The field of credit decisioning is anything but static. As new technologies emerge, underwriters continuously innovate to enhance critical aspects of the process, from predicting the probability of default to pricing loans and effectively managing customers throughout the loan lifecycle.

Today, Generative AI (GenAI) is transforming credit decisioning with unprecedented speed, and its real-world impacts are only beginning to unfold. In this article, Romain Mazoué, Group Chief Risk Officer at Younited, explores the evolution of credit decisioning, delves into Artificial Intelligence (AI) and GenAI use cases shaping the industry, and provides a glimpse into the future as these technologies become integral to lenders’ operations.

The credit underwriting process has evolved significantly over time, from manual, subjective evaluations to highly automated and data-driven systems. Below is a detailed chronology of the decision-making system for credit underwriting:

Manual Underwriting & expert judgment
In the early days of credit underwriting, decisions were made manually by credit managers or underwriters. These individuals would review an applicant's financial information, such as income, debts, and personal characteristics, to assess their creditworthiness. The process was highly subjective, potentially influenced by biases such as race, gender, or personal judgments about an applicant's character. This manual approach was time-consuming and prone to inconsistencies.

Introduction of credit scoring algorithms
The first major shift in credit underwriting came with the development of credit scoring models in the 1950s. FICO (Fair Isaac Corporation), founded in 1956 by Bill Fair and Earl Isaac, introduced one of the first credit score models for lenders. These early models were designed for specific businesses and were not universally applicable across different industries. Three other main players in the credit scoring field were also created during this period.

In the 70’s, the Fair Credit Reporting Act (FCRA) started to regulate the collection of personal credit information and customers’ access to credit reports in the USA.

By the 1980s, credit bureaus began to digitize consumer data, and credit scoring algorithms became more widely adopted. In 1989, FICO released its first "universal" credit score model that could be used by any lender through credit reporting agencies such as Equifax, Experian or TransUnion. This marked the beginning of a more standardized and objective approach to assessing credit risk

At this time, fixed pricing was the norm despite significant differences in customer costs (specifically cost of risk), with loan APR only being changed by expert judgment.

Broader usage of credit score
In the 1990s, more complex underwriting systems began to emerge, initially with credit card products, before gaining significant traction across other lending products, particularly in the mortgage industry. Pioneered by US government-sponsored enterprises like Fannie Mae and Freddie Mac, these systems aimed to streamline the underwriting process by automating the evaluation of borrower creditworthiness based on predefined criteria such as credit scores and debt-to-income ratios. These new underwriting systems significantly reduced the time required for loan approvals and improved consistency in decision-making.

In the early 2000s, advancements in computing power and data analytics led to the development of more sophisticated statistical models, such as logistic regression, to predict default risk more accurately. Banks began hiring data scientists and statisticians to develop these models, which were trained on large datasets containing hundreds of variables related to borrower behavior. Credit decision engines gradually became integral to financial institutions' tech stacks, enabling real-time decision-making based on a wide range of borrower data.

In parallel, multiple events contributed to fostering wider usage of credit scoring and an improvement in loan pricing:

1. US regulation & credit scoring availability
The US Congress passed the Fair and Accurate Credit Transaction Act in 2003. The act allowed consumers to request and obtain a free credit report once every 12 months from each of the three nationwide consumer credit reporting companies (Equifax, Experian, and TransUnion). However, the adoption of this was slow due to the lack of marketing and consumer awareness. 

2. Basel II Regulation
The Basel II Internal Ratings-Based Approach (IRBA) requirements were officially released in June 2004. This framework allowed banks to use their own internal ratings to assess credit risk and calculate regulatory capital requirements, provided they met certain minimum conditions and received approval from their national supervisors.

By allowing banks to develop and use their own models, the IRBA incentivized institutions to improve their risk management practices. Banks were encouraged to adopt more sophisticated modeling techniques to better assess and manage their credit risk exposures through risk parameters estimation. These models allowed for a more dynamic assessment of credit risk, incorporating both observable and unobservable risk factors.

The IRBA introduced stringent regulatory requirements for model validation and approval. Banks had to demonstrate that their internal models were robust, accurate, and consistent with regulatory standards. This led to increased scrutiny of modeling practices and encouraged continuous improvement in model development and validation processes.

3. Capital One & risk-based pricing
Early innovators went beyond credit scoring to undertake sophisticated modeling of multiple additional factors, including churn, marketing efficiency, customer lifetime value, and profitability scoring.

In the 1990s, Capital One developed an innovative targeted marketing approach, focusing on customer profitability analysis, to achieve remarkable success as a leading credit card issuer. It maintained its competitive edge by investing in infrastructure and personnel and continually enhancing its expertise through a 'test-and-learn' methodology. Capital One’s “information-based” strategy consisted of deploying an advanced risk-based pricing approach.

The next leap in credit underwriting came with the integration of Artificial Intelligence (AI) and Machine Learning (ML) technologies. These tools allowed lenders to analyze vast amounts of both traditional and non-traditional data (e.g., social media activity or utility payments) to make more nuanced credit decisions.

AI-driven systems could assess borrower reliability beyond conventional metrics like credit scores, broadening access to credit for individuals with limited credit histories. AI also improved risk assessment by identifying patterns in borrower behavior that traditional models might miss. However, several challenges appeared with the new generation of models, specifically the lack of transparency and the potential risk of model biases leading to discrimination. 

The growing usage of alternative data has fueled the emergence of AI models. By better leveraging large datasets of non-linearly correlated and diverse data, AI models have shown their superiority to traditional statistical models. For credit decisioning, alternative data usually include (i) open banking data, (ii) utilities and rental payment history, (iii) open admin data, and (iv) digital footprint. This approach has proven particularly beneficial for underserved populations, such as recent immigrants or gig economy workers, who might otherwise be excluded from traditional lending systems.

