Implications for risk assessment and compliance

Credit underwriting serves as the fundamental gatekeeper of the global financial system. It is the process through which lenders evaluate a borrower’s creditworthiness and determine whether to extend credit. Historically, the evaluations were based on the “5 Cs” of credit: Character, Capacity, Capital, Collateral, and Conditions. Traditional underwriting rely heavily on manual verification, where human underwriters examine a limited set of structured data points, primarily sourced from historical credit bureau reports such as FICO scores, income statements, and tax returns. While these methods provided a stable framework for decades, they are increasingly defined by their inherent constraints: they are time-consuming, influenced by human subjectivity, and exclusionary. By relying solely on comprehensive credit histories, traditional models often overlook the “credit invisible” populations, such as young adults and immigrants with limited banking records.

Source: Deloitte Digital Transformation Report: AI vs. Traditional Lending Productivity, 2026

The motivation for adopting Artificial Intelligence (AI) in this space is therefore not only an upgrade in speed, but a shift toward precision and financial inclusion. As noted by the Bank for International Settlements (BIS), traditional linear models struggle to capture the complex, non-linear relationships in modern financial behavior. Using machine learning, financial institutions can analyze alternative data such as utility payments and digital behavior to move beyond static reports and adopt predictive risk models that assess risk dynamically and quickly. According to McKinsey & Company, in their report “Designing next-generation credit-decisioning models”, this evolution can reduce default rates by up to 25% while expanding the eligible borrower pool, proving that AI is a driving force for a more equitable financial landscape.

However, implementing these technologies within the European Union introduces an additional layer of regulatory responsibility. Under the EU AI Act (Regulation 2024/1689), AI systems used to evaluate the creditworthiness of natural persons are classified as “high-risk” systems, as outlined in Annex III. As a result, European financial institutions face stricter obligations than those in less regulated environments. AI models therefore cannot be used as fully non-transparent systems. Instead, banks must ensure that such systems meet requirements related to transparency, accountability, and risk management, shaping how AI can be responsibly integrated into credit decision-making processes.

Source: Industry Benchmark for Explainable AI (XAI) in Credit Scoring 2025

The adoption of Artificial Intelligence (AI) in credit underwriting represents more than a technological improvement in processing speed. It reflects a structural shift toward greater precision, adaptability, and financial inclusion in credit risk assessment. Traditional credit scoring methods, which rely heavily on linear statistical models, have long struggled to capture the complexity of modern financial behavior. According to the Bank for International Settlements (BIS), these models are limited in their ability to represent non-linear relationships and evolving patterns in borrower behavior. Machine learning techniques address these limitations by enabling financial institutions to analyze a wider range of data sources, including alternative data such as utility payment histories, transaction records, and digital behavior. This allows lenders to move away from static and backward-looking credit reports and toward predictive models that assess risk dynamically. 

Within the European Union, however, the deployment of AI-based credit underwriting systems is shaped by a particularly strict regulatory environment. The regulation requires compliance with clear standards related to transparency, data quality, and human oversight. In practice, this means that institutions must ensure their training datasets are representative and free from demographic biases in order to avoid automated discrimination. Concerns around biased algorithmic outcomes have already emerged in real-world cases, such as the Apple Card investigation, where credit decisions were alleged to disadvantage women. These developments underline that in the European context, the future of underwriting is not only about improving predictive accuracy, but also about ensuring responsible AI use and safeguarding individuals’ rights, including the right to receive a meaningful explanation when credit is denied.

Understanding the impact of AI in underwriting also requires examining the scale and nature of the data being used. Traditional credit scoring models typically rely on a limited number of variables, often between 20 and 30, primarily sourced from credit bureau reports. In contrast, AI-driven models developed by fintech firms such as Upstart and Zest AI may analyze over 1,000 data points. These models are capable of identifying correlations and behavioral patterns that are difficult for human analysts or conventional statistical approaches to detect. However, the increased complexity of these systems introduces challenges related to interpretability. In regulated financial environments, lenders cannot justify adverse credit decisions by simply stating that an algorithm made the decision. Regulatory guidance from bodies such as the Consumer Financial Protection Bureau in the United States requires lenders to issue Adverse Action Notices that clearly explain the reasons for denying credit. This regulatory pressure has contributed to the growing importance of explainable AI techniques.

Source: Industry Benchmark for Explainable AI (XAI) in Credit Scoring 2025

Explainable AI tools aim to make complex model outputs understandable without significantly reducing predictive performance. One of the most widely adopted methods is SHAP, or Shapley Additive Explanations, which assigns contribution values to individual input features and shows how each factor influenced a specific credit decision. This enables lenders to explain outcomes in concrete terms, such as identifying a recent decline in cash flow as a key reason for denial. Local Interpretable Model-agnostic Explanation (LIME), offers a complementary approach by approximating complex models with simpler, interpretable ones around individual predictions. These tools are particularly valuable in regulatory contexts, where transparency and accountability are essential.

