A Comprehensive Research Report
Executive Summary
The adoption of artificial intelligence (AI) in financial services is accelerating. Industry surveys indicate that the proportion of finance professionals using AI grew from 37% in 2023 to 58% in 2024, and forecasts project that a majority of finance functions will deploy AI-enabled technologies by 2026 [1]. Market research estimates the North American AI-in-finance market at USD 10.6 billion in 2023, with expectations of reaching USD 98.2 billion by 2032 (CAGR 28.1%) [2]. This report surveys the current state of AI in finance, examining technological advances, persistent challenges, and future trajectories. While the evidence suggests transformative potential, careful assessment reveals that many claims rest on limited empirical ground; significant hurdles related to explainability, bias, regulatory fragmentation, and data infrastructure must be addressed.
1. Introduction
The integration of artificial intelligence into financial services has progressed from experimental applications to widespread institutional deployment. Financial institutions now utilize AI across the entire value chain—from back-office automation and risk management to customer-facing advisory and algorithmic trading. The emergence of generative AI, large language models (LLMs), and more sophisticated deep learning architectures has amplified the pace of change, prompting regulatory bodies worldwide to reassess existing governance frameworks. In November 2024, the Financial Stability Board (FSB) noted that “the rapid adoption of AI in finance means that authorities should address information gaps for monitoring, assess the adequacy of current policy frameworks, and enhance supervisory and regulatory capabilities” [3]. This report synthesizes current knowledge in three dimensions: recent technological advancements, key challenges, and future directions, while critically evaluating the robustness of the underlying evidence.
2. Recent Advancements
2.1 Large Language Models and Generative AI in Finance
Domain-specific LLMs have attracted considerable attention. BloombergGPT, a 50-billion-parameter model trained on a financial corpus (FinPile) of 363 billion tokens combined with general data, outperformed comparable open models on financial natural language processing (NLP) tasks while maintaining competitive performance on general benchmarks [4][5]. However, the model’s training cost (utilizing 512 A100 GPUs) and its proprietary nature limit reproducibility and broad access. FinBERT, which fine-tunes BERT on financial texts such as corporate filings and analyst reports, has become a standard tool for financial sentiment analysis, showing superior performance in domain-specific sentiment classification [6][7]. An open-source alternative, FinGPT, employs instruction tuning and data-centric methods to achieve comparable results on sentiment analysis with substantially lower computational requirements (a single RTX 3090 GPU), aiming to democratize access to financial LLMs [8].
GPT-4 and ChatGPT have been applied to financial text analytics. Lopez-Lira and Tang (2023) demonstrated that ChatGPT could assign sentiment scores to news headlines that, in some tests, predicted stock market returns better than earlier models such as GPT-2 and BERT [9]. The MarketSenseAI framework integrated GPT-4 with market data for stock selection, illustrating LLMs’ potential as decision-support tools [10]. In institutional settings, Morgan Stanley deployed GPT-4 to organize internal knowledge for wealth management, while J.P. Morgan explored AI-powered investment advisory capabilities [11]. BondGPT from Broadridge applied generative AI to corporate bond trading [12]. Despite these advances, the robustness of LLM-based trading signals and the risk of hallucinated information remain active research questions.
2.2 Deep Learning in Algorithmic Trading and Market Prediction
Transformer architectures originally developed for NLP have been adapted for financial time series. Models such as the Temporal Fusion Transformer (TFT) and Informer incorporate attention mechanisms and domain-specific inductive biases to forecast financial metrics [13]. The Quantformer framework integrates transformer models with quantitative strategies, while hybrid ConvLSTM architectures combine convolutional and recurrent layers with transformers to capture spatiotemporal patterns [14]. These approaches often show improved forecasting metrics in backtests, though concerns over overfitting and lack of out-of-sample testing persist.
