History of AI in Finance: Before GenAI

IBM 650 Computer

Overview

The U.S. financial industry has long been an early-adopter of innovative technologies due to the vast amount of data amassed from its daily transactions and operations. In 1954, banks and insurers popularized the new IBM 650 — the world’s first mass produced, profitable business computer — to automate tedious accounting tasks. In the decades since, simple rules-based algorithms have given way to more advanced systems spurred on by advancements in machine learning and neural networks that harvest data to continuously improve outputs and hasten decision-making. Today’s frontier of development concerns Generative AI, specifically Large Language Models (LLMs), with its use cases in the financial sector being fleshed out as we speak (more on that in the individual “GenAI Use Cases” pages). Below is a brief evolution of the regulatory and technological changes that have brought us to where we are today: $97 billion is expected to be invested in AI across banking, insurance, capital markets and payments by 2027 (McKinsey), and 70% of financial services executives believe that AI will directly tie to revenue growth in upcoming years (World Economic Forum).

Payments: Automation and Fraud Detection

1950/60s: Banks began using mainframes like the IBM 650 to process checks and transactions in bulk​ (BIS). The advent of credit cards created new payment networks and massive data streams, prompting use of algorithms for tasks such as authorizing charges and early fraud rules.

1980s/90s: As electronic payments volume grew, so did fraud. Payment companies and banks started exploring AI to detect suspicious transactions more effectively. Initial approaches involved expert systems encoding fraud heuristics (e.g. sudden spending sprees or atypical merchant categories), but these had limited success. A major turning point came in the early 1990s with the application of neural networks to credit card fraud detection. In 1992, HNC Software (co-founded by neuroscientist Robert Hecht-Nielsen) introduced the Falcon Fraud Manager, a groundbreaking system that evaluated card transactions in real time using trained neural network models (​KDnuggets) (​ Fresh Brewed Tech).

This neural-network approach became an industry standard; by the mid-1990s, most U.S. credit card companies were using AI models to score each authorization request for fraud risk (​KDnuggets). Payments thus became one of the earliest commercial success stories for AI, using big data to safeguard trillions of dollars in transactions.

2000s: Machine learning (ML) models beyond neural nets (e.g. decision trees and ensemble methods) were adopted to further reduce false positives and adapt to evolving fraud tactics.

Beyond fraud, AI also found roles in payments back-office operations. Routing optimizations and liquidity management benefited from ML models that forecast cash flows and settlement needs, helping payment networks and banks manage liquidity in real time (e.g. predicting which ATM or branch needs cash) (VolanteTech). Regulators encouraged these advances: for instance, the U.S. Financial Crimes Enforcement Network in 2018 issued guidance encouraging banks to explore innovative AI techniques for anti-money laundering compliance.

Asset Management: From Quantitative Models to Robo-Advisors

1950s-80s: Harry Markowitz’s Modern Portfolio Theory introduced mathematical optimization for portfolio selection – a foundational algorithm (though not AI, it set a tone for data-driven decision making). By the 1970s, large financial firms were using mainframes to perform risk analysis and optimize portfolios under various scenarios. The first instances of computer-aided trading emerged as exchanges computerized some operations; for example, program trading and index arbitrage (using software to exploit price differences) were seen on Wall Street by the late 1970s and early 1980s. However, these strategies followed fixed rules or formulas rather than adaptive “intelligent” algorithms. A notable development in the mid-1980s was Wall Street’s exploration of expert systems for complex pricing problems – Lehman Brothers was reported to be developing an AI system to evaluate interest rate swap prices in the mid-1980s, one of the first such AI forays by a major firm ​(Winton)

Around the same time, a new breed of quantitative hedge funds was born: Renaissance Technologies (founded 1982 by James Simons) and D.E. Shaw & Co. (founded 1988 by David Shaw) are prominent examples. These firms employed advanced statistics and computer science (including early machine learning techniques) to systematically trade securities ​(Winton). This era established the principle that algorithms (and by extension AI) could confer an edge in asset management, particularly as market data and electronic trading grew.

