Introduction
As the Chief Technology Officer (CTO) of Anycover, I'm thrilled to share my journey of building a chatbot that revolutionized the e-commerce claims handling process. Anycover serves as a solution for Small and Medium-sized Enterprise (SME) retailers, enabling them to seamlessly offer extended warranties with zero upfront costs and liabilities. In essence, it functions as the equivalent of AppleCare for a wide array of products. The integration process is streamlined: a merchant can swiftly incorporate the Anycover plugin, which seamlessly integrates with popular e-commerce platforms such as Shopify, Lazada, Woocommerce, Magento, and Wix.
Once the Anycover plugin is installed on a retailer's website, the platform undergoes verification, and a user-friendly widget is activated, presenting customers with relevant pricing and plan information on the product page. Upon purchasing Anycover protection, customers receive a confirmation email containing comprehensive coverage details and step-by-step instructions on how to initiate a claim. Should the need arise due to product damage or malfunction, users can conveniently navigate to the chatbot interface to file a claim swiftly and efficiently.
The problem
We started our market research by manually going through stores that offer extended warranties. We purchased products and then went through the whole claims process. We also spoke to users who purchase extended warranties to understand the core issue. It was clear to us to keep track of all the purchases, fill out long forms, and call customer care.
The plan
Having gone through these claims process ourselves. We set out to improve the process to make claims more efficient and user-friendly. I was excited to build a solution to digitize and automate the claims processing. We set out to create a new chatbot to file for the claims interactively.
We created a chatbot with a fixed flow. Contrary to LLMs like ChatGPT, flow-based chatbots cannot recognize and chat beyond what it is programmed to. The main reason to go with this approach is to have the ability to integrate into our existing systems. At the time of writing LLMs can be trained on large amounts of text but they can’t query the real-time data from the database.
Evaluating the tools
We checked around a few tools and yellow.ai was on top of our list to build our chatbot. It provided us with tools to easily build, deploy, and iterate flow-based chatbots. But we didn’t go through this approach because
Limited 3rd party support: Yellow AI allows us to interact with the database via API. It lacked the functionality to upload files. We required that feature to evaluate the images of the claims.
Performance & cost: An additional API layer reduces performance and increases in cost as we scale up.
Customization: It only allowed the chatbot to be embedded in a page. We wanted to create a chatbot with a full window view with Anycover branding.
We eventually settled on building our chatbot on AWS Lex. It had some of the issues that were listed above. However, we were able to overcome this by using other AWS services. This required some upfront development effort in the start, but this helped us to build our custom requirements and helped keep the cost down.
The development stack
Frontend: It was a fork from https://github.com/aws-samples/aws-lex-web-ui. It is a UI wrapper for the AWS Lex. The customization was done via vanilla JS, CSS, and HTML. The deployment was done via their Cloudformation template.
Backend: The backend was handled by AWS Lambda, which allowed us to query the database in real time. The code was initially written in TypeScript and then transpiled to JS.
Database: We used Prisma ORM to connect to the RDS Postgres Database for data storage, providing us with ample flexibility and scalability.
CI/CD: Both frontend and backend had their own CI/CD on GitHub Actions. For AWS Lambda, we configured it via the AWS SAM build.
Operations
All the claims that are placed in the chatbot are notified to the team via the Slack channel in real time so that they can address the claims faster. This allows for easy tracking and addressing of customer concerns and also provides valuable data for future improvements and optimizations.
Insights
Faster processing: On average it takes less than 30 seconds for a customer to make the claim.
Customer satisfaction: We contact the users for feedback and our Net Promoter Score (NPS) is above 80.
Usage: More than 70% of the claims were made through the chatbot. The rest preferred direct human interaction.
If you're considering chatbot development or have any inquiries, I invite you to connect with me via the chat feature at the bottom of the page or through the contact form on our website.
留言