a16z 10,000-Word Deep Dive: How Can Financial Services Leverage Generative AI?
TechFlow Selected TechFlow Selected
a16z 10,000-Word Deep Dive: How Can Financial Services Leverage Generative AI?
Let's delve into the five goals of personalized consumer experiences, efficient operations, better compliance, improved risk management, and dynamic forecasting and reporting to see how established companies and startups are leveraging generative AI.
Navigating Web3 tides with focused insightsAuthors: Angela Strange, Seema Amble, et al., a16z
Translation: InvestmentAI
Today, as technology advances at an unprecedented pace, AI (artificial intelligence) and ML (machine learning) have already been applied in the financial services industry for over a decade—ranging from improved risk controls to foundational anti-fraud scoring. Now, generative AI based on large language models (LLMs) represents a historic leap forward, transforming fields such as education, gaming, and business. Unlike traditional AI/ML, which primarily focuses on prediction or classification using existing data, generative AI can create entirely new content.
Imagine training on vast amounts of unstructured data with nearly unlimited computing power—this could bring about the biggest transformation in financial services markets in decades. Unlike previous platform shifts—such as the internet, mobile, or cloud—where financial services often lagged behind, here we expect both emerging startups and established players to embrace generative AI immediately.
Financial services companies possess massive historical financial data. By leveraging this data to fine-tune LLMs (or even train them from scratch like BloombergGPT), they could rapidly answer almost any financial question. For example, an LLM trained on customer chat logs and additional product specifications should be able to instantly answer all inquiries about a company’s products. An LLM trained on ten years of suspicious activity reports (SARs) should be able to identify transaction patterns indicative of money laundering. We believe that financial services are poised to leverage generative AI toward five key goals: personalized consumer experiences, efficient operations, better compliance, improved risk management, and dynamic forecasting and reporting.
In the competition between incumbents and newcomers leveraging AI to launch new products and enhance operations, incumbents will enjoy an early advantage due to access to proprietary financial data, but may eventually be hindered by high standards for accuracy and privacy. In contrast, new entrants might initially rely on public financial data to train their models, but will quickly begin generating their own data, using it as a wedge to distribute new products.
Let’s now dive into these five goals and examine how both incumbents and startups are harnessing generative AI.
Personalized Consumer Experiences
Despite tremendous success among consumer fintech companies over the past decade, they have yet to fulfill their most ambitious promise: optimizing consumers’ balance sheets and income statements without human intervention. This promise remains unfulfilled because user interfaces fail to fully capture the human context influencing financial decisions, and cannot offer advice or cross-selling in ways that help people make appropriate trade-offs.
One non-obvious example of critical human context is how consumers prioritize bill payments during difficult times. When making such decisions, consumers typically weigh utility and brand loyalty—two intertwined factors that complicate creating an experience capable of fully capturing optimal decision-making. This makes delivering top-tier credit education without human employees particularly challenging. While experiences like Credit Karma can guide customers through 80% of the journey, the remaining 20% feels like a magical abyss—further attempts to capture context often become too narrow or rely on false precision, undermining consumer trust.
Similar shortcomings exist in modern wealth management and tax preparation. In wealth management, human advisors outperform fintech solutions—even those narrowly focused on specific asset classes and strategies—because individuals are deeply influenced by unique hopes, dreams, and fears. This is why human advisors have historically provided more tailored advice than most fintech systems. In taxation, despite modern software, Americans still spend over 6 billion hours annually on taxes, commit 12 million errors, and frequently miss income or overlook benefits they don’t know about—such as deducting work-related travel expenses.
Large language models (LLMs) offer an elegant solution to these problems by better understanding and navigating consumers’ financial decisions. These systems can answer questions (“Why do I hold municipal bonds in my portfolio?”), evaluate trade-offs (“How should I think about term risk versus yield?”), and ultimately incorporate human context into decision-making (“Can you create a flexible plan to help me financially support my aging parents at some future point?”). These capabilities should transform consumer fintech from a set of high-value but narrow use cases into applications that help consumers optimize their entire financial lives.