The late 1980s and 1990s marked a pivotal period where academic research and industry experiments brought machine learning into asset management. With the invention of backpropagation (1986) making neural networks feasible, researchers began applying them to financial prediction. Early successes included using neural nets to forecast corporate bankruptcy from financial statements and to predict stock price movements from historical data ​(Restack)

1990s: Several financial firms began piloting AI models, but mainstream uptake was gradual; traditional asset managers remained cautious due to the unpredictability of AI in volatile markets (the 1987 “Black Monday” crash was exacerbated by automated trading algorithms). Still, by the late 90s, AI had proven valuable in certain niches: commodity trading advisors used neural nets for price pattern recognition, and some mutual funds employed AI to allocate assets or screen stocks.

Fidelity Investments in the late 1990s rolled out an online “Portfolio Review” tool using a rule-based expert system to suggest fund allocations to clients, foreshadowing the robo-advisors to come.

2000s: The 21st century saw an explosion of data in finance – news, social media, high-frequency market data – and AI techniques evolved to harness it. Support vector machines (SVMs) and decision-tree ensembles (Random Forests, Gradient Boosting) joined neural networks as tools for asset pricing forecasts and risk modeling. Hedge funds like Bridgewater, Two Sigma, and Citadel invested heavily in machine learning talent to enhance their trading algorithms. At the same time, computational power allowed high-frequency trading (HFT) to flourish (firms executing thousands of trades per second based on algorithms). While HFT relied more on speed and less on “learning,” some firms began integrating AI to optimize execution strategies (e.g. using reinforcement learning to minimize market impact when selling large positions).

2010s: A transformative innovation was the rise of robo-advisors – automated investment advisory platforms. The first robo-advisor services, Betterment and Wealthfront, launched around 2008–2010 with the goal of providing algorithm-driven portfolio management to consumers ​(Investopedia). These platforms use online questionnaires to gauge a client’s risk tolerance and goals, then apply optimization algorithms (typically grounded in modern portfolio theory) to allocate across ETFs and periodically rebalance.

By the mid-2010s, robo-advisors had amassed billions in assets, forcing traditional wealth managers (Schwab, Vanguard, etc.) to launch their own automated offerings. In parallel, natural language processing started to play a role in asset management (The Blueberry Fund). Firms began analyzing news articles, earnings call transcripts, and even social media sentiment to inform trading decisions. For example, hedge funds deployed NLP models to gauge market sentiment from news feeds (a form of AI-driven alternative data analysis). In 2013, the SEC enabled companies to use social media for disclosures, and shortly after, traders were mining Twitter (e.g. tweets from CEOs or broader sentiment) using AI to predict stock moves. Deep learning gained attention after 2012; by late 2010s, deep neural networks were being tested for tasks like stock trend prediction, options pricing, and portfolio risk estimation.

Banking and Financial Intermediation: Credit Scoring, Lending, and Service

1960s-90s: In banking, one of the earliest uses of computational algorithms was credit scoring – quantifying a borrower’s credit risk with a numeric score. The concept gained traction over the 1960s as banks realized computers could consistently analyze large loan portfolios. By the 1970s, regulatory changes spurred wider use of credit scoring – notably the Equal Credit Opportunity Act of 1974, which barred lending discrimination on race, gender, and other grounds, prompting banks to use data-driven scoring models as a defense against bias claims (Jora Credit). By 1995 it became mainstream when Fannie Mae and Freddie Mac (the giant mortgage GSEs) required FICO scores on all conforming mortgage applications. In 1997 Frannie Mae and Freddie Mac fully launched Desktop Underwriter (DU) and Loan Prospector (LP), AI-driven underwriting platforms that they distributed to lenders nationwide​ (predatorylending.duke.edu). Automated underwriting cut mortgage approval times from weeks to minutes, reduced human error, and enforced consistent standards ​(predatorylending.duke.edu). Many banks and mortgage companies adopted these or built their own variants, making algorithmic loan decisioning an industry norm by the 2000s.

In the 1990s as Neural Networks made inroads, banks started deploying AI to detect suspicious patterns in checking account transactions or unusual access to accounts, using similar neural-network techniques as credit card fraud systems.

2000s: As banking services moved online, AI found new applications in front-office customer interaction and internal process automation. Call centers began using basic NLP in the form of phone menu “voice response” systems that could understand spoken requests (Conversation Design Institute). By the 2010s, banks introduced more advanced virtual assistants or chatbots. For instance, Bank of America’s chatbot Erica (launched 2018) uses AI/NLP to help customers with tasks like checking balances, scheduling payments, or even providing personal finance advice.

