Finance
Last reviewed
May 13, 2026
Sources
33 citations
Review status
Source-backed
Revision
v2 ยท 4,806 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 13, 2026
Sources
33 citations
Review status
Source-backed
Revision
v2 ยท 4,806 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Finance ChatGPT Plugins
Artificial intelligence in finance refers to the use of machine learning, deep learning, and generative AI systems across banking, capital markets, insurance, fintech, and personal finance. Finance was one of the first industries to adopt statistical and machine learning models at scale, and by 2024 surveys from the Bank of England and the Financial Conduct Authority found that 75% of UK financial firms were already using AI in some part of their business, up from 58% in 2022.
The industry's relationship with AI is older than most people assume. Renaissance Technologies was applying hidden Markov models and pattern recognition to commodity prices in the late 1980s. FICO scores, the dominant US consumer credit benchmark introduced in 1989, are themselves a statistical model. What changed after 2017 was the arrival of transformer models and, after 2022, the rapid push to deploy large language models inside trading floors, wealth-management businesses, and customer-service operations.
AI applications in finance fall into roughly four buckets:
| Category | Typical techniques | Representative use cases |
|---|---|---|
| Predictive AI | Gradient boosting, neural networks, time-series models | Credit scoring, fraud detection, churn prediction, default risk |
| Algorithmic trading | Statistical learning, reinforcement learning, signal processing | Execution algorithms, market making, deep hedging, alpha research |
| Generative AI | Large language models, retrieval-augmented generation | Research summarization, advisor copilots, document review, customer chat |
| Operational AI | Computer vision, NLP, classification | Document automation, KYC, claims triage, accounts payable |
McKinsey estimated in 2023 that generative AI alone could add between $200 billion and $340 billion in annual value to global banking, roughly 9% to 15% of operating profits, mostly through productivity gains in coding, customer service, and document handling. The actual realized value to date is much smaller. A November 2024 Financial Stability Board report concluded that AI adoption is widespread but uneven, with foundation models still representing only about 17% of AI use cases in surveyed UK firms.
The sector is also unusually regulated, which shapes what gets deployed. United States banks have run quantitative models under the Federal Reserve's SR 11-7 model risk management guidance since 2011, and that framework now applies to every machine learning system used in lending, capital, and operational decisions. The EU AI Act classifies credit scoring and life-and-health insurance pricing as high-risk uses requiring conformity assessment.
Quantitative finance and AI have intertwined histories. The Black-Scholes option pricing model, published in 1973, did not use machine learning, but it established the precedent that exotic mathematics could move from academia to trading desks within a few years. Statistical credit scoring followed shortly after, with FICO launching its first general-purpose consumer score in 1989. Through the 1990s and 2000s, banks deployed expert systems for compliance, neural networks for fraud detection, and rule-based engines for AML alerting.
A more aggressive form of machine learning appeared at hedge funds. James Simons founded Renaissance Technologies in 1982 and launched the Medallion Fund in 1988. The firm relied on pattern recognition, hidden Markov models, and statistical learning rather than fundamental analysis, and reportedly returned roughly 66% gross annual returns from 1988 through 2018, according to a Bradford Cornell analysis cited in the popular history The Man Who Solved the Market by Gregory Zuckerman. Two Sigma, founded in 2001 by David Siegel and John Overdeck, scaled this approach further, running multi-strategy quant funds with more than 250 PhDs on staff.
The deep learning era brought new tools to mainstream banks. JPMorgan's Contract Intelligence (COIN) program, deployed around 2017, used natural language processing to extract data from commercial loan agreements. The bank later disclosed that COIN replaced an estimated 360,000 hours of legal review per year, although the work it absorbed was largely template extraction rather than legal judgment.
Generative AI arrived in 2022 and 2023. Bloomberg published BloombergGPT, a 50-billion-parameter model trained on a 363-billion-token financial corpus, in a March 2023 arXiv paper. Morgan Stanley launched AI @ Morgan Stanley Assistant, a GPT-4-based internal tool for wealth advisors, in September 2023. JPMorgan filed a trademark application for IndexGPT in May 2023 and rolled out Quest IndexGPT for thematic index construction in 2024.
