Future of banking? AI multi-agent systems streamline fraud detection and credit analysis

At the core of the study is the concept of agentic AI systems, where multiple AI agents with specialized roles work together under a structured multi-agent collaboration framework.

The financial services industry has long been driven by data, analytics, and automation. However, with the rise of large language models (LLMs) and multi-agent AI systems, a new frontier of autonomous decision-making is emerging. These agentic AI systems go beyond traditional AI applications by enabling AI agents to collaborate, specialize, and work together in structured workflows to solve complex problems.

A recent study titled "Agentic AI Systems Applied to Tasks in Financial Services: Modeling and Model Risk Management Crews", authored by Izunna Okpala, Ashkan Golgoon, and Arjun Ravi Kannan, published by Discover Financial Services Inc., explores the transformative impact of AI-driven agentic systems in financial modeling and risk management. The study introduces a structured AI framework where multiple agents collaborate to perform financial modeling tasks, assess risks, and ensure regulatory compliance.

How agentic AI systems work in finance

At the core of the study is the concept of agentic AI systems, where multiple AI agents with specialized roles work together under a structured multi-agent collaboration framework. In financial services, these AI agents are assigned to perform specific tasks such as data exploration, feature engineering, model selection, hyperparameter tuning, model training, and evaluation. Unlike traditional AI, which relies on a single model to handle all tasks, agentic AI systems divide the workload among specialized agents, enabling faster execution, improved accuracy, and more adaptable decision-making.

The study focuses on two primary applications in financial services: financial modeling crews and model risk management (MRM) crews. The financial modeling crew consists of agents responsible for data preparation, model training, and optimization, while the MRM crew ensures compliance, performs validation, and assesses model risks. These AI teams function similarly to human teams, with managers assigning tasks to different AI agents, ensuring a smooth and structured workflow.

One of the key innovations introduced in the study is the ability of AI agents to store memory, retrieve past actions, and learn from interactions. The memory module enables AI agents to recall previous tasks, ensuring consistency and reducing redundant computations. This capability enhances the accuracy and efficiency of financial models, making them more reliable for credit risk analysis, fraud detection, and credit approval processes.

Applications of agentic AI in financial modeling

The study demonstrates the effectiveness of agentic AI systems in three key financial applications: credit card fraud detection, credit card approval prediction, and portfolio credit risk modeling. Each use case highlights how AI agents can work together to improve financial decision-making.

In the credit card fraud detection scenario, the AI modeling crew was tasked with analyzing over 280,000 transactions, identifying fraudulent patterns, and selecting the best machine learning model for fraud detection. The feature engineering agent identified missing values and performed dataset transformations, while the model selection agent used GridSearchCV to find the best-performing algorithm. The results showed a fraud detection accuracy of 94.39%, outperforming traditional approaches used in industry.

For credit card approval prediction, AI agents collaborated to analyze a dataset containing 36,457 applications, assessing applicant creditworthiness based on factors such as income, employment status, and credit history. The AI-driven model achieved an accuracy of 95.48%, surpassing previous benchmarks while improving processing speed and efficiency.

The portfolio credit risk modeling use case involved analyzing loan data to predict default probabilities. Using the agentic AI framework, the modeling team processed over 32,500 loan applications, identifying key risk factors and ensuring compliance with financial regulations. The XGBoost model selected by the AI agents achieved a 95.37% accuracy rate, demonstrating the system’s effectiveness in predicting credit risk.

Ensuring model risk management and compliance

Beyond model development, model risk management (MRM) is a crucial component of financial AI applications. The study introduces an MRM crew designed to ensure that financial models are accurate, unbiased, and compliant with industry regulations. This specialized team includes AI agents responsible for documentation review, model replication, conceptual soundness evaluation, and outcome analysis.

The MRM crew operates as a safeguard by validating models before they are deployed. One of the major challenges in financial modeling is ensuring that models remain robust in the face of changing financial conditions. To address this, the outcome analyzer agent in the MRM crew simulates extreme market conditions by introducing shifted and adversarial inputs. These stress tests evaluate how well financial models perform under unexpected scenarios, such as economic downturns or data anomalies.

Additionally, the documentation compliance checker agent ensures that all AI-driven decisions align with regulatory frameworks by validating model outputs against predefined compliance guidelines. This reduces regulatory risks and enhances transparency, making AI-driven financial decision-making more accountable and auditable.

Future of AI-driven finance

The introduction of agentic AI systems represents a major shift in how financial services approach AI-driven decision-making. By allowing AI agents to collaborate, specialize, and self-improve, financial institutions can achieve higher accuracy, faster processing speeds, and improved regulatory compliance.

However, as the study highlights, several challenges remain. One major concern is AI interpretability, as multi-agent systems can introduce complexity in understanding decision-making processes. To address this, researchers emphasize the importance of human oversight, ensuring that AI-driven decisions remain transparent and aligned with ethical financial practices.

Looking ahead, future research could focus on self-improving AI agents, where models adapt dynamically based on market conditions and continuously refine their decision-making abilities. Additionally, the development of crew-generating AI systems - where AI agents create and manage their own specialized teams - could revolutionize automated trading, investment management, and risk assessment in the financial sector.