AI in finance
Last reviewed
Sources
28 citations
Review status
Source-backed
Revision
v5 ยท 6,280 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
28 citations
Review status
Source-backed
Revision
v5 ยท 6,280 words
Add missing citations, update stale details, or suggest a clearer explanation.
AI in finance is the use of artificial intelligence and machine learning across financial services to automate decisions and analyze data at scale. Banks, insurers, asset managers, and fintech firms apply AI to detect fraud, execute trades, score credit, manage portfolios, answer customer questions, and meet regulatory obligations, increasingly using natural language processing and large language models to read filings, summarize research, and draft reports in seconds. The global AI in finance market was valued at approximately $38.36 billion in 2024, and MarketsandMarkets forecasts it will reach $190.33 billion by 2030, a compound annual growth rate (CAGR) of 30.6% [1]. As of 2025, over 85% of financial firms actively apply AI in some capacity, and more than 80% plan to increase their AI investments in the near term [2].
The financial industry was among the earliest adopters of computational methods for decision-making. Quantitative trading models emerged in the 1980s, and banks began using statistical models for credit scoring decades ago. What has changed in the 2020s is the scale, sophistication, and pervasiveness of AI in finance. Modern systems process millions of transactions per second, parse earnings calls in real time, and execute trades in microseconds based on patterns identified by neural networks.
AI adoption in finance is driven by several factors: the availability of massive datasets (transaction logs, market feeds, customer records), intense competitive pressure to reduce costs and improve returns, regulatory demands for better risk management, and the falling cost of cloud computing and specialized hardware such as GPUs. Financial regulators have also begun to grapple with the implications of AI, introducing new frameworks for model risk management, algorithmic accountability, and data governance.
The integration of large language models (LLMs) into financial workflows, beginning in earnest in 2023, has accelerated adoption further. Banks and asset managers now deploy LLM-powered assistants that can summarize research, draft reports, answer compliance questions, and interpret regulatory filings in seconds rather than hours.
AI is deployed across virtually every segment of the financial industry. The following table summarizes the major application areas, the AI techniques involved, and representative examples.
| Application | Description | Key AI techniques | Examples |
|---|---|---|---|
| Algorithmic trading | Automated execution of trades based on quantitative models and real-time data analysis | Reinforcement learning, deep neural networks, time-series forecasting | Renaissance Technologies, Citadel, Two Sigma |
| Fraud detection | Real-time identification of suspicious transactions, account takeovers, and payment fraud | Anomaly detection, random forests, deep neural networks | Stripe Radar, Featurespace, FICO Falcon |
| Credit scoring | Assessment of borrower creditworthiness using alternative data sources beyond traditional credit history | Gradient boosting, logistic regression, neural networks | Upstart, Zest AI, LenddoEFL |
| Robo-advisors | Automated portfolio management and financial planning with minimal human intervention | Optimization algorithms, modern portfolio theory, ML-based rebalancing | Betterment, Wealthfront, Schwab Intelligent Portfolios |
| Risk management | Modeling and quantification of market, credit, and operational risks | Monte Carlo simulation, deep learning, stress testing models | JPMorgan Athena, Goldman Sachs Marquee |
| Insurance underwriting | Automated assessment of insurance risk and policy pricing | Predictive modeling, computer vision (damage assessment), NLP (claims processing) | Lemonade, Tractable, Shift Technology |
| Regulatory compliance (RegTech) | Automated monitoring of regulatory obligations, reporting, and audit trails | NLP for regulatory text parsing, rule engines, anomaly detection | ComplyAdvantage, Behavox, Ascent RegTech |
| Customer service chatbots | AI-powered conversational agents handling customer inquiries, account management, and product recommendations | NLP, dialogue systems, transformer models | Bank of America's Erica, Capital One's Eno |
| Anti-money laundering (AML) | Detection of complex money laundering patterns across transaction networks | Graph neural networks, network analysis, anomaly detection | Napier AI, Hawk AI, NICE Actimize |
Algorithmic trading, sometimes called algo trading or automated trading, uses computer programs to execute orders at speeds and frequencies that human traders cannot match. AI-driven trading systems analyze market microstructure, news sentiment, macroeconomic indicators, and even satellite imagery to identify trading opportunities.
