or traditional credit standards so institutions can better understand and manage risk. It is also important because regulators increasingly expect ongoing AI governance, monitoring, and documentation around decisioning, including explainability across each phase of AI model implementation. Regular output testing helps institutions support those expectations by creating a clearer view of how AI underwriting is performing in practice and where additional review or controls may be needed.
Does testing evaluate only the AI model itself?
No. CrossCheck’s approach evaluates the full credit decision-making process, including model outputs, score cutoffs, exclusion rules, manual review activity, and final decisions. That broader view helps institutions identify at what stages AI may influence fair lending disparities within the underwriting funnel.
What does CrossCheck review as part of AI model output testing?
The review can include AI model outputs, AI configuration thresholds, manual reviews, and final credit decisions. CrossCheck also applies statistical testing at different stages of the process to isolate where disparities may be introduced or amplified.
What is AI Fair Lending Funnel Analysis℠?
AI Fair Lending Funnel Analysis℠ is CrossCheck’s framework for evaluating how applicants move through an AI-enabled credit decision process from initial model scoring to final decision. The goal is to isolate where fair lending disparities emerge and support stronger transparency, documentation, and file review planning.
What is challenger model benchmarking?
Challenger model benchmarking uses a legacy, or benchmark, model as a baseline reference point. By establishing baseline disparities using traditional underwriting criteria, institutions can better assess whether disparities in the AI decision process reflect existing risk patterns or are introduced by the use of AI.
Can AI model output testing help distinguish model risk from policy or process risk?
Yes. The framework is designed to evaluate whether fair lending risk arises from the AI model itself, lender-defined business rules, human discretion, or traditional credit standards. That helps institutions separate output-level issues from configuration, governance, policy, or review-stage issues.
Does fair lending AI testing include manual review analysis?
Yes. Manual review is an important part of the AI model output testing analysis because disparities can arise or change after an underwriter reviews documentation, applies overrides, or makes subjective decisions. Testing manual review outcomes alongside model output provides a more complete view of fair lending compliance risk.
What do institutions gain from AI fair lending testing?
Institutions gain a clearer understanding of where disparities emerge in the AI-enabled decision process, and which parts of the funnel may warrant further review. The analysis is intended to support transparency, documentation, file review planning, and stronger AI governance.
How often should AI models be tested for fair lending risk?
AI model output testing is typically most effective as an ongoing process. As underwriting models, decision thresholds, and review practices evolve, institutions benefit from cyclical monitoring that helps identify emerging fair lending compliance risk, track changes in outcomes over time, and support robust governance around AI underwriting.