Deep reinforcement learning (DRL) has emerged as a paradigm for algorithmic trading and portfolio optimization. The FinRL ecosystem provides standardized market environments and benchmarks [8]. Several studies report that DRL-based portfolios using Deep Deterministic Policy Gradient (DDPG) achieve higher average returns and Sharpe ratios than traditional Mean-Variance Optimization in specific backtesting periods, but translating these gains to live trading remains challenging due to transaction costs, market impact, and non-stationarity [15]. Multi-agent DRL frameworks have been proposed for multi-asset adaptive trading [16]. The integration of LLM-derived sentiment scores with RL agents, as seen in the Financial News-Driven LLM Reinforcement Learning (FNRL) framework, introduces both opportunities and additional layers of complexity [17].
2.3 Fraud Detection and Anti-Money Laundering (AML)
AI-driven fraud detection has seen significant progress. In controlled settings, machine learning models such as Random Forest, XGBoost, and deep neural networks can achieve detection accuracy exceeding 99% for money laundering patterns [18]. However, these metrics often reflect imbalanced datasets and may not generalize to evolving fraud tactics. Graph neural networks (GNNs) leverage transactional relationships to identify complex layering schemes [19]. Explainable AI techniques like SHAP and LIME are increasingly embedded to meet regulatory requirements for transparency [20][21]. Dynamic graph convolutional networks combined with LSTM layers capture both structural relationships and temporal dynamics, improving adaptability to emerging fraud patterns [22]. The European Central Bank (ECB) noted a sharp increase in AI use cases for fraud detection among significant institutions between 2023 and 2024 [23]. Nonetheless, adversarial attacks and data quality issues remain critical challenges.
2.4 Credit Risk Assessment and Scoring
AI-powered credit scoring has moved beyond traditional logistic regression to incorporate deep learning and alternative data (e.g., utility payments, social media). LSTM-based networks have been adapted for credit scoring with 3D tensor modeling [24]. The EU AI Act classifies credit scoring as a high-risk application, mandating transparency, human oversight, and detailed documentation [25]. Recent studies suggest that combining conventional bureau data with alternative signals can improve risk assessment for underbanked populations, but concerns about algorithmic bias and the potential for discriminatory outcomes demand careful model governance [26].
2.5 Robo-Advisory and Wealth Management
The global robo-advisory market was estimated at USD 8.3 billion in 2024 and is projected to reach USD 33.6 billion by 2030, implying a CAGR of 26.4% [27]. Assets under management by robo-advisors grew by 58% year-on-year in 2024, partly due to AI-enhanced portfolio management [28]. A notable trend is the hybrid model that combines automated advice with human oversight, addressing findings that many high-net-worth individuals seek personalized guidance [29]. For instance, Goldman Sachs partnered with Betterment in 2024 to blend wealth management expertise with digital platforms [30]. Despite growth, robo-advisory faces challenges related to algorithmic bias, suitability assessments, and the need for transparent fee structures.
2.6 Sentiment Analysis and NLP for Financial Markets
Financial sentiment analysis has transitioned from lexicon-based dictionaries (e.g., Loughran-McDonald) to transformer-based models. FinBERT remains a popular benchmark, but LLMs such as GPT-3.5, GPT-4, and Llama2-7B have shown improvements in extracting sentiment from financial disclosures [31][32]. Research published in 2024 indicates that LLM-derived sentiment scores significantly predict stock returns in fixed-effect regression frameworks, though the economic magnitude of these effects depends on model specification and data quality [33]. Extensions to central bank communications—using FinBERT to parse ECB speeches—demonstrate the broadening scope of NLP in finance, yet the causal interpretation of such sentiment impacts remains debated [9].