Consumer credit scoring has been algorithmic since FICO introduced its score in 1989, but until the 2010s the dominant approach was logistic regression on a small set of credit bureau variables. Machine learning challengers, including Zest AI (founded in 2009 by former Google CIO Douglas Merrill), Upstart, and Pagaya, market gradient-boosted models that incorporate cash-flow data, education, employment history, and other non-traditional variables. Zest AI says its FairBoost tool helps lenders identify less discriminatory alternative models, a concept that the Consumer Financial Protection Bureau endorsed in its March 2023 fair lending guidance.
The central legal constraint is the Equal Credit Opportunity Act and Regulation B in the United States, which require creditors to provide adverse action notices explaining specific reasons for credit denials. On September 19, 2023, the CFPB issued a Consumer Protection Circular making clear that this obligation applies regardless of whether the decision was made by AI or by simpler models, a statement aimed at lenders that had treated complex models as black boxes.
The most cited cautionary tale is the Apple Card. In November 2019, software developer David Heinemeier Hansson tweeted that the card, underwritten by Goldman Sachs, had given him 20 times the credit limit of his wife despite shared finances. Steve Wozniak reported a similar disparity. The New York Department of Financial Services opened an investigation and reviewed underwriting data for roughly 400,000 New York applicants. Its March 2021 report did not find unlawful discrimination, concluding that men and women with similar credit characteristics generally received similar treatment, but criticized Goldman for poor customer-facing transparency about why limits were set as they were. The episode is still cited in regulatory speeches as evidence that algorithmic credit decisions are accountable under fair-lending law whether or not bias is intentional.
Payment fraud is the longest-running production use of machine learning in finance, going back to FICO's Falcon system in 1992. The volumes are enormous: Mastercard processes roughly 125 billion transactions per year on its network, and its Decision Intelligence platform scores each one against fraud risk. In May 2024, Mastercard announced Decision Intelligence Pro, a recurrent neural network architecture that the company said could improve fraud detection rates by an average of 20%, with some banks seeing far higher uplifts on specific attack types.
Stripe Radar uses a fraud model trained on transactions across the Stripe network, which processed more than $1.4 trillion in payments in 2024 according to the company's annual letter. Stripe says Radar's models reduce fraud by about 32% on average against a rule-based baseline and that around 92% of cards seen by Stripe have appeared elsewhere on the network, giving the system unusual training-data depth. PayPal acquired the fraud-detection startup Simility in 2018 and combines linear models, neural networks, and deep learning ensembles. The company has reported keeping its fraud rate near 0.32% of revenue, well below the merchant-industry average it cites of about 1.32%.
| Vendor / product | Founded | Specialty | Notable detail |
|---|---|---|---|
| FICO Falcon | 1992 | Card fraud scoring | Used by issuers covering more than 2.6 billion cards globally |
| Mastercard Decision Intelligence Pro | 2024 | Network-level transaction scoring | Recurrent neural network on 125 billion annual transactions |
| Stripe Radar | 2016 | Online payments fraud | Network effect from cross-merchant card history |
| Featurespace ARIC | 2008 | Adaptive behavioural analytics | Acquired by Visa in 2024 |
| Quantexa | 2016 | Contextual decision intelligence | Entity resolution and graph analytics for AML |
| ComplyAdvantage | 2014 | Sanctions and PEP screening | Continuously updated risk database |
| ThetaRay | 2013 | Unsupervised AML for cross-border payments | Used by Santander and Mastercard |
Anti-money-laundering work has shifted from rule-based transaction monitoring (which generates extreme false-positive rates, often above 95%) toward graph-based and supervised learning systems. Quantexa, a UK firm founded in 2016, uses entity resolution to stitch together customer records across data sources and exposes networks of related accounts. HSBC has been a public reference customer. Featurespace, a Cambridge spinout, was acquired by Visa in 2024 for a reported $1 billion; its ARIC platform uses adaptive behavioural analytics to flag deviations from per-customer norms. Compliance vendors do not generally release verifiable false-positive reduction figures because the relevant numbers are confidential to each bank.
