Developing and operating future-proof, trustworthy AI systems – how is that possible? Compound AI systems as a potential solution for sustainable AI integration.
This article was published in August 2024 as a guest contribution in the “Versicherungsforen Leipzig Themendossiers” (PDF available in German).
Generative AI (GenAI) is increasingly being used in both private and professional contexts. The results are often based on Large Language Models (LLMs), which require immense amounts of data for training and rely on an enormously powerful infrastructure.
Insurers are particularly concerned about protecting their (customer) data, not least due to regulatory requirements. The new EU DORA regulation is just another step by the EU to strengthen IT security in the face of increasingly complex cyber-attacks on financial companies. AI applications and the technologies used should be sustainable and future-proof. Most importantly, GenAI should deliver trustworthy results.
This article focuses on how insurers can tackle and solve these challenges with a suitable design of their AI use case.
What is a compound AI system?
Using the results of an LLM involves risks, as the model’s training data can have an unforeseen effect on its answers. This is why Retrieval Augmented Generation (RAG) has become essential. The necessary information is provided as additional context during result generation instead of in the model training. RAG thus reduces the adaptation effort and increases the output quality.
RAG systems are part of a new trend known as Compound AI Systems (CAS). The idea of CAS is based on a modular approach with the aim of using integrated, connected components to solve a specific AI use case. In such a system, each component is responsible for the element for which it is best suited. The focus moves away from optimizing comprehensive models that solve the overall problem (Fig. 1, left) towards LLMs as part of a system to solve the use case (Fig. 1, right). Within CAS, each component solves a sub-problem and thus makes a valuable contribution to realizing an overall solution. The components can be AI models – classic and generative – or other applications that are connected via interfaces and together form the CAS.
Advantages of modularity:
CAS delivers higher quality outputs that can be improved upon easily and in a more targeted manner, as the changes take place within the components.
CAS are dynamic (e.g., basic RAG). Why? They can integrate current data via components.
CAS offer a high level of transparency with regard to creating results. For example, this simplifies quality assurance, where intermediate results of the individual components can be viewed.
But there are some disadvantages: The use case must be broken down into suitable sub-problems, and the corresponding components must be developed and integrated. Each additional component also creates additional interfaces that may increase testing complexity and error potential.
Data protection regulatory requirements
The requirements for data protection and data use depend on the type of data. The strictest regulations apply to particularly sensitive personal data. The most powerful LLMs are often only available via hyper-scalers and not freely accessible to the general public. In order to benefit from this, companies must relinquish their data sovereignty – a challenge for use cases involving personal data.
One solution is the internal operation of local language models. However, this is expensive and time-consuming in your own data center. New approaches make it possible to use smaller, open models for specific use cases. A CAS can help to achieve optimal results while maintaining compliance requirements, as the modularity allows data access in the individual components to be controlled in a targeted manner.
Creating trust
The business model of insurers is the diversification of an individual risk to a larger collective, essentially selling security. Customers have a practical need for trustworthy, error-free insurers.
Making mistakes is in the nature of LLMs - how to deal with them?
A CAS offers various ways to increase the robustness of results, e.g., RAG systems. By combining LLMs with supplementary data sources, CAS can produce outputs that are more specific and therefore more accurate. Additional components, such as automated filters, can reduce the risk of errors. Human feedback should also be integrated as a CAS component.
Future-proofing
The rapid development of LLMs and generative AI makes it difficult to design future-proof systems. To solve this challenge, it is crucial to make the system as flexible as possible through a well-thought-out design.
Hyper-scalers such as Microsoft and Google offer ready-made but closed CAS as complete solutions. This means that the components of these systems are hidden behind the provider’s interfaces (Fig. 2, left). The advantage: quick and easy implementation. However, this increases dependency on the provider and reduces the flexibility to control each component individually.
Alternatively, a separate CAS can be set up (Fig. 2, right). It offers complete flexibility, as each component is selected individually. This approach makes it easier to integrate technological innovations quickly. However, the effort required for initial development and updates is higher than using a ready-made solution.
Application example
A chatbot should be accessible to the policyholder via the homepage of an insurer and able to answer questions about the policyholder’s contracts and the like. A CAS, based on a simple LLM, can increase the quality and security of the chatbot.
In the initial variant, the questions are answered directly by an LLM. The LLM does not have information about the insurance company’s contracts, so the answers are only general. The policyholder’s questions cannot be answered adequately.
Therefore, the system is extended by a component that provides the LLM with this information as context from a vector database. It is possible to decide whether the vector database should be provided as a service or in the company’s own infrastructure.
In the third variant, the “Check” component is used to trigger manual random checks of the chat extracts to service employees for further quality assurance. The results are then used to optimize the system.
The addition of just two components turns an error-prone chatbot lacking specific customer knowledge into a trustworthy assistant that can also reliably answer individual questions.
Conclusion and outlook
A compound AI system offers clients a new opportunity to extend the limits of individual LLMs through the clever integration of additional components. In this way, high data protection and security requirements can also be met.
When implementing a CAS, it is important to adapt the architecture to the use case and take into account the required flexibility of the individual components in order to create a future-proof system.
In addition, the integration of a quality assurance instance (e.g., via the human-in-the-loop component) in a CAS makes it possible to increase confidence in the results and ensure quality performance over the long term.
Let's get talking
Niklas Lümmen
Business Owner AI Informationsfabrik PART OF X1F
Dr. Leschek Homann
Cluster Lead AI Informationsfabrik PART OF X1F
