AI Customer Service in Banking: Deployments and Regulatory Risks

Deploying artificial intelligence (AI) customer service in banking allows financial institutions to handle routine account inquiries, process payments, and manage card settings at lower cost per interaction. Research from the Consumer Financial Protection Bureau report shows that all ten of the largest commercial banks in the United States operate automated chat assistants.

Reports cited by the CFPB estimate $8 billion per year in cost savings, about $0.70 saved per customer interaction. Over ninety-eight million consumers used bank chatbots in 2022, with adoption projected to reach 110.9 million users by 2026.

Realizing these efficiencies depends on technical accuracy, clear dispute escalation protocols, and secure data handling. When automated tools fail, consumers face repetitive conversational loops, unaddressed account disputes, and privacy risks. The CFPB report is from 2023, and it does not say what adoption looks like beyond the 2026 projection.

What AI Customer Service in Banking Does Today

Banks use chatbots to check balances, review transactions, transfer money, lock or replace cards and take dispute requests. The CFPB's June 2023 report found that each of the top 10 largest commercial banks has deployed chatbots. Its risk findings center on accuracy, a route to a human, dispute recognition and chat-log protection.

Most top ten institutions also support direct messaging or business chat on social platforms including Twitter, Facebook, and Instagram. Morgan Stanley is testing a GPT-4-powered tool that feeds scripts to financial advisors. Goldman Sachs considered an internal "ChatGS" for employees.

A related page covers best AI tools for banking.

• Capital One Eno: Can check account balances, review recent transactions and available credit, know when payments are due, pay a bill, activate, lock or replace a card, update personal information, find account numbers, and add authorized users.
• Bank of America Erica: Used by nearly 32 million customers in over 1 billion interactions by October 2022.
• U.S. Bank Smart Assistant: Handles finding a credit score, transferring money, disputing transactions, and paying others through Zelle.
• Wells Fargo's Fargo: Uses LLMs to process input and give tailored responses.

Rule-Based Systems and Large Language Model Architectures

Most customer service chatbots deployed across the banking industry rely on rule-based logic rather than generative language models. Rule-based systems follow decision trees or keyword databases to trigger pre-scripted replies to specific inquiries. Users are typically limited to predefined inputs, often menu choices. Some complainants found FAQ redirection, a common behavior, unhelpful.

Complex conversational systems use machine learning and large language model (LLM) architectures to analyze word patterns in large datasets and simulate natural dialogue. Fargo by Wells Fargo uses this approach to process conversational input and produce custom responses for banking customers. Unlike rule-based systems, large language models predict upcoming words statistically to construct novel text replies.

Generative models create operational risks because statistical methods are not well-positioned to distinguish factually correct data from false statements. Generated text can be inaccurate, which the CFPB notes can lead to inappropriate fees, default, or other harms. Technical teams evaluating generative customer support systems can review ChatGPT customer service.

Cost Reductions Driving AI Customer Service in Banking

Commercial banks adopt automated chat tools to lower the cost of servicing routine customer inquiries. The CFPB report notes that financial institutions view automated chat as a cost-effective alternative to human customer service. Reports cited by the CFPB estimate that chatbots deliver $8 billion per annum in cost savings, approximately $0.70 saved per customer interaction.

In 2022, over 98 million consumers, representing thirty-seven percent of the United States population, used a bank chatbot. Projections cited in the report indicate adoption will grow to 110.9 million consumers by 2026, though post-2026 metrics are not published.

A related page covers AI customer service outsourcing.

Consumer Protection Risks in Deficient Banking Chatbots

Deficient banking chatbots create risks that the CFPB classifies into three categories. These are risk of noncompliance with federal consumer financial laws, risk of diminished customer service and trust, and risk of harming people. The CFPB says a bot may give inaccurate information, fail to recognize a consumer invoking federal rights, or fail to protect privacy and data.

Customer trust declines when automated systems trap consumers in repetitive conversational loops, known as doom loops. In these loops, software delivers circular scripted responses without offering a working path to a live human agent. A survey cited by the report found 80% of consumers who used a chatbot felt more frustrated. The survey also found 78% needed a human after it failed. Consumers may have limited bargaining power, for example with mortgage servicers and credit reporting companies.

Direct consumer harm occurs when inaccurate bot statements lead to inappropriate fees, which can cause default or other harms. Technology trained on a limited number of dialects makes it difficult for consumers with diverse dialect needs. The CFPB says the tradeoff from reduced human support is especially true for groups with limited technical availability or limited English proficiency, with higher rates of failed resolution. The report cites the 2018 Ticketmaster UK breach through its chatbot vendor. That incident affected 9.4 million data subjects, including 60,000 individual payment card details. The report also cites a May 2022 DHL-impersonation scam as an example of impersonation bots and phishing.

Regulatory Oversight of AI Customer Service in Banking

The CFPB is actively monitoring chatbot use and expects institutions using chatbots to do so in a manner consistent with customer and legal obligations. The CFPB says chatbots must comply with all applicable federal consumer financial laws. Legal requirements include that entities accurately identify when a customer is raising a concern or dispute.

The CFPB says chat logs holding personal data should be considered sensitive consumer information. The CFPB cites its Circular 2022-04 on insufficient data protection or security for sensitive consumer information.

The report states it is not intended to impose any obligations or define any rights, and is not an interpretation of a regulation or statute. Consumers who cannot get answers are encouraged to submit a complaint to the CFPB. A related page covers AI customer service automation.

Operational Checklist for Bank Chatbot Deployments

Questions a bank can ask before launching a chatbot, drawn from the risks the CFPB describes.

• Human Handoff: Is there a clear route to a human without unreasonable waits when automated replies trap customers in repetitive responses?
• Dispute Recognition: Can the chatbot accurately identify when a customer is raising a concern or dispute?
• Accuracy: Is the information accurate, including where federal consumer financial law requires accuracy?
• Chat Logs: Are chat logs holding personal data treated as sensitive consumer information?
• Language and Dialect Access: Has the system accounted for technology trained on a limited number of dialects and higher rates of failed resolution for consumers with limited English proficiency?