An AI assistant in banking customer service manages with surprisingly little customer data. It answers most questions using public product information, terms and conditions, and user guides. It only requires customer-specific data, such as the status of an application, when the customer is logged in and the assistant is integrated with the relevant system. It never sees access credentials: no PIN, no TAN, no password, no full card number. Anyone who categorizes data into these three classes has already made the most crucial decisions for data protection and information security.

Three Data Classes

Freely usable. Product information, terms and conditions, fees, user guides for online banking and the app, opening hours, forms. This content is public or the same for all customers. It forms the knowledge base and supports the majority of conversations.

Only after login and via an integration. Application status, personal product portfolio, branch appointments. This data belongs to an individual. The assistant is only permitted to display it if the person is securely authenticated, retrieving it from the primary system at the moment the question is asked.

Never in chat. PIN, TAN, passwords, security questions, full card numbers. No customer should enter this data into a chat, and no assistant asks for it. If a customer asks whether they should enter it anyway, the correct response is a clear "no," accompanied by guidance on the secure path.

Assisting with Access Without Seeing Credentials

A large portion of banking customer service queries revolves around access: first-time login to online banking, setting up TAN procedures, locked accounts. An assistant can resolve these queries without ever seeing access credentials. It explains the steps, directs users to the right place in the app, and hands off to a human agent if an account unlock is required.

This is precisely where it becomes evident whether an assistant is bound to verified content. A language model without authoritative sources might helpfully ask for a TAN because it recognizes the pattern from other conversations. An assistant that responds exclusively from approved guides strictly follows the bank's established procedures.

What Happens to Conversation Data

Even when the assistant only uses freely available content, customers still share personal information in the chat. Four key questions apply to this conversation data:

1. How long are chat histories stored? The retention period is defined by the bank, ranging from deletion immediately after the conversation to a fixed retention timeframe for auditability.

2. Who has access to them? Role-based access controls determine which employees can view chat histories, while an audit log records every access event.

3. Where do they go? If requests are sent to an external AI provider, the data leaves the bank's sphere of control.

4. What are they used for? If conversations are used to train external models, they deviate from their intended purpose.

For banks, regulatory requirements for outsourced IT services also apply, which the vendor should support through comprehensive documentation and audit rights. To see what documentation a vendor should provide, consult the compliance checklist for chatbot vendors.

How Mercury.ai Implements the Data Classes

Mercury.ai generates responses from approved sources such as product information, terms and conditions, and user guides. The assistant retrieves customer-specific data via integrations during the live conversation once the customer is logged into the portal. Knowledge content can be filtered by permissions so that it only appears under specified conditions. The bank configures its own deletion periods for conversation histories, with roles and a comprehensive audit log governing access.

Mercury.ai hosts and operates the language models itself, based on licensed Mistral models. Requests are never routed to external AI providers, conversations are not used for model training, and all processing takes place exclusively in Frankfurt. When an issue requires a human decision, the assistant hands off the interaction via the Agent Desk along with the full conversation history. To see how this has been deployed at Volkswagen Bank since 2018, read the Case Study, or explore typical use cases in our article on the AI chatbot for banks.

Looking to define data classes for your banking customer service? Get in touch with us.