2010s: A critical area of use has been anti-money laundering (AML) and “Know Your Customer” compliance. Traditionally, AML transaction monitoring in banks relied on hard-coded rules (e.g. “flag any transfer above $10,000” or “if an account receives multiple international wires then…”) which produced many false positives. In the 2010s, major banks started adding machine learning on top of these systems to prioritize alerts and detect complex behaviors that rules missed.​ (McKinsey​) Unsupervised learning models, for example, could identify a network of accounts engaged in an unusual pattern of circular money movement – a possible laundering ring – which individual rule triggers might not catch. U.S. regulators actually encouraged this innovation; a 2018 joint statement by the Treasury, Fed, FDIC, and OCC urged banks to consider “innovative approaches” like AI to strengthen AML compliance​ (McKinsey).

In banking (particularly investment banking), AI also influenced trading and risk management similar to the asset management story. Large banks’ trading desks in the 2000s deployed algorithmic trading bots for market-making and hedging. Post-2008, banks had to enhance risk oversight, and AI helped by stress-testing portfolios under simulated scenarios more efficiently. For example, JP Morgan developed a program called LOXM in 2017 (an AI for trade execution) and Goldman Sachs invested in AI to automate parts of its trading and reduce costs. Additionally, market risk models started to incorporate machine learning to better account for nonlinear risks. One academic contribution was using support vector machines to predict bank failures or rating downgrades, which informed risk officers at banks and regulators.

Insurance

1980s: Insurers began experimenting with expert systems to automate underwriting – the process of evaluating insurance applications and determining coverage terms. A landmark project was CLUES (Comprehensive Life Underwriting Expert System) developed by the Mutual Life Insurance Company of New York around 1986​. CLUES was a rule-based AI system designed to replicate the reasoning of experienced life insurance underwriters. It took in an applicant’s data (age, medical history, lifestyle, etc.) and ran it through hundreds of coded underwriting rules – about 800 rules amounting to 8,000 logical evaluation steps – to arrive at a decision. Around the same time, property and casualty (P&C) insurers were also exploring expert systems.

1990s/2000s: Forward-looking insurers, who traditionally relied on statistical models, began to incorporate machine learning techniques to find additional patterns in underwriting and claims. Decision tree algorithms were applied to auto insurance data to segment customers by risk more finely than traditional rating tiers. Neural networks were tested for fraud detection in claims – identifying telltale signs of fraudulent claims by training on past known fraud cases. Early successes in fraud analytics for insurance mirrored those in banking: by late 1990s, companies like CAP Index and others were offering data-driven fraud scoring for property claims. Health insurers used neural nets to flag anomalies in medical billing that could indicate provider fraud (e.g. upcoding or phantom billing). Meanwhile, life insurers in the late 90s introduced predictive models to supplement underwriting for certain products – for instance, using an applicant’s credit score or other non-medical data as inputs to mortality risk models

On the regulatory side, the use of new data and models in insurance began to be monitored for fairness. State insurance regulators required that any underwriting or rating factor be justifiable – which in practice meant that black-box AI models had to be used carefully. As a result, insurers in the 2000s typically confined AI to advisory roles or internal risk assessment, while final premium calculations still often relied on interpretable models.

2010s: claims processing became the hotbed of AI innovation in the 2010s. Processing an insurance claim involves verifying details, detecting fraud, determining payouts, and coordinating service (repairs, medical care, etc.). Insurers deployed machine vision and NLP to expedite claims. For example, after an auto accident, instead of waiting for an adjuster, policyholders could upload photos of vehicle damage; AI models (trained on thousands of crash images) analyze the damage and predict repair costs, enabling immediate preliminary estimates. Several big auto insurers (like Progressive and Allstate) partnered with startups in computer vision by 2016–2018 to roll out such image-based estimating tools.

During the late 2010s, many insurers launched virtual assistants for customers. Geico’s app introduced an AI virtual assistant (“Kate”) to answer policy questions. Progressive’s Flo chatbot (leveraging its marketing character) could help file simple claims or give quotes. These AI-driven interfaces improved response times and cut customer service costs.