Breakthrough web experience: Migdal's gen AI chatbot is scriptless & evolving from every interaction Qorus Innovation in Insurance Awards 2024

At Migdal, we’re always aiming to improve the customer experience and are continuously exploring new innovative solutions to level-up our service. That's why we embarked on a pilot program – a limited beta test – to introduce a new generative AI chatbot feature on our website.

Migdal's chatbot stands out by prioritizing real-world learning. While competitors may be stuck in an endless cycle of QA testing before launching a gen AI chatbot, Migdal's chatbot is actively learning from real-life interactions. Migdal decided to embrace the potential for initial hiccups in favor of faster development and continuous improvement. This "live learning" approach allows the chatbot to learn faster with each real-world customer interaction.

In simple terms, this is how the generative AI chatbot gathers content and assembles and delivers a response to the questions it receives: 1. The original content is generated by a content manager and added to our public website, migdal.co.il. 2. The crawler runs once a day and indexes all content to the search engine. 3. When a chatbot user asks a question, the search engine finds the most relevant content from the website and summarizes it to provide an answer to the question. The chatbot also provides a link to the relevant digital service to provide additional related and helpful information, for example sharing a link to start a claim.

To understand how this gen AI chatbot comes to life, here's a breakdown of the underlying technology. The gen AI chatbot is a hybrid AI solution consisting of RAG design pattern and an advanced intent model based on the BERT language model. [RAG (Retrieval Augmented Generation) improves factual accuracy in complex tasks by letting large language models access external knowledge sources before generating responses. And BERT (Bidirectional Encoder Representations from Transformers) is a language model that can be fine-tuned for specific natural language processing tasks like question answering and sentiment analysis.]

The chatbot understands the needs and wants of the user, finds the relevant information from our content management system and provides detailed answers to any relevant question. In addition, the bot finds the most relevant action related to the question and directs the customer to the right digital service to fulfill its intent. The chatbot’s functionality hinges on some core technological components that all must work together. To achieve this, we:

1. Developed a crawler solution to get all our public website content and index it to our search engine solution based on AWS Open Search. 2. Used AWS Open Search to find the most relevant content related to the user’s question. We used a sophisticated hybrid search solution based on algorithms such as word embedding, KNN [K-Nearest Neighbors is an algorithm used in machine learning for both classification and regression tasks] and more. 3. Used AWS Bedrock based on Antropic Claude 2.1 LLM (Large Language Model) to summarize the content and answer the question. 4. Used Langchain open source to orchestrate a RAG design pattern (LangChain is an open-source framework that simplifies building applications powered by LLMs). 5. Used BERT language model to create an intent recognition model to deal with small talks and direct the customer to the right digital services when relevant, for example to issue a claim. 6. Used AWS Open Search to aggregate usage analytics, reporting and quality testing.

We also incorporated an automated feedback loop where users can like or dislike answers. This feedback loop improves both the search results based on the feedback and the answers.

There's inherent risk in being the first to adopt new technology, but Migdal recognized the potential for a transformative customer experience, and took a calculated risk by seizing the opportunity to boldly be the first in the country to embrace this technology.