The global AI trading market was valued at approximately $11.2 billion in 2024, with projections to nearly triple that figure by 2030 [3]. Algorithmic trading now accounts for an estimated 60% to 73% of all equity trading volume in the United States, depending on the measure used.
Modern trading algorithms go well beyond simple rule-based strategies. They employ deep reinforcement learning to adapt to changing market conditions, recurrent neural networks to model temporal dependencies in price data, and attention mechanisms to weigh the relevance of incoming information. Some firms use generative models to simulate market scenarios and stress-test strategies before deployment.
AI-driven trading has expanded significantly into cryptocurrency and decentralized finance (DeFi) markets. In 2025, AI-driven crypto hedge funds delivered average annual returns of 36%, compared to 21% for long-only crypto funds and 13% for market-neutral strategies. Quantitative AI-driven crypto funds achieved an average annual return of 48%, outpacing traditional crypto strategies by 12-15% [22].
As of 2025, there are over 400 active crypto hedge funds worldwide, and more than 55% of traditional hedge funds now hold crypto assets, up from roughly a third in 2022. AI enhances crypto trading through several mechanisms:
| AI technique | Application in crypto/DeFi |
|---|---|
| Reinforcement learning | Simulating thousands of market scenarios to optimize asset allocations and reduce risk |
| Execution algorithms | Managing liquidity and reducing slippage across centralized and decentralized exchanges |
| Sentiment analysis | Monitoring social media, on-chain data, and news feeds for trading signals |
| Smart contract analysis | AI-powered auditing of DeFi smart contracts for vulnerabilities before deploying capital |
| Cross-exchange arbitrage | AI directing trades across multiple exchanges for optimal execution |
Transaction speeds for crypto hedge funds have improved by approximately 20% due to advancements in blockchain infrastructure and AI-enhanced execution, reducing latency in high-frequency trading [22]. Notable AI-focused crypto funds include Multicoin Capital, BlockTower Capital, and IAESIR Crypto Hedge Fund.
Fraud detection is one of the most mature and impactful applications of AI in finance. By 2025, approximately 87% of global financial institutions had implemented AI-based fraud detection systems, up from 72% in early 2024 [2]. These systems analyze hundreds of signals per transaction, including transaction amount, location, device fingerprint, behavioral biometrics, and historical patterns, to assign a risk score in real time.
Stripe Radar, one of the most widely used fraud prevention tools, processes payments for millions of businesses and learns from the collective network of over $1.4 trillion in annual payment volume. Stripe migrated from an ensemble model combining XGBoost and a deep neural network to a pure deep neural network architecture in mid-2022, achieving a 38% average reduction in fraud across its merchant base [4]. In 2025, Stripe introduced AI-powered features including natural language rule creation via its Radar Assistant and dynamic rules that combine machine learning with real-time issuer responses [4].
AI-based fraud detection significantly outperforms traditional rule-based systems because it can identify novel fraud patterns that have never been explicitly programmed as rules. However, it also creates challenges around false positives (legitimate transactions flagged as fraudulent) and the need to explain decisions to regulators and customers.
Traditional credit scoring relies on a limited set of variables, primarily payment history, outstanding debt, length of credit history, and types of credit used. AI-based credit scoring models incorporate a far broader range of signals, sometimes called alternative data, including utility payments, rent history, employment data, education, and even behavioral patterns such as how a borrower fills out a loan application.
Companies like Upstart have demonstrated that AI-driven credit models can approve 27% more borrowers and reduce loss rates by 16% compared to traditional models, according to the company's regulatory filings. Upstart's model uses over 1,600 variables and is trained on more than 55 million repayment events [5].
The use of AI in credit scoring raises important questions about fairness and explainability. Under the U.S. Equal Credit Opportunity Act and the EU's General Data Protection Regulation (GDPR), lenders must be able to explain why a credit application was denied. This creates tension with complex ML models whose internal logic may not be easily interpretable.