Electronic trading and algorithmic execution predate machine learning. The Knight Capital incident of August 1, 2012 remains the clearest reminder of operational risk in this domain: a faulty deployment to one of eight servers re-activated a dormant test routine, leading to roughly $440 million in losses over 45 minutes and forcing the firm into a rescue financing arranged by Jefferies and others. The SEC later charged Knight $12 million for violating market-access rules. Knight Capital's case is not about AI, but it shaped how regulators think about automated trading controls under Rule 15c3-5.
At JPMorgan, the LOXM execution algorithm, deployed in equities in 2017, is trained on a large historical dataset of orders and trades to optimize the timing and venue of large block executions. JPMorgan researchers Hans Buehler, Lukas Gonon, Josef Teichmann, and Ben Wood published the foundational paper on "deep hedging" in February 2018 on arXiv, framing the problem of hedging derivative portfolios under transaction costs and trading frictions as a reinforcement-learning task. A 2019 expanded paper added market frictions and a 2022 "Deep Bellman Hedging" paper introduced an actor-critic algorithm. JPMorgan has confirmed in public talks that it uses these models in production for some vanilla index-option books.
The public quant funds keep most details opaque. What is reasonably documented is that Two Sigma hired former Google scientist Mike Schuster (a co-creator of bidirectional RNNs) in 2018 and that the firm uses deep sequence models in its research stack. Citadel and Citadel Securities have been more reticent about specifics but openly recruit from machine-learning research groups and run substantial GPU clusters. Market-making firms including Jump Trading, Hudson River Trading, Optiver, IMC, and Jane Street all describe machine learning as part of their pricing and risk infrastructure in published recruiting and engineering content.
Academic claims of profitable deep-learning stock prediction should be treated with skepticism. Systematic reviews in journals such as Expert Systems with Applications and Artificial Intelligence Review have repeatedly found that many published models are evaluated without transaction costs, on overlapping training and test sets, or with leak from future data. A 2022 survey by Hu, Liu, and others in Artificial Intelligence Review called for finance-aware evaluation frameworks because most academic stock-prediction papers do not test whether their edge survives realistic costs.
Robo-advisors automate portfolio construction and rebalancing for retail investors, typically using modern portfolio theory under the hood rather than anything more exotic. Betterment, founded in 2008, reported $56.4 billion in assets under management in 2024, making it the largest US independent robo-advisor. Wealthfront, founded in 2008 as kaChing and rebranded in 2010, manages roughly $42.9 billion. Both firms charge 0.25% of assets per year for their core offerings. Their actual portfolio choices are not particularly AI-driven; the marketing language has shifted toward AI in 2023 and 2024, but the core allocation engines are tax-loss harvesting and goal-based rebalancing.
BlackRock sits at the high end of the market. Its Aladdin platform, which the company says runs across roughly $25 trillion in assets under administration as of December 2025, is more risk management than AI. After acquiring eFront in 2019, BlackRock added private-markets analytics; in 2024 it integrated an eFront Copilot generative-AI assistant. Aladdin Wealth launched an AI-enabled commentary tool with Morgan Stanley in 2024 to draft portfolio narratives.
Morgan Stanley's wealth-management deployment is the most-watched. AI @ Morgan Stanley Assistant, launched in September 2023, is built on OpenAI's GPT-4 and indexes more than 100,000 internal research documents and policy materials. The firm later reported that more than 98% of advisor teams were using the tool. A second product, AI @ Morgan Stanley Debrief, was launched in 2024 and uses Whisper plus GPT-4 to summarize Zoom client meetings into CRM notes and follow-up drafts.
Bloomberg's BloombergGPT paper, released March 30, 2023, was the first widely cited domain-specific financial LLM. The 50-billion-parameter model was trained on a 363-billion-token Bloomberg corpus combined with 345 billion tokens of public text, and outperformed similar-size general models on financial benchmarks including FPB, FiQA SA, and Headline classification. Bloomberg later integrated parts of the work into Bloomberg Terminal's AI-Powered News Summaries feature.