3. Key Challenges
3.1 Explainability and the “Black Box” Problem
The inherent opacity of deep neural networks poses a fundamental challenge for financial applications where auditability and regulatory compliance are paramount. The Bank for International Settlements (BIS) observes that strict interpretation of model risk management guidance may be incompatible with deep learning methods, because one cannot conclusively attribute outputs to specific input combinations [34]. A 2024 systematic review identified three main obstacles: the ambiguity of target audiences for explanations, a scarcity of explainable AI (XAI) tools designed for financial contexts, and the lack of standardized metrics to evaluate explanation quality [35]. The S.A.F.E. framework (Sustainability, Accuracy, Fairness, Explainability) proposed by Giudici and Raffinetti (2023) represents one attempt to operationalize AI trustworthiness metrics [36]. The CFA Institute further highlights concerns about privacy leakage through detailed explanations and the challenge of providing real-time interpretability [37].
3.2 Bias, Fairness, and Discrimination
AI models trained on historical financial data may perpetuate and even amplify existing biases. The U.S. Department of the Treasury’s 2024 report underscores bias risks in credit decisioning, insurance pricing, and fraud detection [38]. An OECD survey of 49 jurisdictions likewise identified bias and discrimination as the most frequently reported risk area [39]. The problem is exacerbated when models exploit proxy variables (e.g., ZIP codes correlating with race or income) that can lead to discriminatory outcomes even when protected attributes are excluded [37]. Addressing these concerns requires not only technical debiasing but also ongoing fairness audits and alignment with evolving legal standards.
3.3 Data Privacy, Security, and Cybersecurity
Financial AI systems process vast quantities of sensitive data. The global average cost of a data breach reached USD 4.88 million in 2024 [40]. AI introduces novel attack vectors, including data poisoning, adversarial examples, and model inversion attacks that can expose training data. The 2024 incident in which a finance worker in Hong Kong was allegedly tricked by AI-generated deepfake video and audio into authorizing a USD 25 million transfer illustrates how generative AI can be weaponized for social engineering [41]. The FSB, in its November 2024 report, specifically flagged cybersecurity as a critical vulnerability, noting that AI tools can be used as vehicles for attacks by manipulating outputs [3]. Canada’s Office of the Superintendent of Financial Institutions (OSFI) warned that financial institutions customizing LLMs with internal data risk unintended release of consumer information or trade secrets [42].
3.4 Model Risk, Hallucinations, and Reliability
LLMs are known to generate plausible but incorrect content; estimated hallucination rates range from 1.4% to 4.2%, which is unacceptable in high-stakes financial contexts [43]. Consequently, many institutions limit AI use to internal operational efficiency rather than core business activities, due to “uncertainties surrounding regulatory expectations related to accountability, ethics, data privacy, fairness, transparency, and explainability” [43]. Additionally, concentration risk arises from dependence on a small number of technology providers—Stanford’s 2024 AI Index Report noted that 97 out of 163 foundation models released between 2019 and 2023 were controlled by just four firms [43]. The July 2024 global IT outage, with estimated financial losses of USD 5.4 billion, highlighted systemic vulnerabilities stemming from third-party dependencies [42].
3.5 Regulatory Fragmentation
The regulatory environment for AI in finance is fragmented across jurisdictions, creating compliance complexities for global institutions. The EU AI Act, effective mid-2025, classifies financial AI applications by risk tier, imposing strict requirements on high-risk systems [40]. The United States relies on a patchwork of federal and state laws (e.g., Equal Credit Opportunity Act, Fair Credit Reporting Act) and agency guidance, without a comprehensive AI statute [38]. The OECD AI Principles, updated in 2024 to address generative AI, provide non-binding guidance for 49 jurisdictions [39]. International bodies such as the FSB and BIS have published assessments and recommendations, yet binding global standards remain elusive [3][43]. Deloitte has described the landscape as “fragmented and fast-changing, creating uncertainty for financial institutions” [44].
3.6 Data Infrastructure Deficits
Despite rising AI adoption, foundational data capabilities lag. An industry survey found that 67% of senior executives view inadequate data infrastructure as a barrier to AI initiatives, and 83% believe stronger data foundations would accelerate deployment [1]. Persistent issues include data silos, inconsistent formatting, insufficient labeling for supervised learning, and legacy systems that hinder integration with modern AI workflows.