Robo-advisors are digital platforms that provide automated, algorithm-driven financial planning and investment management with little to no human supervision. They typically use modern portfolio theory to construct diversified portfolios based on a client's risk tolerance, time horizon, and financial goals, then continuously rebalance holdings as market conditions change.
As of 2025, global assets under management by robo-advisors exceeded $2.5 trillion, up from approximately $1.4 trillion in 2022 [6]. Major platforms include Betterment (managing approximately $50 billion), Wealthfront, and robo-advisory offerings from traditional firms like Schwab, Vanguard, and Fidelity.
Increasingly, robo-advisors are incorporating AI techniques beyond simple portfolio optimization, including tax-loss harvesting algorithms, NLP-based financial planning chatbots, and personalized asset allocation models that learn from individual client behavior.
Anti-money laundering (AML) represents a critical and rapidly growing application of AI in finance. Financial institutions globally spend over $180 billion annually on compliance, a substantial portion of which goes to AML activities [7]. Traditional rule-based AML systems generate enormous numbers of false positives, with industry estimates suggesting that 95% or more of alerts are false alarms, wasting investigator time and resources.
AI-powered AML systems use graph neural networks to analyze transaction networks, identify unusual flow patterns, and detect layering techniques (where illicit funds are moved through multiple accounts to obscure their origin). A 2023 PwC survey found that 62% of financial institutions already use AI and ML for AML, with adoption expected to reach 90% by 2025. AI-powered solutions reduce false positives by 90% to 95% while improving detection of sophisticated laundering patterns [7].
The global RegTech market, which encompasses AML along with broader regulatory compliance technology, is projected to exceed $22 billion by mid-2025, growing at a CAGR of 23.5% [7].
AI is transforming the insurance industry across underwriting, claims processing, customer service, and fraud detection. Analysts project that by late 2026, more than 35% of insurers will deploy AI agents across at least three core functions, cutting processing time by up to 70% [23].
Traditional insurance underwriting involved manual review of application data, which could take days or weeks. AI-powered underwriting systems now analyze applications in minutes or seconds, incorporating data from multiple sources including public records, social media, IoT devices, telematics, and satellite imagery. This reduces cycle times from weeks to minutes while improving risk assessment accuracy.
| Insurance AI application | Description | Impact |
|---|---|---|
| Automated underwriting | AI models assess risk in real time using alternative data sources, enabling instant policy issuance | Reduces underwriting time from days to minutes; improves risk selection |
| Claims processing | Computer vision analyzes photos of vehicle damage; NLP extracts information from claims documents | Tractable's AI assesses vehicle damage from photos within seconds; reduces claims cycle by 40-60% |
| Customer service | AI chatbots handle policy inquiries, quote generation, and first notice of loss | One insurer reported 80% of transactions moving online with customer satisfaction scores rising 36% |
| Fraud detection | AI identifies suspicious claim patterns, duplicate claims, and staged accidents | Shift Technology's AI flags fraudulent claims with accuracy rates exceeding 95% |
| Telematics and usage-based pricing | AI analyzes driving behavior data from connected vehicles to personalize auto insurance premiums | Enables pay-per-mile and behavior-based pricing models |
Lemonade, the AI-native insurance company, uses AI across its entire operations. Its claims bot, Jim, can review and pay claims in as little as three seconds for straightforward cases. The company reports that AI handles the majority of its customer interactions without human intervention [24].
