AI Agents in Finance and Fintech: A Scientific Review of Agent-Based Systems, Applications, and Future Horizons

In one sentence

Summary

The paper addresses the rapid integration of autonomous AI agents into financial systems, positioning them as a distinct paradigm beyond traditional statistical models. It aims to consolidate knowledge on agent-based methodologies, specifically focusing on reinforcement learning, multi-agent systems, and LLM-driven frameworks. The review highlights the shift towards agents capable of reasoning, learning, and autonomous action in high-stakes environments, while identifying critical gaps in standardized evaluation, scalability, and systemic risk management. It serves as a comprehensive survey for researchers and practitioners navigating the technical and ethical complexities of agentic finance. The methodology involves a systematic literature review of peer-reviewed publications and industry reports, categorizing applications into algorithmic trading, credit risk, fraud detection, robo-advisory, and RegTech. The authors analyze architectural frameworks, contrasting rule-based, utility-based, and BDI agents with modern deep reinforcement learning and LLM-based systems. Key experiments and findings cited include the use of LLMs to simulate investor behavior (StockAgent), the superiority of multi-agent vertical structures for complex risk analysis, and the performance of CatBoost in credit scoring. The review also examines the use of Graph Neural Networks for fraud detection and the limitations of LLMs in high-frequency trading due to latency and reliability issues. Findings indicate that while AI agents improve efficiency and accuracy in areas like fraud detection (40% false positive reduction) and credit risk, they introduce significant challenges. LLM-based agents face issues with hallucinations, bias, and prompt sensitivity, making them less suitable for deterministic, low-latency trading compared to specialized RL agents. Ethical and regulatory concerns, including algorithmic bias, lack of explainability, and potential for AI-driven collusion, are prominent. The paper concludes that responsible AI deployment requires human-in-the-loop oversight, robust stress testing, and the development of explainable, neuro-symbolic architectures to ensure trust and compliance in financial ecosystems.

Research problem

The paper addresses the lack of a consolidated understanding of autonomous AI agents in finance, distinguishing them from traditional AI. It seeks to map the current state of agentic architectures, evaluate their performance across key financial domains, and identify the technical and ethical challenges hindering their widespread, responsible adoption.

Core contributions

• Synthesizes agent-based methodologies in finance, contrasting traditional RL/MAS with emerging LLM-driven agents.
• Categorizes and evaluates AI agent applications across five core domains: trading, credit, fraud, advisory, and RegTech.
• Analyzes architectural frameworks, including rule-based, BDI, and neuro-symbolic agents, highlighting their financial suitability.
• Identifies critical challenges such as LLM hallucinations, systemic risk, algorithmic bias, and the need for explainability.
• Proposes future research directions focusing on responsible AI, standardized evaluation, and hybrid neuro-symbolic systems.

Method

• Systematic literature review using PRISMA-like criteria for selection and analysis.
• Categorization of agents by architecture: rule-based, utility-based, BDI, RL, and LLM-based.
• Cross-comparison of agent capabilities across trading, credit, fraud, advisory, and RegTech.
• Analysis of specific frameworks like StockAgent, FinRobot, and FinSearch for LLM applications.
• Evaluation of multi-agent system configurations (vertical vs. horizontal) for task efficiency.
• Assessment of ethical implications including bias, fairness, and accountability in agentic systems.

Results & findings

• LLM-based agents can simulate investor trading behaviors but show variability and sensitivity to external factors.
• Multi-agent vertical structures outperform single agents in complex risk analysis tasks.
• AI agents in credit scoring can reduce loan defaults by 20% and improve risk model accuracy by 25%.
• Fraud detection agents using deep learning and GNNs can decrease false positives by 40%.
• LLM-advisors match human performance for some profiles but struggle with complex risk-taking cases.
• Users may be insensitive to advice quality, preferring extroverted LLM personas even with worse advice.
• RL agents face challenges with sample inefficiency and training instability in dynamic markets.
• RegTech agents automate compliance checks but require human oversight for complex regulatory interpretation.

Data & experiments

• Systematic review of peer-reviewed literature from IEEE, ACM, Springer, Elsevier, and arXiv.
• Analysis of specific case studies including StockAgent, FinRobot, and FinSearch frameworks.
• Evaluation of performance metrics in credit scoring (recall, default reduction) and fraud detection (false positive rates).
• Assessment of LLM capabilities in simulating trading behaviors and providing financial advice.
• Review of multi-agent system performance in complex financial tasks like risk analysis and model risk management.

Limitations & caveats

• LLM-based agents suffer from hallucinations, bias, and prompt sensitivity, limiting reliability in critical tasks.
• RL agents face challenges with sample inefficiency, training instability, and overfitting to historical data.
• Multi-agent systems introduce complexity in orchestration and potential for emergent, unpredictable behaviors.
• Lack of standardized evaluation frameworks for financial AI agents hinders comparability across studies.
• Regulatory and ethical challenges, including bias and accountability, remain significant barriers to deployment.
• High computational costs and latency of LLMs make them unsuitable for high-frequency trading scenarios.

Curator notes

Discovery status: round3_promoted. Source links: Enhancing Investment Analysis: Optimizing AI-Agent Collaboration in Financial Research; Large Language Models in equity markets: applications, techniques, and insights; When AI Meets Finance (StockAgent): Large Language Model-based Stock Trading in Simulated Real-world Environments