4. Future Directions
4.1 Agentic AI and Autonomous Finance
A frontier research area is agentic AI—systems that combine LLMs with planning, memory, and tool-use capabilities to execute multi-step financial tasks autonomously. Boston Consulting Group projects that 75% of finance leaders expect such agents to be routine by 2028, though these projections hinge on unresolved questions of reliability and oversight [1]. The convergence of generative AI, agentic AI, and blockchain could enable programmable finance where AI agents interact with decentralized protocols, but it also raises governance challenges [45].
4.2 Quantum Computing and Quantum-AI Integration
Quantum computing holds long-term potential for portfolio optimization, risk modeling, and Monte Carlo simulations. HSBC, in collaboration with IBM, reported a 34% improvement in predicting trade execution likelihood for bond trading using quantum algorithms, marking early empirical evidence of quantum advantage in finance [46]. Vanguard and IBM are exploring hybrid quantum–classical methods for portfolio construction [46]. However, practical, large-scale quantum applications remain years away, and quantum computing simultaneously threatens current cryptographic standards, prompting institutions such as Lloyds Banking Group to begin migrating to post-quantum cryptography [47].
4.3 Explainable and Responsible AI by Design
Future AI systems are expected to embed explainability, fairness, and robustness as design principles. The KAIRI framework for AI risk measurement exemplifies the movement toward operationalizing responsible AI [36]. Regulatory sandboxes, such as the Monetary Authority of Singapore’s Project MindForge and various EU initiatives, offer controlled environments for testing AI innovations under supervisory oversight [48].
4.4 Multimodal and Foundation Models for Finance
The next generation of financial AI is moving toward multimodal models capable of processing text, numerical time series, chart images, and audio (e.g., earnings calls). Foundation models for time series analysis may unify forecasting across diverse economic indicators, but the added complexity may exacerbate interpretability and data governance concerns [9].
4.5 ESG Integration and Sustainable Finance
AI is increasingly deployed to extract ESG signals from corporate disclosures, news, and alternative data. Lim (2024) provided a comprehensive review of this intersection [49]. Robo-advisory platforms are beginning to incorporate ESG screening, yet the inconsistency and limited standardization of ESG data remain significant obstacles [50].
4.6 Democratization and Open-Source AI
Open-source initiatives like FinGPT and FinRL lower entry barriers for smaller institutions and fintech startups, potentially fostering innovation. At the same time, the uncontrolled proliferation of sophisticated AI capabilities raises governance and safety concerns, as open models may be misused or inadequately tested [8].
5. Conclusion
Artificial intelligence is reshaping finance, but the transition from experimental promise to safe, equitable, and reliable deployment is far from complete. Recent advances in LLMs, deep reinforcement learning, and graph-based models have demonstrated notable capabilities in tasks ranging from fraud detection to portfolio management. However, the evidence base remains uneven: many performance claims derive from backtests or controlled settings that may not hold in live markets, and concerns about explainability, bias, cybersecurity, and regulatory fragmentation persist. The future trajectory points toward more autonomous, multimodal, and integrated AI systems, yet the pace of adoption will depend on institutions’ ability to strengthen data infrastructure, implement robust governance, and engage in multilateral regulatory coordination. As the FSB concluded, “while existing regulatory and supervisory frameworks address many of the vulnerabilities associated with AI adoption, more work may be needed to ensure that these frameworks are sufficient” [3].