AI-powered chatbots and virtual assistants have become standard in retail banking. Approximately 92% of North American banks use AI chatbots in customer service as of 2025 [2].
| Banking chatbot | Bank | Users / Scale | Key capabilities |
|---|---|---|---|
| Erica | Bank of America | Over 2 billion interactions since 2018 launch | Account management, spending insights, bill reminders, credit score monitoring, investment guidance |
| Eno | Capital One | Available to all Capital One customers | Transaction alerts, fraud notifications, account inquiries, virtual card numbers |
| NOMI | Royal Bank of Canada | Integrated into RBC mobile banking | Spending insights, budget recommendations, predictive cash flow |
| Amy | HSBC | Deployed across Hong Kong operations | Account inquiries, product information, service requests |
| Ceba | Commonwealth Bank (Australia) | Handles 200+ banking tasks | Account management, card services, payment assistance |
These chatbots handle routine inquiries, freeing human agents to focus on complex issues. Bank of America's Erica has processed over 2 billion customer interactions since its 2018 launch, demonstrating the scale at which AI-powered banking assistants now operate.
The roots of AI in finance trace back to the early adoption of quantitative methods on Wall Street. In the 1970s and 1980s, a small number of mathematicians and physicists began applying statistical models to financial markets, laying the groundwork for what would become algorithmic trading.
Edward Thorp, a mathematics professor at UC Irvine, is widely credited as one of the pioneers. After using probability theory to beat blackjack casinos (documented in his 1962 book "Beat the Dealer"), Thorp turned his methods to the stock market. In 1969, he co-founded Princeton/Newport Partners, one of the first quantitative hedge funds, which used statistical arbitrage strategies and achieved consistently strong returns for nearly two decades [8].
In the 1980s, Wall Street firms began hiring physicists and mathematicians, colloquially known as "quants," to build pricing models for increasingly complex financial derivatives. The Black-Scholes model, published in 1973, had provided a mathematical framework for options pricing, and the growing derivatives market demanded ever more sophisticated computational approaches.
Renaissance Technologies, founded in 1982 by mathematician James Simons, represents perhaps the most successful application of quantitative and AI methods to finance in history. Simons, who had previously worked as a Cold War code breaker at the Institute for Defense Analysis (IDA) and later chaired the mathematics department at Stony Brook University, brought a fundamentally different approach to investing: treating financial markets as complex systems amenable to mathematical modeling rather than relying on human judgment about company fundamentals.
Simons initially started a fund management firm called Monemetrics in 1978, trading currencies. After renaming the firm Renaissance Technologies in 1982, he began recruiting scientists rather than finance professionals. His first key hire was Leonard Baum, a cryptanalyst from the IDA who co-authored the Baum-Welch algorithm for hidden Markov models. Other early recruits included algebraist James Ax and, later, computational linguists Peter Brown and Robert Mercer from IBM's speech recognition group [9].
In 1988, Renaissance launched the Medallion Fund, which would go on to produce arguably the most extraordinary track record in investment history. From 1988 through 2018, the Medallion Fund generated average annual returns of approximately 66% before fees (39% after the fund's steep 5-and-44 fee structure). From 1994 through mid-2014, the fund returned 71.8% annually before fees [9]. To put this in perspective, $1 invested in the Medallion Fund at its inception would have been worth over $40,000 by 2018, compared to approximately $20 for a passive S&P 500 index investment.
Renaissance's approach involved collecting vast amounts of market data, storing it in petabyte-scale data warehouses, and using machine learning models to identify statistical patterns and correlations that could be profitably traded. The firm adapted methods from IBM's speech recognition research, originally designed to predict the next sound in a sequence based on prior sounds, and applied them to predict price movements based on historical sequences [9].
The fund's strategies are executed at very short time horizons (often holding positions for seconds to days), allowing the statistical edge of each individual trade to be small but highly repeatable. Renaissance's success demonstrated that machine learning and data-driven methods could consistently outperform traditional investment approaches, catalyzing the growth of the quantitative hedge fund industry.
The 1990s saw early experiments with neural networks in financial applications, though the results were mixed and the technology was not yet mature enough for widespread adoption. Researchers explored neural networks for stock price prediction, credit risk assessment, and foreign exchange forecasting.
The limitations of early neural networks, including insufficient computing power, limited training data, and the "vanishing gradient" problem, constrained their practical utility. It was not until the deep learning revolution of the 2010s, enabled by advances in GPU computing, the availability of massive datasets, and architectural innovations like dropout and batch normalization, that neural networks became truly viable for financial applications at scale.
