A Policy Memo on AI in Financial Services

Takeaway

Generative AI (Gen AI) holds the potential to democratize access to financial services and enhance the efficiency of US banks.

Policy should continue to:

• Enable the adoption of new technologies to benefit millions of underbanked American consumers and small businesses (financial inclusion, SMB development, competitiveness of US banks)
• Provide guidance for model risk management (MRM) and data protection
• Promote competition so that the gains do not accrive only to the biggest tech companies and biggest banks

Background

Tens of thousands of organizations are using generative AI to become more productive. For instance, 1.5 million programmers across 27,000 organizations use Microsoft’s GitHub co-pilot to write code 55% faster. 88% feel more productive and 96% are faster with repetitive tasks. This is the fastest growing product in GitHub’s history. It is a tool that increases the productivity of Americans. Gen AI has the potential to do the same for financial services.

Machine learning is not new - it is already widely used in financial services for document processing e.g., OCR/digitizing and extracting data from checks, bank statements, income statements, etc and risk modeling e.g., in underwriting processes.

Possible new use cases include making loan and mortgage applications faster and less complicated through conversational AI and making back-office processing less manual and more efficient (i.e., lowering the cost to provide financial services) through AI co-pilots, just like for coding.

Current State of Financial Services

Today, financial services are far from perfect:

• Underbanked Americans: 5.9M Americans lack access to financial services (underbanked)
• Customer complaints: Over 1M complaints are registered per year at the consumer financial protection bureau.
• Offshoring operations: Banks have moved their operations out of the US to reduce cost. The combination of AI and human-in-the-loop has the potential to keep sensitive data in the US and employ Americans as qualified AI trainers, while still making our banks more competitive.

Potential of Generative AI in Financial Services

Generative AI has the potential to improve financial services by better serving millions of underbanked American consumers and small businesses, and increasing the competitiveness of US banks through better technology, instead of offshoring banking services e.g., call centers and back office processing.

Banks are eager to adopt generative AI, but are lagging behind other industries. Currently, big organizations are experimenting with gen AI e.g., BloombergGPT, FinBERT (JP Morgan), etc. Smaller organizations are looking to partner with Fintechs, bringing top talent from the best universities of the country to strengthen our community and regional banks.

MRM Guidance and AI Risks

MRM Guidance (model risk management) promotes proper monitoring of models to enable technological innovation and competition to benefit the American consumer.

Known risks of AI can be addressed:

• Without proper context, language models like ChatGPT can hallucinate answers: If you give the model proper context, you get answers rooted in this context (retrieval augmentation)
• Without human feedback, models may make mistakes: If you build human-in-the-loop systems, wrong responses get caught and corrected by humans. The model then learns from those mistakes and becomes better over time.
• Without proper monitoring, generative AI may be perceived as a black box: collecting an auditable log of model responses and the context that led to these responses allows banks and regulators to monitor how these systems improve financial services.

Policy Goals and AI Adoption

Policy goals and how AI adoption in financial services helps achieve them:

• Financial inclusion: do not gatekeep a personal banking experience to ultra-high-net-worth individuals; help the unbanked and underbanked by making financial services more affordable and accessible.
• Promoting SMB development: faster loan provision; better intelligence.
• Competitiveness of US banks: improve service quality; increase addressable market by making it profitable to serve previously underserved; manage cost.
• Keeping jobs in America: Many banks have offshored functions such as customer service to other countries to save cost - the combination of AI and human-in-the-loop has the potential to keep sensitive data in the US and employ Americans as qualified AI trainers, while still making our banks more competitive.
• Data protection: ensure adherence to existing laws and good software development practices.