Today, many lenders use real-time data analytics to make instant decisions on loan applications. Automated systems can now integrate alternative data sources—such as rental payment history or utility bills—to assess borrowers who might not have a robust credit history. This shift has made lending more inclusive while maintaining rigorous risk management practices.

Additionally, modern decisioning systems now allow for greater agility in decision-making through configurable rules that can be adjusted without extensive reprogramming. SaaS decisioning platforms, for instance, enable lenders to design and test highly segmented, real-time decision workflows and models tailored to their customer base.

While AI models have existed for decades—dating back to the 1980s with logistic regressions—they were limited by smaller datasets and less computational power. In contrast, the foundations of GenAI models are relatively recent, beginning with the development of Transformer models by Google in 2017 and the release of BERT in 2018. BERT marked a breakthrough in language processing, revolutionizing the ability to understand context through its transformer-based architecture

In November 2022, OpenAI released ChatGPT, which quickly became a groundbreaking product due to its ability to generate human-like text. Within two months, it gained over 100 million users, making it the fastest-growing consumer application in history. Two years later, GenAI models have continued to improve, with multiple new major players competing with ChatGPT: Gemini, Android, and Mistral, for example.

Additionally, hardware capabilities have followed the Gen AI hype, with Nvidia becoming the largest corporate valuation in the world—going from $500 billion to $3,200 billion!

These examples illustrate how fast GenAI has moved from a novel technology to a widely adopted tool across various industries, enhancing efficiency and personalization in numerous applications.

Over the past years, structural trends have reshaped how credit decisions are processed through three phases. We first shifted from deterministic behavior based on expert rules to probabilistic behavior with machine learning and deep learning models, which can learn from data without being explicitly programmed. Then, the third phase is the GenAI models that use content to create new content and process complex workflows. These three generations of models can be used together in a complementary way.

Another noticeable trend is the switch from proprietary software such as SPSS or SAS to open-source tools and libraries (R, python). The open-source community has fostered innovation and continuous improvement by making new tools and algorithms, such as XG Boost or CAR, available.

Zooming in on the second and third phases mentioned above, in the context of credit decisioning and loan applications, AI and Generative AI (GenAI) offer distinct capabilities that complement each other to enhance the loan approval process.

Traditional AI models think as a deductive brain to make decisions. They can be used in multiple steps of a credit decisioning framework, for example:

GenAI models work similarly to the inductive brain, as they can demonstrate creativity and problem-solving approaches. The new set of capabilities brought by GenAI models can be summarized as follows:

While this new set of new capabilities is generic, we see more and more examples of successful Gen AI implementations to tackle risk management topics:

There are multiple domains where both types of AI complement each other, such as recommendation engines and demand forecasting. Importantly, GenAI will complement traditional AI opportunities, not replace them!

In summary, while traditional AI focuses on predictive analytics, fraud detection, and automation to streamline credit decisioning processes, GenAI brings advanced capabilities in handling unstructured data, enhancing personalization, and generating synthetic data for model training.

These complementary technologies enable more accurate, efficient, and customer-centric lending practices.

Model risk is perhaps the flipside of increasing sophistication and the growing complexity of credit decisioning models with more parameters. Challenges such as lack of transparency, the risk of discriminatory bias, sensitivity to data quality, and limited explainability become more pronounced, posing significant concerns—particularly in the highly regulated financial industry.

The European Union's Artificial Intelligence Act (AIA) has significant implications for credit decisioning models, as it classifies AI systems used for credit scoring and creditworthiness assessments as "high-risk." The AIA imposes stringent requirements on these high-risk systems, including robust risk management, transparency, accuracy, human oversight, and measures to prevent discrimination or unfair outcomes. These rules aim to mitigate potential risks like bias or lack of interpretability while ensuring that AI-driven credit models operate responsibly. 

By fostering accountability and fairness, the AIA seeks to balance the benefits of AI - such as improved efficiency and inclusivity - with the protection of fundamental rights and consumer trust in financial decision-making processes. However, this may also pose challenges for European model developers as compliance challenges and cost might hinder innovation and model performance.

GenAI models are known to sometimes provide unexpected outcomes. For highly regulated and sensitive processes such as credit decisioning, this limitation can represent a challenging roadblock. Innovations such as Retrieval-Augmented Generation (RAG) could address challenges like hallucinations, ensuring that AI-generated outputs are grounded in factual data. Moreover, GenAI's ability to generate transparent explanations for decisions could improve trust among consumers and regulators while providing rejected applicants with more humanistic feedback.

As open finance expands, GenAI will play a pivotal role in integrating data from multiple financial institutions to provide comprehensive credit assessments. With improved reliability and gradual adaptations in internal workflows and software to accommodate this new generation of models, we anticipate significant transformations in operational processes. GenAI models and agents will become increasingly autonomous in the credit decisioning process through the possibility of dynamically managed safety nets triggered by human-in-the-loop checks or controls.

GenAI-based credit decisioning could lead to more accurate and inclusive lending decisions, expanding access to credit for underserved populations while reducing default risks. However, widespread adoption will require financial institutions to address risks related to bias, governance, and system reliability.GenAI models have the power to not only streamline credit decisioning but also foster innovation, inclusivity, and efficiency in the financial sector, setting new standards for customer-centric lending practices.

Disclaimer
This information provided in this article does not, and is not intended to constitute professional advice; instead, all information, content, and material are for general informational and educational purposes only. Accordingly, before taking any actions based upon such information, we encourage you to consult with the appropriate professionals.