By the mid-2010s, hedge funds and investment banks were investing heavily in deep learning talent and infrastructure. Firms like Two Sigma, D.E. Shaw, and Citadel built large data science teams, and the competition for AI researchers between Wall Street and Silicon Valley intensified.
Hedge funds deploying AI-driven trading strategies have consistently outperformed their peers. A 2024 SEC report found that AI-driven funds outperformed their peers by an average of 12%. The performance gap has been particularly notable in systematic and quantitative strategies, where AI's ability to process vast amounts of data and execute at speed provides the greatest advantage [22].
The release of ChatGPT in November 2022 and the subsequent wave of large language model development triggered rapid adoption of generative AI across the financial industry. LLMs have proven particularly useful for tasks that involve processing and synthesizing large volumes of unstructured text, a pervasive need in finance.
In March 2023, Bloomberg unveiled BloombergGPT, a 50-billion-parameter language model specifically built for finance. The model was trained on a 363-billion-token dataset drawn from Bloomberg's proprietary financial data sources, combined with 345 billion tokens from general-purpose datasets [10]. Bloomberg designed the model to excel at financial NLP tasks, including sentiment analysis, named entity recognition, news classification, and question answering on financial topics.
Gideon Mann, Head of Bloomberg's ML Product and Research team, explained the rationale for building a finance-specific model: "The quality of machine learning and NLP models comes down to the data you put into them. Thanks to the collection of financial documents Bloomberg has curated over four decades, we were able to carefully create a large and clean, domain-specific dataset to train a LLM that is best suited for financial use cases" [26].
BloombergGPT demonstrated that domain-specific LLMs trained on specialized data can significantly outperform general-purpose models on industry-specific benchmarks while retaining strong general language capabilities. The model's release sparked a broader trend toward sector-specific LLMs across finance and other industries.
FinGPT, an open-source alternative to BloombergGPT, emerged from the AI4Finance Foundation in 2023. Unlike Bloomberg's proprietary model, FinGPT takes a data-centric approach, providing researchers and practitioners with accessible, transparent resources for developing financial LLMs. FinGPT prioritizes lightweight adaptation, leveraging the best available open-source base models and fine-tuning them on publicly available financial data [11].
The project represents an effort to democratize financial AI and reduce the barrier to entry for smaller firms and academic researchers who lack access to Bloomberg's proprietary data and infrastructure.
Morgan Stanley became one of the first major banks to deploy an LLM-powered assistant at scale. In September 2023, the firm launched an AI assistant built on OpenAI's GPT-4 for its approximately 16,000 financial advisors. The system provides instant access to over 100,000 research reports, investment strategy documents, and market commentaries, allowing advisors to retrieve relevant information through natural language queries rather than manual search [12].
More than 98% of Morgan Stanley's financial advisor teams have adopted the AI @ Morgan Stanley Assistant [27]. Jeff McMillan, the firm's Head of Firmwide AI, summarized the value to advisors: "This technology makes you as smart as the smartest person in the organization" [27]. The firm has continued to expand the system's capabilities, developing role-specific copilots for investment banking, trading, and risk management functions [12].
In January 2025, Goldman Sachs began rolling out its internal GS AI Assistant to 10,000 employees across investment banking, trading, and research divisions, with the stated goal of reaching all knowledge workers firm-wide. The tool assists with tasks ranging from document summarization and code generation to regulatory research and client communication drafting [13].
JPMorgan Chase introduced a generative AI assistant called the LLM Suite to its employees in collaboration with OpenAI. The system is designed to harness LLMs for a wide range of tasks across the firm. Separately, JPMorgan developed IndexGPT, an AI-powered tool for thematic investing that combines GPT-4 with advanced NLP to curate thematic investment baskets [14].