References
[1] Pigment. (2024). The State of AI in Finance: 10 Statistics FP&A Leaders Should Know
[2] Credence Research. (2024). North America Artificial Intelligence in Finance Market Size and Share 2032
[3] Financial Stability Board. (2024, November 14). The Financial Stability Implications of Artificial Intelligence
[4] Wu, S., Irsoy, O., Lu, S., et al. (2023). BloombergGPT: A Large Language Model for Finance. arXiv preprint. Bloomberg
[5] Alpha Architect. (2024). Where Large Language Models and Finance Meet
[6] Araci, D. (2019). FinBERT: Financial Sentiment Analysis with Pre-trained Language Models. arXiv preprint.
[7] Du, M. (2024). Research on Application of Financial Large Language Models. Highlights in Business, Economics and Management, 45, 628–633. Article
[8] AI4Finance Foundation. (2024). FinGPT: Open-Source Financial Large Language Models. GitHub repository
[9] Dong, X. (2024). Large Language Models for Financial and Investment Management: Applications. Journal of Portfolio Management. PDF
[10] Fatouros, G., et al. (2024). Can Large Language Models Beat Wall Street? Evaluating GPT-4’s Impact on Financial Decision-Making with MarketSenseAI. Neural Computing and Applications. Article
[11] CNBC. (2023). JPMorgan developing AI investment advisor; OpenAI. (2023). Morgan Stanley Wealth Management deploys GPT-4. Press releases.
[12] LTX Trading. (2023). BondGPT: Introducing LTX’s Generative AI Application for Corporate Bond Trading.
[13] Cacciatore, E.J. (2024). The Transformer Revolution in Financial Markets: Technical Insights, Applications, and Caveats. Medium article
[14] Quantformer: From Attention to Profit with a Quantitative Transformer Trading Strategy. (2024). arXiv:2404.00424
[15] Yan, R., Jin, J., Han, K. (2024). Reinforcement Learning for Deep Portfolio Optimization. Electronic Research Archive, 32(9), 5176–5200. Article
[16] Cheng, L.C., Sun, J.S. (2024). Multiagent-Based Deep Reinforcement Learning Framework for Multi-Asset Adaptive Trading and Portfolio Management. Neurocomputing, 594, 127800.
[17] Financial News-Driven LLM Reinforcement Learning for Portfolio Management. (2024). arXiv:2411.11059
[18] Gandhi, H., et al. (2024). Navigating the Complexity of Money Laundering: Anti–Money Laundering Advancements with AI/ML Insights. International Journal on Smart Sensing and Intelligent Systems, 17(1).
[19] Bakhshinejad, N., et al. (2024). A Graph-Based Deep Learning Model for the Anti-Money Laundering Task of Transaction Monitoring. In IJCCI 2024 (pp. 496–507).
[20] Kapale, et al. (2024). Explainable AI for Fraud Detection: Enhancing Transparency and Trust in Financial Decision-Making. 2nd DMIHER International Conference on AI in Healthcare, Education, and Finance.
[21] Li, P.Y., et al. (2024). Unveiling the Black Box: An XAI-Based Anti-Money Laundering Model. 2024 International Conference on Consumer Electronics-Taiwan (pp. 293–294). IEEE.
[22] Wan, F., Li, P. (2024). A Novel Money Laundering Prediction Model Based on a Dynamic Graph Convolutional Neural Network and Long Short-Term Memory. Symmetry, 16(3), 378.
[23] European Central Bank. (2025, November). AI’s Impact on Banking: Use Cases for Credit Scoring and Fraud Detection. Supervisory Newsletter. Article
[24] Revolutionizing Credit Risk Management: Opportunities and Challenges in Credit Scoring with AI and Deep Learning. (2024). IntechOpen chapter
[25] Oscilar. (2024). How AI is Changing the Credit Scoring Game
[26] Addy, W., et al. (2024). AI in Credit Scoring: A Comprehensive Review of Models and Predictive Analytics. Global Journal of Engineering and Technology Advances, 18(2), 118–129.