Modern credit underwriting systems rely on a broad range of analytical models, spanning from traditional statistical methods to advanced machine learning architectures. Logistic regression remains a foundational tool due to its interpretability and regulatory acceptance, but institutions increasingly adopt more sophisticated approaches to enhance predictive performance. Ensemble methods, including Random Forests and gradient boosting algorithms such as XGBoost, LightGBM, and CatBoost, are widely used because they combine multiple models to reduce overfitting and improve accuracy. Beyond ensemble techniques, Artificial Neural Networks are employed to capture complex non-linear relationships in borrower behavior. Convolutional Neural Networks are applied to structured data to identify hierarchical patterns, while Recurrent Neural Networks, particularly Long Short-Term Memory models, are effective for analyzing sequential time-series data such as transaction histories and cash flow dynamics. Graph Neural Networks further extend these capabilities by modeling relationships between borrowers, accounts, and financial entities, enabling the detection of fraud networks, correlated defaults, and systemic risks. In parallel, Natural Language Processing models such as BERT are increasingly used to analyze unstructured text, including financial news, customer communications, and internal notes, allowing lenders to extract qualitative risk indicators that traditional models often overlook.

The effectiveness of these AI-driven systems depends heavily on the breadth and quality of the data they consume. While traditional sources such as credit reports, loan applications, and financial statements remain essential, they are now commonly supplemented with alternative data to improve predictive power and reach individuals with limited or non-traditional credit histories. Transactional and utility data, including rental payments, utility bills, and cash flow records, provide valuable insights into financial discipline. Digital footprint data, such as telecommunications usage, digital wallet activity, and e-commerce behavior, offers real-time signals of economic stability. Some lenders also experiment with behavioral and psychometric data, such as typing patterns or application completion speed. Although controversial, these data sources may enhance predictive accuracy when used responsibly and within regulatory boundaries.

AI-based underwriting differs fundamentally from traditional credit scoring in its ability to evolve continuously and uncover latent patterns without explicit programming. Unlike conventional scorecards that assume linear relationships, AI models identify complex interactions among thousands of variables, such as the combined effects of digital behavior and repayment history. Techniques such as autoencoders support advanced feature engineering by reducing high-dimensional data into compact representations that highlight the most relevant predictors of default risk. Additionally, reinforcement learning and online model updates allow underwriting systems to adapt lending rules in response to real-world outcomes and changing macroeconomic conditions.

Despite these benefits, the use of AI in credit underwriting raises significant ethical, privacy, and accountability concerns. Machine learning models are highly dependent on the data used to train them, and when historical data reflects existing socioeconomic inequalities, AI systems may unintentionally reproduce or amplify these biases. Even when protected characteristics such as gender or ethnicity are excluded, proxy variables may still lead to discriminatory outcomes. Privacy concerns also arise from the growing reliance on alternative data sources, as questions emerge around consent, proportionality, and data protection. Furthermore, the automation of credit decisions complicates accountability, making it more difficult to determine responsibility when errors or unfair outcomes occur.

These challenges have prompted regulatory responses across jurisdictions. In the European Union, the AI Act establishes comprehensive requirements for high-risk AI systems, including representative training data, regular audits, and continuous human oversight. The regulation explicitly guarantees individuals the right to a meaningful explanation of automated credit decisions. Similar regulatory momentum exists in the United States, where the Consumer Financial Protection Bureau has reinforced the application of fair lending laws to algorithmic underwriting. Under the Equal Credit Opportunity Act, lenders must provide specific and actionable reasons for adverse credit decisions, and regulatory authorities have made clear that algorithmic complexity does not exempt institutions from compliance. International bodies such as the Basel Committee on Banking Supervision and the Financial Stability Board have also emphasized the importance of AI governance, model validation, and ongoing monitoring to mitigate systemic risks.

To meet these regulatory and ethical expectations, financial institutions increasingly rely on explainable AI tools and, in some cases, inherently interpretable models such as decision trees, rule-based systems, and generalized additive models. Counterfactual explanations are also gaining attention, as they provide borrowers with actionable guidance on what changes could lead to credit approval. Explainability is therefore not merely a regulatory requirement, but a fundamental aspect of responsible AI design.

Looking ahead, the future of AI in credit underwriting will likely be shaped by regulatory convergence, technological innovation, and evolving societal expectations. Advances in federated learning and privacy-preserving computation may enable institutions to develop robust models while protecting sensitive data, while the integration of real-time economic indicators may further transform underwriting into a forward-looking and adaptive process. Ultimately, AI presents both an opportunity and an obligation for financial institutions. While it offers powerful tools to improve risk assessment and expand access to credit, these benefits can only be realized through sustained commitments to fairness, transparency, and accountability. The success of AI-driven underwriting will therefore be measured not only by predictive accuracy and efficiency, but by the extent to which these systems earn and maintain public trust.

Written by:

• Domenico Agostino
• Keti Aznaurashvili
• Nicolò ballabio
• Matilde Sommese