JPMorgan Chase has been one of the most aggressive adopters of AI among global banks. The firm's COiN (Contract Intelligence) platform, deployed in 2017, uses AI to analyze commercial loan agreements and extract critical data points from legal documents. The system processes approximately 12,000 contracts annually, work that previously required approximately 360,000 hours of manual review by lawyers and loan officers. COiN achieves a near-zero error rate on these tasks, a level of accuracy that manual processing could not reliably match [14].
Beyond COiN, JPMorgan has invested broadly in AI. The firm expects to spend roughly $19.8 billion on technology in 2026, an increase of nearly $2 billion year over year, with AI a central priority [28]. JPMorgan spends about $2 billion a year specifically on AI, and chief executive Jamie Dimon has said the program's costs are now offset by a similar amount in annual savings, calling current gains "the tip of the iceberg" [28]. As of 2025, JPMorgan employs more than 2,000 AI and ML specialists and has filed for numerous AI-related patents, including trademarks for AI-powered financial tools [14].
Stripe's Radar system exemplifies how network effects can amplify the power of AI in finance. Because Stripe processes payments for millions of businesses globally (handling over $1.4 trillion in annual payment volume), its fraud detection models benefit from an exceptionally large and diverse training dataset. Every payment on the Stripe network generates signals that feed back into Radar's models, improving detection for all merchants [4].
Radar uses hundreds of signals per transaction, including behavioral patterns, device characteristics, and network-level information. Stripe has continued to invest in advancing its AI capabilities, including building a Payments Foundation Model designed to improve each step of the payments process with AI [4].
Ant Group, the financial technology affiliate of Alibaba Group that operates Alipay, has embedded AI throughout its financial services ecosystem. Serving over 1.3 billion users, Ant Group uses machine learning algorithms and custom-designed chips to process credit decisions in seconds, analyzing vast quantities of user data to determine loan eligibility [15].
In 2023, Ant Group invested 21.19 billion yuan ($2.92 billion) in technology research and development, with a heavy focus on AI. The company has launched several AI-powered services on the Alipay platform, including Zhixiaobao (an AI life assistant), Maxiaocai (an AI financial manager), and Angel (an AI healthcare manager that uses augmented reality to assist hospital patients). In 2025, Ant Group launched Alipay AI Pay, an AI-native payment solution that enables secure transactions through AI agents. By February 2025, it had surpassed 100 million users, becoming the world's first AI-native payment product to reach that milestone [15].
Ant International also launched the Alipay+ GenAI Cockpit, an AI-as-a-Service platform that enables fintech companies and super apps to build AI-native financial services [15].
Citadel, the multi-strategy hedge fund founded by Ken Griffin managing over $60 billion in assets, represents a pragmatic approach to AI in finance. Machine learning has been part of Citadel's quantitative trading operations for over a decade, with the firm employing reinforcement learning models to optimize trading strategies and statistical arbitrage approaches where ML models identify complex patterns invisible to human analysts [16].
However, Citadel's leadership has taken a measured public position on AI's impact. Ken Griffin has described AI as a productivity enhancement tool rather than a revolutionary force in finance, noting that while generative AI streamlines workflows and improves efficiency, it has not yet "revolutionized" the firm's core strategies [16].
Citadel has developed an AI Assistant for its equity investors that can rapidly process SEC filings, earnings transcripts, and research reports. The system understands portfolio context, highlights risks, and builds customized reading lists. Notably, the assistant does not make buy or sell decisions; it serves as an information processing and retrieval tool for human decision-makers [16].
Regulatory technology (RegTech) powered by AI has become one of the fastest-growing segments in financial technology. Financial institutions face an increasingly complex web of regulations across jurisdictions, and the cost of non-compliance can be severe, both in fines and reputational damage.