JPMorgan's IndexGPT, announced in 2024 under the brand Quest IndexGPT, uses GPT-4 to translate investment themes into keyword sets, then runs a separate natural-language model over news articles to identify constituent companies for thematic baskets. JPMorgan also rolled out its internal LLM Suite to roughly 200,000 employees by early 2025, making it the largest disclosed internal generative-AI deployment at a US bank. The platform connects to OpenAI and Anthropic models behind the firm's own gateway and is used for document drafting, code, earnings-call analysis, and email summarization.
Goldman Sachs followed a similar path. The firm completed the rollout of a generative-AI code-assistance tool to thousands of developers in mid-2024 and began deploying the GS AI Assistant firm-wide in 2025. Goldman's setup routes prompts through an internal compliance gateway with prompt filtering and data masking before requests reach OpenAI, Google, or Meta models.
Klarna, the Swedish buy-now-pay-later firm, generated the most public attention for a customer-facing deployment. In February 2024, Klarna announced that an OpenAI-powered assistant had handled 2.3 million customer conversations during its first month, roughly two-thirds of its customer-service volume. Klarna said the bot did work equivalent to about 700 full-time agents, reduced repeat inquiries by 25%, and contributed an expected $40 million in profit improvement. The claims were widely questioned. In May 2025, Klarna's CEO Sebastian Siemiatkowski told Bloomberg the firm was hiring human agents back after customer-quality complaints, an episode that became a reference point for the limits of full-replacement chatbots in financial services.
| Product | Vendor | Year launched | Underlying model | Primary use case |
|---|---|---|---|---|
| BloombergGPT | Bloomberg LP | 2023 (research) | 50B parameter custom LLM | Financial NLP benchmarks, internal tooling |
| AI @ Morgan Stanley Assistant | Morgan Stanley | September 2023 | GPT-4 | Advisor research retrieval |
| AI @ Morgan Stanley Debrief | Morgan Stanley | 2024 | Whisper + GPT-4 | Meeting transcription and CRM notes |
| AskResearchGPT | Morgan Stanley | 2024 | GPT-4 | Institutional research retrieval |
| LLM Suite | JPMorgan Chase | Summer 2024 | OpenAI, Anthropic via gateway | General-purpose employee tool, 200K+ users |
| Quest IndexGPT | JPMorgan Chase | 2024 | GPT-4 plus internal NLP | Thematic index construction |
| GS AI Assistant | Goldman Sachs | 2024 to 2025 | OpenAI, Google, Meta via gateway | Bankers, traders, asset managers |
| eFront Copilot | BlackRock | 2024 | Generative AI on private markets data | Private markets analytics |
| Klarna AI Assistant | Klarna | February 2024 | OpenAI | Consumer customer service (partially rolled back 2025) |
| AlphaSense | AlphaSense | 2011 (AI features 2023) | Custom plus partner LLMs | Market intelligence search |
| Hebbia | Hebbia | 2020 | Partner LLMs | Document research for finance and law |
Research platforms aimed at financial analysts grew quickly during 2023 and 2024. AlphaSense, founded in 2011, aggregates SEC filings, expert call transcripts, broker research, and news; its AI search and summarization layer expanded after a 2023 acquisition of Tegus. Hebbia, founded in 2020, focuses on multi-step document research over user-provided corpora and is used widely by buy-side analysts and law firms.
Insurance was an early adopter of predictive modeling for actuarial pricing and is now among the most aggressive deployers of computer vision and LLM tools. The Bank of England and FCA's 2024 survey found the insurance sector with the highest AI usage of any UK financial industry, at 95%.