[27] Research and Markets. (2025, April). Robo Advisory Market Report; SNS Insider. Robo Advisory Market Size, Share, Growth & Trends Report 2035. SNS Insider report
[28] IDEX Consulting. (2025). How Leading Wealth Advisors Are Using AI to Stay Competitive
[29] Capgemini. (2025). Financial Services Top Trends 2025: Wealth Management. PDF
[30] Market.us. (2024). Robo Advisory Market Size, Share, Growth
[31] Todd, et al. (2024). Text-Based Sentiment Analysis in Finance. Intelligent Systems in Accounting, Finance and Management. PDF
[32] Luo and Gong. (2025). Leveraging Large Language Model as News Sentiment Predictor in Stock Markets. Discover Computing. Article
[33] Kirtac, K., Germano, G. (2024). Sentiment Trading with Large Language Models. Finance Research Letters, 62, 105227. PDF
[34] Bank for International Settlements. (2024). Managing Explanations: How Regulators Can Address AI Explainability Issues. FSI Papers. PDF
[35] Model-Agnostic Explainable Artificial Intelligence Methods in Finance: A Systematic Review. (2025). Artificial Intelligence Review. Article
[36] Giudici, P., Raffinetti, E. (2023). SAFE Artificial Intelligence in Finance. Finance Research Letters, 56, 104088; Giudici, P., et al. (2024). Artificial Intelligence Risk Measurement. Expert Systems with Applications, 235, 121220.
[37] CFA Institute. (2025). Explainable AI in Finance: Meeting Stakeholder Needs
[38] U.S. Department of the Treasury. (2024). Artificial Intelligence in Financial Services. PDF
[39] OECD. (2024). Regulatory Approaches to Artificial Intelligence in Finance (OECD Artificial Intelligence Papers, No. 9). PDF
[40] Ideas2IT. (2024). AI Governance in Finance: Key Strategies and Challenges
[41] IMF. (2024, October). Advances in Artificial Intelligence: Implications for Capital Market. In Global Financial Stability Report (Chapter 3). PDF
[42] OSFI. (2024). AI Uses and Risks at Federally Regulated Financial Institutions
[43] Crisanto, J.C., et al. (2024). Regulating AI in the Financial Sector: Recent Developments and Main Challenges (BIS FSI Insights). PDF
[44] Deloitte. (2024). AI Regulation in Financial Services. Cited in Medium/Axonect Blog, Regulatory Challenges of Using AI in Financial Services. Medium blog
[45] KPMG. (2025). 2025 Futures Report: Seizing Opportunities in an Era of Disruption. PDF
[46] CFA Institute. (2026). Quantum Computing vs. AI. Enterprising Investor Blog.
[47] Lloyds Banking Group. (2025, October). What Is Quantum Computing and How Might It Impact Financial Services?
[48] Monetary Authority of Singapore. (2024). Project MindForge; European Commission. (2024). Launch Event for the Spanish Regulatory Sandbox on Artificial Intelligence.
[49] Lim, T. (2024). Environmental, Social, and Governance (ESG) and Artificial Intelligence in Finance: State-of-the-Art and Research Takeaways. Artificial Intelligence Review, 57(4), 76.
[50] Barile, et al. (2025). A New Era: Managing Green Investments Through Robo-Advisors. Management Decision. In Springer Nature Link, Current Trends for Robo-Advisors. Chapter
[51] Mienye, E., et al. (2024). Deep Learning in Finance: A Survey of Applications and Techniques. AI, 5(4), 2066–2091. Article
[52] A Review of Artificial Intelligence for Financial Fraud Detection. (2025). Applied Sciences, 16(4), 1931. Article
[53] European Central Bank. (2024, May). The Rise of Artificial Intelligence: Benefits and Risks for Financial Stability. Financial Stability Review. Article
[54] Reserve Bank of Australia. (2024, September). Focus Topic: Financial Stability Implications of Artificial Intelligence. Financial Stability Review. Article
[55] Nature. (2025). AI Integration in Financial Services: A Systematic Review. Humanities and Social Sciences Communications. Article