The regulatory environment for AI in finance continues to evolve rapidly:
| Regulator / Framework | Key AI-related requirements (2025-2026) |
|---|---|
| SEC (U.S.) | FY 2026 Examination Priorities sharpen expectations around AI, cybersecurity, and controls |
| FINRA (U.S.) | 2026 Oversight Report elevates GenAI risk and cyber-enabled fraud as examination priorities |
| EU AI Act | Classifies credit scoring and insurance underwriting AI as high-risk; requires transparency, documentation, and human oversight |
| NIST (U.S.) | Draft Cyber AI Profile accelerating standards for what "secure AI" looks like in practice |
| Bank of England (UK) | Published expectations for AI model risk management in supervised firms |
| European Central Bank | Issued supervisory expectations for banks' use of AI in risk management and compliance |
AI-powered compliance tools monitor regulatory changes across jurisdictions, parse new regulations using NLP, assess their impact on the firm's operations, and generate compliance reports. Rather than large-scale system overhauls, firms are increasingly adopting agentic AI for compliance one process at a time, allowing them to realize value without major infrastructure changes [25].
Notable RegTech AI providers include ComplyAdvantage (real-time financial crime risk data), Behavox (employee communications monitoring), Ascent RegTech (automated regulatory obligation tracking), and Chainalysis (blockchain analytics for cryptocurrency compliance).
Financial regulators increasingly require that AI-driven decisions be explainable. Under the U.S. Equal Credit Opportunity Act, lenders must provide specific reasons when denying credit. The EU's GDPR grants individuals a right to explanation for automated decisions. The EU AI Act, which classifies AI systems used in credit scoring and insurance underwriting as high-risk, imposes additional transparency and documentation requirements [17].
This creates a fundamental tension with modern deep learning models, which often function as "black boxes." The field of explainable AI (XAI) has developed techniques like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) to provide post-hoc explanations of model decisions, but achieving genuine interpretability in complex financial models remains an active area of research.
Financial institutions operate under some of the most stringent regulatory regimes of any industry. AI systems must comply with regulations governing fair lending, consumer protection, market manipulation, data privacy, capital adequacy, and anti-money laundering. Different jurisdictions impose different requirements, adding complexity for global institutions.
The U.S. Federal Reserve, the Office of the Comptroller of the Currency (OCC), and the FDIC issued guidance on model risk management (SR 11-7) that applies to AI models. The European Central Bank has published expectations for the use of AI by supervised banks. In 2025, regulators across jurisdictions are actively developing more specific frameworks for AI in finance, though harmonization remains limited [17].
Model risk, the risk that a model produces inaccurate outputs leading to poor decisions, is amplified when complex AI systems are used for high-stakes financial decisions. AI models can fail in ways that are difficult to anticipate, particularly when market conditions shift outside the range of their training data (a phenomenon known as distribution shift).
The 2010 Flash Crash, where the Dow Jones Industrial Average plunged approximately 1,000 points in minutes before recovering, illustrated how automated trading systems can interact in unpredictable ways. While the Flash Crash was triggered by a large sell order and amplified by high-frequency trading algorithms rather than AI specifically, it underscored the systemic risks of automated decision-making in financial markets.
AI models are only as good as the data they are trained on. In finance, data quality issues include missing values, survivorship bias (historical datasets that exclude failed companies), look-ahead bias (inadvertently using information that was not available at the time of the decision), and selection bias. Financial data is also inherently non-stationary: the statistical properties of market data change over time, meaning models trained on historical data may not generalize to future conditions.
Bias in training data can lead to discriminatory outcomes, particularly in credit scoring and insurance underwriting. If historical lending data reflects past discrimination against minority groups, an AI model trained on that data may perpetuate those patterns. Addressing this requires careful data curation, fairness-aware model training, and ongoing monitoring of model outcomes across demographic groups.
Trading algorithms can be vulnerable to adversarial attacks, deliberate attempts to manipulate the inputs or environment that AI systems rely on. Examples include "spoofing" (placing and then canceling large orders to create a false impression of supply or demand), "layering" (a more sophisticated version of spoofing involving multiple price levels), and feeding misleading information into news feeds or social media channels that sentiment-analysis models monitor.
In 2024, researchers demonstrated that coordinated adversarial strategies could manipulate certain types of reinforcement learning-based trading agents by exploiting their predictable patterns of behavior. As AI systems become more prevalent in markets, the incentives and opportunities for adversarial exploitation grow correspondingly [18].