Efficient Operations
In a world where generative AI tools permeate banking, Sally should undergo continuous credit evaluation so that when she decides to buy a house, she already has a pre-approved mortgage.
Yet this world has not yet materialized, for three main reasons:
-
First, consumer information is scattered across disparate databases. This makes cross-selling and predicting consumer needs extremely challenging.
-
Second, financial services are highly emotional purchases, often involving complex and hard-to-automate decision trees. This means banks must employ large customer service teams to answer numerous questions about which financial products best suit individual circumstances.
-
Finally, financial services are heavily regulated. This requires human staff such as loan officers and processors to remain involved in every cycle of available products (e.g., mortgages) to ensure compliance with complex, unstructured laws.
Generative AI will increase the efficiency of labor-intensive tasks—such as extracting data from multiple sources, understanding unstructured personal contexts, and interpreting unstructured compliance regulations—by a thousandfold. For instance:
-
Customer Service Representatives: At every bank, thousands of representatives must deeply understand the bank's products and related compliance requirements to answer customer queries. Now imagine a new representative who can use an LLM trained on the past decade of the bank’s customer service calls. This agent could use the model to quickly generate correct answers to any question, enabling deeper discussions across broader product lines while reducing training time. Incumbents will want to ensure their proprietary data and specific customer information aren't used to improve general-purpose LLMs accessible to competitors. New entrants will need to be creative in constructing their datasets.
-
Loan Officers: Existing loan officers typically pull data from nearly a dozen different systems to compile a loan file. A generative AI model trained on all these systems’ data would allow an officer to simply input a customer name and instantly generate the full loan file. Human oversight may still be needed to ensure 100% accuracy, but the data collection process would become vastly more efficient and accurate.
-
Quality Assurance: Much of the QA work at banks and fintech firms involves ensuring full compliance with numerous regulatory bodies. Generative AI can dramatically accelerate this process. For example, Vesta could use a generative AI model trained on Fannie Mae’s selling guide to immediately alert mortgage officers to compliance issues. Since many regulatory guidelines are publicly available, this could provide an interesting entry point for new market entrants. However, real value will still accrue to companies with robust workflow engines.
All of these steps move us closer to a world where Sally can instantly receive a potential mortgage approval.
Better Compliance
In the future, if compliance departments adopt and utilize generative AI, the $800 billion to $2 trillion in illicit money laundering occurring globally each year could be effectively curbed. Drug trafficking, organized crime, and other illegal activities might consequently see their largest decline in decades.
We currently spend hundreds of billions of dollars annually on compliance, yet only prevent around 3% of criminal money laundering. Most compliance software is built on “hard-coded” rules. For example, anti-money laundering systems allow compliance officers to implement rules like “flag any transaction over $10,000” or search for other predefined suspicious activities. But applying these rules often proves ineffective, as many institutions are legally required to investigate a large volume of false positives, which are often complex and cumbersome to handle. To avoid severe penalties, compliance departments hire thousands of employees—often exceeding 10% of a bank’s total workforce.
With generative AI, the future could look very different:
-
More Efficient Screening: Generative AI models can rapidly aggregate key information about any individual across various systems and present it to compliance officers, enabling faster risk assessment of transactions.
-
Better Prediction of Money Launderers: Imagine a model trained on the past 10 years of SARs that, without explicit instructions, discovers new patterns in the reports and autonomously identifies behavioral patterns indicative of money laundering.
-
Faster Document Analysis: Compliance departments must ensure adherence to internal policies and external regulations. Generative AI can analyze vast volumes of documents—contracts, reports, emails—and flag potential issues or areas requiring deeper investigation.
-
Training and Education: Generative AI can also develop training materials and simulate real-world scenarios to teach compliance officers best practices and how to recognize potential risks and non-compliant behaviors.
New entrants can leverage publicly available compliance data from dozens of institutions to make searching and integration faster and easier. Meanwhile, large incumbents with years of accumulated data will need to design appropriate privacy safeguards.