Lemonade, founded in 2015, built its renters and pet insurance products on a chatbot interface from launch. Its AI Maya handles intake and quoting; AI Jim handles claims first notice of loss. Lemonade has stated that the majority of its first notices of loss are taken without human intervention. The company's high-AI-content marketing has occasionally backfired; in 2021, a viral Twitter thread about its claims AI prompted regulatory scrutiny, and Lemonade clarified that its models do not use facial-analysis features to deny claims, only voice transcription.
Tractable, a London-based vendor founded in 2014, applies computer vision to auto-damage photos. The company says its models process around $2 billion worth of vehicle repairs per year across major insurers. Geico, Tokio Marine, Mitsui Sumitomo, and Polish carrier Beesafe have been disclosed customers. Tractable expanded into property damage assessment in 2022 with a system that can produce repair estimates from drone or smartphone photos.
| Vendor / insurer | Country | AI focus | Notable deployment |
|---|---|---|---|
| Lemonade | US | AI-native distribution and claims | AI Maya intake and AI Jim claims |
| Tractable | UK / US | Computer vision for damage assessment | Auto claims for Geico, Tokio Marine, others |
| Shift Technology | France | AI for fraud, claims, underwriting | Generative-AI claims handling on Azure OpenAI |
| Cape Analytics | US | Aerial and satellite imagery underwriting | Property risk scoring from imagery |
| Zesty.ai | Canada | Wildfire and climate risk modeling | Cited by California insurers and the CDI |
| Hyperexponential | UK | Pricing decision platform | Used by specialty London-market reinsurers |
| Akur8 | France | Actuarial pricing automation | GLM and ML for non-life pricing |
Shift Technology, a Paris-based vendor founded in 2014, sells claims-fraud detection and underwriting risk-detection tools used by more than 100 insurers globally. The firm has been a public Microsoft Azure OpenAI customer, using GPT-class models inside its fraud-investigation workflow. Cape Analytics applies computer vision to satellite and aerial imagery to assess property condition, roof quality, and lot features for property underwriting.
Consumer personal-finance apps were among the first to apply LLMs to user-facing chat. Mint, the budgeting app launched in 2006 and acquired by Intuit in 2009, used machine learning mainly for transaction categorization. Intuit retired Mint in early 2024 and pushed users to Credit Karma. The vacuum it left attracted alternatives.
Cleo is a London-based budgeting app built around a chatbot persona. It uses LLM-style chat over user account data to answer spending questions and nudge saving behavior, and connects to underlying accounts through aggregators including Plaid. Monarch Money, founded by former Mint product staff, connects to more than 13,000 financial institutions and uses machine learning for transaction categorization and cash-flow forecasting. Copilot Money and Origin are similar in shape, with stronger emphasis on iOS-native experiences. None of these apps disclose model details, and all are heavily dependent on Plaid or comparable data infrastructure for account linkage.
Plaid itself plays an important infrastructure role. The company uses LLMs internally to label anonymized transactions for merchant identification and category tagging, and in 2024 began exposing AI-friendly tooling, including a Plaid CLI for AI agents and a Model Context Protocol server that allows Anthropic and OpenAI agents to call its APIs.
Large enterprises have rolled AI into corporate finance work, mostly in accounts payable, expense reporting, and contract review. Vic.ai, founded in 2017 by Alexander Hagerup, sells an autonomous accounts-payable platform that the company says is trained on more than one billion invoices. Tipalti, founded in 2010, added an AI assistant for finance teams in 2024 that surfaces invoice and PO data through a chat interface. AppZen and Anaplan have similar offerings in expense audit and planning, respectively. These tools mostly automate previously manual classification work; their bigger effect is restructuring finance-team headcount rather than producing new insight.
Financial regulation around AI is fragmented and changes quickly.
United States. The Federal Reserve's SR 11-7 guidance, issued April 4, 2011, defines a "model" broadly enough to cover almost any machine learning system used in lending, capital, valuation, or operational decisions. Banks are expected to maintain a model inventory, validate models independently, and monitor performance. A recurring 2024 to 2025 examination finding is that supervised LLMs deployed in customer service or compliance had not been added to model inventories. The Consumer Financial Protection Bureau issued chatbot guidance in June 2023 and an adverse-action-notice circular on September 19, 2023, both signaling that AI does not change the underlying duty to give borrowers specific reasons for denials.
