The era of AI-Agents is expected to arrive in the 2030s. AI-Agents are envisaged both for individuals and for companies. Agents for individuals will be referred to as ‘personal agents’ and those for companies as ‘corporate agents’. In multi-agent systems, agents are expected to communicate with each other and negotiate, buy and sell on behalf of humans. Corporate agents are expected to act standing on the side of the company, while personal agents are expected to act on the side of the individual (consumer). If corporate agents act without taking into account the company’s strategy and policies, it will have disastrous consequences. (The same applies to personal agents.)
The industry is currently developing RAG (Retrieval-Augmented Generation) systems. This is clearly a prelude to the above. We therefore discuss how to build in a corporate strategy into the RAG system. Traditionally, corporate strategy has been the subject of business administration. It’s a world where MBAs play on Peter Drucker’s words and use cross-charts. However, the strategies they produce are often only written in thick reports and not implemented. In contrast, we have demonstrated over the past two decades that strategies can be represented using Self-Organising Maps (SOMs) and decisions can be implemented on a computer.
Let’s focus here on the interaction between companies and their customers.
1. Customer Segmentation Model
In order to optimise customer interaction, it is assumed that a customer segmentation model has first been created using data science methods. We strongly recommend SOM here. (However, the SOM must be correctly implemented and many SOMs published in open-source libraries are not recommended.) Based on historical transaction data, each customer is segmented. Analysing the statistical characteristics of each customer segment (profile analysis) and defining in detail what treatment is given to which segments. This is the basis for everything that follows.
2. VoC Analysis
Before describing the RAG system, it is necessary to analyze the content of customer contacts that the RAG system processes. Using LLM, we convert customer contacts into text, split them into chunks, and get embedding vectors, which we then segment with SOM, allowing us to analyze how many and what types of inquiries, questions, requests, and complaints customers have. This information is integrated with the customer segmentation mentioned above.
3. Extension for Strategic RAG System
The purpose of the RAG system, which is currently being discussed in the world, is to enable chatbots using LLM to have more effective conversations with customers by taking into account the unique information held by the company, as the information learned by LLM alone is insufficient. It is expected that it may be possible to some extent to include corporate strategies and policies in the prompts given to LLMs, but there are limitations to this and, primarily, it is expected to be difficult to maintain. Therefore, we propose to add an extension system that makes strategic decisions based on the segment to which the customer belongs and the content of the conversation with the customer.
We are also considering using Bayesian belief networks to accurately judge the context of conversations with customers based on limited information. Typical RAG systems assume that users (customers) ask questions and the system answers them, but it is difficult for people to ask questions about things they do not know in the first place. By using BBNs (as already achieved in applications such as troubleshooting), the system can ask the customer questions and guide the customer to the goal the company wants to achieve.
We provide the software tools and Python code to achieve this.