Leobit
AI-Powered Knowledge Base Chatbot
Custom AI development for an insurtech company to reduce support workload
ABOUT the project
- Client:
- Insurance and financial services holding company
- Location:
-
Bermuda
- Company Size:
- 250+ Employees
- Industry:
- InsurTech
- Solution:
- Custom Software
Leobit’s team is building an AI-powered chatbot for an insurtech company to improve its customer support workflows. The chatbot provides quick answers about the company’s services and policies, using content from its SharePoint environment and public website. Each response includes source attribution, so users can verify the information themselves. The solution is being developed in phased sprints, starting with a grounded discovery phase followed by chatbot development and optimization.
We were really satisfied with Leobit’s depth of expertise and strong business analysis. Their delivery model is very convenient, as they drive progress and provide updates consistently.
Customer
Our client is a Bermuda-based insurtech company specializing primarily in health insurance, while also covering other areas, from vehicle and real estate insurance to fund administration services. As a notable player in the Caribbean region, the company operates a high-traffic website that places significant demands on its support teams.
Business Challenge
The customer needed to reduce the load on its support team. They required a solution that would improve customer communication through an AI agent capable of providing fluent, well-grounded responses. The agent also needed to be connected to a company-specific knowledge base. In addition, the client wanted visibility into the content the chatbot was indexing.
Why Leobit
Leobit’s team is building an AI-powered chatbot for an insurtech company to improve its customer support workflows. The chatbot provides quick answers about the company’s services and policies, using content from its SharePoint environment and public website. Each response includes source attribution, so users can verify the information themselves. The solution is being developed in phased sprints, starting with a grounded discovery phase followed by chatbot development and optimization.
Project
in detail
The project is ongoing, as it has reached its testing and stabilization stage. Before that, we ran a discovery and developed the solution’s core and AI capabilities in several sprints.
We started a project with a discovery phase during which our business analyst and technology specialists analyzed all the customer’s requirements and came up with an implementation roadmap. Particularly, we outlined the key knowledge sources on which the solution would base its responses, namely SharePoint content and the client’s website. The team also planned an admin tool for indexing visibility, source attribution in chat replies, and a roadmap toward deeper business-process integrations, such as CRM escalation.
We started by building the chatbot’s core with a .NET back end and Preact 10 for the front end. Preact was chosen for its lightweight architecture and compatibility with isolated CSS, which helped us keep the chatbot UI self-contained. Our specialists also built an admin panel with React.js and Tailwind CSS for managing data sources and reviewing basic index analytics.
The team built a SharePoint content indexing pipeline using Azure Blob Storage to store unindexed content and Azure AI Search to make that content searchable. Website content is kept current through an Abot crawler, which powers a separate crawling and indexing pipeline so the chatbot’s knowledge base stays up to date with both SharePoint and public web content.
For the model layer, the team used Azure OpenAI to set up three separate deployments: one for chat, one for general-purpose utility tasks, and one for embeddings. At this stage, we did some custom configurations using the OpenAI SDK to ensure effective connectivity with the data sources.
Source-Grounded Chatbot Responses
The chatbot’s answers are tied directly to indexed SharePoint and website content rather than the model’s general training data. This reduces the risk of misleading answers that generic chatbots are prone to, particularly on edge-case queries or in areas where content coverage is thin. The tool also has a source attribution functionality. As a result, every chat response displays where its answer came from, giving users a way to verify accuracy.
Admin Visibility and Control
An admin tool shows indexing status across both SharePoint and website sources, giving the client’s team direct insight into what content is powering the chatbot and where indexing gaps might exist. The admin panel also includes knowledge base management features. The admin can remove and filter sources, keeping the chatbot’s responses relevant and accurate.
Quality AI Self-Verification Workflow
The tool uses the GPT-5.1 model for core chatbot functionality and source attribution, while a GPT-4.1 judge model verifies responses against pre-set company configurations. This test suite generates two response versions, each from a different model, compares them, and uses the comparison results to refine the final output. This feature sets an additional validation layer that ensures higher quality of chatbot outputs.
The Journey
Behind Our Success
Technology Solutions
- Reliable data pipeline supporting the chatbot, including Azure AI Search for vector searching of information, Azure Blob Storage for storing data that awaits indexation, which retrieves information from the corporate SharePoint and
- Convenient admin panel for the chatbot built with React.js and Tailwind CSS
- Chatbot response test suite involving a GPT-4.1 judge model that verifies chatbot responses according to corporate rules
Value Delivered
- Complex AI integration project handled in around 3 months
- Complete ownership of the project, from project discovery to release
- Additional technology inputs aimed at improving the quality of chatbot responses
- Source-grounded AI chatbot that can reduce the workload of the customer support team by up to 40%