The SEC under Chair Gary Gensler proposed a predictive data analytics rule on July 26, 2023, that would have required broker-dealers and advisers to eliminate or neutralize conflicts of interest in any AI or predictive system used in investor interactions. The proposal drew unusually heavy industry opposition, including a withdrawal request from the Investment Adviser Association, and was formally withdrawn in June 2025.
Europe. The EU AI Act, finalized in 2024, classifies AI systems for evaluating natural-person creditworthiness, setting credit scores, and pricing life and health insurance as high-risk. Fraud-detection systems are explicitly exempted. High-risk systems require conformity assessment, technical documentation, data governance, and registration in the EU database. The Digital Omnibus on AI agreed in May 2026 proposed deferring the Annex III stand-alone high-risk deadline to December 2, 2027.
United Kingdom. The Bank of England and the FCA have run a joint AI in financial services survey three times (2019, 2022, 2024). The November 2024 report found 75% adoption, identified third-party model dependency as a growing concern, and noted that 46% of firms had only partial understanding of the AI they were using because so much of it was outsourced. The FCA has also signalled it will police AI use through existing principles (Consumer Duty, SM&CR) rather than create new AI-specific rulebooks.
International. The Financial Stability Board's November 2024 report, The Financial Stability Implications of Artificial Intelligence, identified four systemic-risk channels: third-party concentration in model providers, increased market correlation as firms adopt similar models, cyber risk, and traditional model risk. The BIS and the ECB have published parallel work. The FSB issued a follow-up monitoring note in October 2025.
The largest risk in financial AI is not a single catastrophic event; it is the accumulation of model decisions whose accountability is unclear. A few episodes are worth singling out because they have shaped policy.
Knight Capital (August 1, 2012). Not an AI incident, but the first widely studied case of an automated trading system causing a near-fatal loss at a major broker. The firm lost roughly $440 million in 45 minutes due to a deployment error that re-activated a dormant test routine. Regulators cite Knight in almost every discussion of automated trading controls. The SEC's Rule 15c3-5 market-access controls were already in force at the time but did not prevent the incident, which led to more aggressive examination of pre-trade risk checks.
Apple Card / Goldman Sachs (November 2019). Allegations of gender bias in credit limits set off a New York DFS investigation. The agency found no statutory violation but criticized the lack of customer transparency. The episode is still the most cited example of how poorly explained algorithmic credit decisions can produce reputational and regulatory damage even when they are technically lawful.
Optiver and other market-maker fines. Optiver paid a $14 million CFTC penalty in 2008 for crude-oil market manipulation, before machine learning was central. More recent fines against high-frequency firms have generally been for spoofing or quote stuffing rather than for AI-specific behavior, but FINRA and the CFTC have warned that surveillance systems must keep pace with the speed and complexity of trading algorithms.
Klarna AI assistant rollback (2024 to 2025). Klarna's early announcements claimed the AI was doing the work of 700 agents, but by mid-2025 the company publicly stated that quality issues had pushed it back toward human-led service. The reversal is now a standard caution in industry commentary about the limits of full chatbot replacement in regulated consumer-finance contexts.
Model herding and correlation. A recurring concern in the FSB and BIS literature is that if banks and asset managers all use the same handful of foundation models, supplied by the same handful of providers, the financial system gains a new concentration risk. The Bank of England has flagged this in its 2024 survey: a third of all AI use cases in UK firms are third-party implementations. A single model bug, a sudden change in fine-tuning, or a vendor outage could affect many institutions at once.
Explainability is the other persistent issue. Deep learning models can outperform linear baselines on credit, fraud, and pricing tasks, but they are difficult to explain to a regulator or to a denied applicant. Tools like SHAP and LIME provide post-hoc rationales, but most regulators treat them as helpful rather than sufficient. The general direction of US, EU, and UK guidance is to push more responsibility back onto the deploying institution: if you can't explain the decision, you should not have made it that way.