Hybrid knowledge agents in insurance companies

Knowledge agents in action: quick access to valuable knowledge

Insurance companies are seeking to use knowledge agents that provide employees with precise and quick answers to complex questions. To do this, the agents access internal data sources, such as document archives, intranet content and external information – usually via APIs and RAG (retrieval-augmented generation). Language models such as GPT4o from OpenAI serve as the basis for understanding queries and answering them in an understandable way.

However, the use of such knowledge agents also presents challenges, particularly in terms of data protection.

Data protection as a priority: protecting sensitive information

For insurers, the protection of sensitive information is essential. Information must not leave the German or European area uncontrolled and in certain cases, disclosure outside the company network is strictly prohibited. So how can a reliable knowledge solution be developed that meets these requirements?

One option is to use open-source language models in an on-premise private cloud. This enables a high level of control over data flows, but requires specialised expertise in operating and maintaining these models. Furthermore, open-source solutions lag behind commercial solutions when it comes to connecting to complex data sources.

The hybrid approach: combining data protection and efficiency

A sensible approach for insurers is a hybrid solution: a high-performance external model such as GPT4o takes over the initial question processing, and anonymised data transfer takes place in this step, while sensitive information is processed internally only. An open-source model such as Llama can be used here, which delivers high-quality answers.

This hybrid approach makes it possible to combine performance and data protection: insurance companies can meet regulatory requirements while taking advantage of the power of modern language models.

Technical implementation of the hybrid knowledge system

The hybrid knowledge system in insurance companies is based on a multi-agent system (MAS) that distributes the individual steps of question processing to specialised agents. This structure increases flexibility when processing complex queries and enables secure information processing (see figure below: multi-agent system (MAS)).

In the following, we will explain an MAS using the example of an SQL agent. The SQL agent makes it possible to chat or talk to a database without any knowledge of SQL (VoiceBot variant).

The advantage of such a solution, which is based on a language model, is that the user can have a dialogue with the database in which the questions and answers build on each other.

• 1) Incoming requests: The user request is fed into the system via a proxy interface to ensure smooth interaction between user and system.
• 2) Data engineer: The data engineer agent analyses the request and develops a query plan based on the database structure to retrieve the relevant data in a targeted manner.
• 3) SQL Executer: The SQL Executer agent formulates and executes the database query to obtain the necessary information from various sources.
• 4) Answer Writer (Llama): Finally, the Answer Writer takes over the interpretation of the query results. It processes and formats the data into a response that the user can understand.

Data protection through local data management

A central function of the system is the local administration of data output. Table descriptions and result data output as results are only saved locally and are not transferred to external systems. This ensures that no sensitive information leaves the company network. This local administration solves the problem of uncontrolled data transfer and thus guarantees data protection.

Future prospects: Optimisations and new functions

In the further development of hybrid knowledge agents, additional improvements could be implemented to increase efficiency and user-friendliness:

Automated database descriptions:

Systems for the automatic provision of database descriptions could save employees time by creating structured descriptions for frequently used databases and providing exemplary values.

Advisory agents for large databases:

For very large databases, the use of specialised advisory agents that filter out targeted information from thousands of tables and make it available to employees could be helpful.

Conclusion: the hybrid approach as a future-proof solution

By using hybrid knowledge agents, insurance companies can not only meet regulatory requirements, but also significantly increase their efficiency in knowledge management. The combination of external and internal language models enables high performance while adhering to strict data protection requirements. This results in a future-proof solution that efficiently handles complex tasks while ensuring the protection of sensitive data.

Also interesting:

• Large Language Models (LLM) in the insurance industry
• A brief introduction to automatic document processing with Large Language Models