Everyone is into AI now.
A year after the ChatGPT release:
- many articles popped up
- new companies appeared
- lots of ideas emerged
But it is impossible to find articles with:
- practical information
- real examples
- obstacles identified
- and how to overcome them
We built our first support chatbot using GPT.
We tested it on actual e-shop data.
We faced many challenges.
And here’s how we solved them:
1. Identify Your Use-case
ChatGPT can provide many benefits to your e-commerce business:
- Idea creation for marketing
- Copy-write improvements
- Reviews analysis
- Support template generation
Yet, the most impactful for your e-shop is to use ChatGPT for:
- Customer service automation
- Product recommendations
Here is a chatbot that we were dreaming of:
- Answers common questions
- Answers questions that have an answer on the websites
- Answers product questions
- Recommend products
2. Understand AI Chatbot Flaws
There are good chatbots and bad chatbots. The bottom line for automation is to improve customer experience. If automation lowers CX, you should rethink your concept.
We recently wrote an article about the most common chatbot mistakes. Keep them in mind. But beware of the new challenges ChatGPT brings:
- How to add product data to the GPT engine
- How to keep them up-to-date
- GPT recommends competitors
- It provides wrong links
- It leads customers out of the e-shop
- Copy-write doesn't relate to your brand
- How to answer non-complete questions
- Misleading or wrong information - GPT hallucination
- No connection to a live agent
- GDPR compliance
3. Overcome AI Challenges
ChatGPT is one of many large language model (LLM) chatbots on the market. But, it was the first publicly offered, with a large user base, ongoing development, and great results.
Microsoft offers a GDPR-compliant ChatGPT variant through their Azure ecosystem. We decided to integrate the Azure ChatGPT version. It connects well with Azure Cognitive Search which will serve as an internal knowledge base.
Your chatbot should know your products but also information on your website available to your live agents.
It should combine the general knowledge…
… with your business-specific knowledge. Consider to provide following data to GPT engine:
- Product data - excel, pdfs, product description, etc.
- XML Feeds - Google Merchants, Amazon, etc.
- Historical support questions & answers
- Answer templates
- Live agent guides & how-tos
You can consider scraping tools like Apify to scrape the whole e-shop. Use some computer-readable format (csv, JSON, txt) and include website URLs.
Scraping is good for initial data collection. But not suited for regular updates - takes a long time, and loads your servers.
To keep the data up-to-date, use product feeds (XML feeds). It is an existing technology you use to share product information with search engines, price comparators, etc.
The remaining challenges mentioned above need to be solved with smart prompt design. Your GPT prompt that goes before the customer’s question should address:
- Do not recommend competitors
- Do not lead customers outside of your e-shop
- Define copy-write and branding GPT should follow
- Prefer internal knowledge base over general knowledge
- Ask for more information when a question is unclear
- Provide accurate information
A natural transition to life agent is triggered by negative feedback from customers. Collect customers’ feedback after the GPT answer.
ChatGPT is a powerful AI tool that is already changing how e-commerce companies operate. By understanding human language and leveraging large data sets, it is capable of automating customer service and tailoring product recommendations.
If you want to have GPT but don’t want to go into technical details, schedule a demo with us. We will show you how you can use our AI conversational designer with GPT integration to transform your customer service.