If many financial institutions use similar AI models and datasets, they may converge on similar trading strategies, risk assessments, and portfolio allocations. This herding effect could amplify market volatility and increase systemic risk, as correlated algorithmic behaviors might trigger cascading sell-offs or liquidity crises. Regulators have flagged this concentration risk as a growing concern [17].
The AI in finance market has experienced rapid expansion and is projected to continue growing at a strong pace.
| Metric | Value | Source |
|---|---|---|
| Global AI in finance market (2024) | $38.36 billion | MarketsandMarkets [1] |
| Projected market size (2030) | $190.33 billion | MarketsandMarkets [1] |
| CAGR (2024 to 2030) | 30.6% | MarketsandMarkets [1] |
| Financial sector AI spending (2024) | $45 billion | Statista [19] |
| Global AI trading market (2024) | $11.2 billion | Industry estimates [3] |
| Generative AI in financial services (2024) | $2.21 billion | Grand View Research [20] |
| Global RegTech market (projected 2025) | $22+ billion | Industry estimates [7] |
| AI agents in financial services (2025) | $691.3 million | Precedence Research [21] |
| AI-driven crypto hedge fund average return (2025) | 36% | Industry estimates [22] |
As of 2025, 91% of asset managers are using or plan to use AI for portfolio construction and research, up sharply from 55% in 2023. Approximately 92% of North American banks use AI chatbots in customer service [2].
The financial industry's adoption of AI has reached an inflection point. Several trends define the current landscape:
LLM integration has become standard. Every major global bank, including JPMorgan Chase, Goldman Sachs, Morgan Stanley, Bank of America, and Citigroup, has deployed or is deploying LLM-powered internal assistants. These tools are used for research summarization, regulatory document analysis, code generation, and client communication. The focus has shifted from proof-of-concept experiments to production deployment at scale.
AI in compliance is accelerating. Regulatory pressure and the rising cost of compliance (estimated at over $180 billion annually for the global financial industry) are driving rapid adoption of AI-powered RegTech solutions. AML systems using graph neural networks and advanced anomaly detection are replacing legacy rule-based approaches, dramatically reducing false positive rates while improving detection of sophisticated financial crimes [7].
Generative AI is reshaping investment research. Equity analysts and portfolio managers increasingly use AI tools to parse earnings calls, summarize 10-K filings, monitor news flow, and generate preliminary research reports. A 2025 survey found that the majority of equity research teams at major investment banks use some form of AI assistance in their workflow [13].
Insurance AI is reaching operational scale. By late 2026, analysts project more than 35% of insurers will deploy AI agents across at least three core functions, moving beyond chatbots to end-to-end automation of submissions, claims, and policy management [23].
Crypto and DeFi AI is maturing. AI-driven crypto trading strategies have demonstrated significant outperformance, and the integration of AI into DeFi protocols for automated market making, risk assessment, and portfolio management is accelerating.
The explainability gap persists. Despite progress in XAI techniques, the tension between model complexity and regulatory explainability requirements remains unresolved. This is particularly acute in credit scoring and insurance underwriting, where regulators require clear reasons for adverse decisions.
Regulators are catching up. The EU AI Act, which enters full applicability for most financial AI systems in 2026, will impose significant new requirements on financial institutions using AI for credit decisions, insurance underwriting, and other high-risk applications. In the United States, regulators are updating model risk management guidance to address AI-specific concerns, though a comprehensive federal AI law remains absent [17].
AI-powered fraud and AI-powered fraud detection are in an arms race. As AI tools become more accessible, fraudsters are using generative AI to create more convincing phishing emails, synthetic identities, and deepfake voice calls for social engineering. Financial institutions are responding by deploying AI systems that can detect AI-generated fraud, creating an ongoing technological arms race.
The financial industry's experience with AI illustrates both the transformative potential and the genuine risks of deploying sophisticated AI systems in high-stakes, heavily regulated environments. The coming years will be defined by how well institutions and regulators navigate the tension between innovation and oversight.