Improved Risk Management
While Archegos and the London Whale sound like creatures from Greek mythology, they actually represent catastrophic failures in risk management that cost the world’s largest banks billions of dollars. Combined with recent examples like Silicon Valley Bank, it’s clear that risk management remains a major challenge for many leading financial institutions.
While advances in AI cannot completely eliminate credit, market, liquidity, and operational risks, we believe this technology can play a significant role in helping financial institutions identify, plan for, and respond to these inevitable risks more quickly. Specifically, here are several areas where we believe AI can enhance risk management efficiency:
-
Natural Language Processing: LLMs like ChatGPT can process vast volumes of unstructured data—news articles, market reports, analyst research—to provide a more comprehensive view of market and counterparty risks.
-
Real-Time Insights: Immediate awareness of market conditions, geopolitical events, and other risk factors enables companies to adapt more quickly to changing environments.
-
Predictive Analytics: The ability to run more complex scenarios and provide early warnings can help companies manage risk more proactively.
-
Integration: Consolidating siloed systems and using AI to unify information can provide a more complete picture of risk exposure and streamline risk management workflows.
Dynamic Forecasting and Reporting
Beyond solving financial questions, LLMs can also help financial services teams improve internal operations and simplify daily tasks. Despite progress in other areas of finance, modern finance teams still rely heavily on Excel, email, and manual business intelligence tools. Automation of basic tasks is hampered by shortages in data science resources, leaving CFOs and their teams bogged down in tedious record-keeping and reporting instead of focusing on strategic decision-making.
Generally, generative AI can help these teams pull data from more sources and automate trend identification, forecasting, and reporting. Here are some concrete applications:
-
Forecasting: Generative AI can assist in writing formulas and queries in Excel, SQL, and BI tools to automate analysis. These tools can also uncover patterns, extract predictive signals from larger, more complex datasets (including macroeconomic factors), and suggest ways to refine models for better decision-making.
-
Reporting: Instead of manually compiling information for reports (board presentations, investor updates, weekly dashboards), generative AI can automatically generate text, charts, and visuals, adapting flexibly based on templates.
-
Accounting and Taxation: Accounting and tax teams spend significant time interpreting regulations and applying them to real situations. Generative AI can summarize, synthesize, and suggest possible interpretations of tax laws and potential deductions.
-
Procurement and Accounts Payable: Generative AI can help auto-generate and adjust contracts, purchase orders, invoices, and reminders.
However, it's important to note that current generative AI still has limitations in domains requiring judgment or precise answers—areas that are often essential for finance teams. While generative AI models continue to advance computationally, we cannot yet fully rely on their accuracy and must still perform human review. As models rapidly improve—with more training data and integration of mathematical modules—new use cases will emerge.
Challenges
Within these five major trends, both new entrants and incumbent market participants face two primary challenges in realizing this generative-AI-powered future.
-
Training Large Language Models (LLMs) on Financial Data: Current LLMs are primarily trained on web data. To meet the specific needs of financial services, these models must be fine-tuned using financial data. New entrants may start with public corporate financial data, regulatory filings, and other easily accessible public financial data, progressively optimizing their models and eventually incorporating proprietary data they collect. Incumbents such as banks—or large platforms with financial services arms (e.g., Lyft)—can leverage their existing proprietary data, giving them an initial edge. However, established financial institutions tend to be overly cautious when adopting new platform shifts—giving unconstrained newcomers a competitive advantage.
-
Accuracy of Model Outputs: Given that answers to financial questions can impact individuals, companies, or even society at large, these new AI models must be as accurate as possible. They must not fabricate incorrect answers or deliver confidently stated but false responses. For critical issues involving taxation or financial health, they must be far more accurate than for pop culture queries or generic high school essays. Initially, a human-in-the-loop will likely be necessary to validate AI-generated outputs.
The rise of generative AI undoubtedly represents a major platform shift for financial services companies—one that could deliver more personalized solutions to customers, make operations more cost-effective, improve compliance, enhance risk management, and enable more agile forecasting and reporting. On the two key challenges outlined above, incumbents and startups will compete fiercely. While we don’t yet know who will ultimately win, one thing is certain: there is already a clear winner—the future consumer of financial services.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
a16z
