GPT-3 making life easier for customer service agents
ActiveChat is a smart conversational AI platform with natural language understanding for customer support, e-commerce and marketing chatbots.
Activechat has partnered with OpenAI team to build something that looks like real magic. Once you describe the context of your business in a plain natural language and provide a couple of example questions and answers that your human live chat agents may encounter in their day-to-day conversations with customers, we can use the power of GPT-3 to provide instant hints to your agents to help them answer almost any question that a customer may have.
Let’s look at a real-life example:
GPT-3 as a bike shop assistant
These settings are used as default in Activechat’s GPT-3 integration, and the context for this use case is pretty straightforward:
“This is a conversation between a website visitor and a smart virtual assistant. The conversations are happening across various pages of the website that is selling bicycles, and virtual assistant can answer all kinds of questions about various bike brands and models and help choose the right one according to the description made by the visitor.”
To seed the AI with some question examples, we used the following:
Q: What’s the difference between a carbon and regular bike? A: The primary difference between carbon and aluminum comes down to weight and ride quality. Carbon frames are usually a bit lighter than aluminum — up to a pound for mountain frames and up to a half-pound on road frames.
Q: How much does a good bike cost? A: Road bikes range between $350 and $700, Mountain bikes around $1000, Single-speed bike — $400, Beach cruiser $200–300, Recumbent bike $1000–2000, and Kids’ bike (3–8 yrs) will be $140–200.
Q: How do I service my bike? A: You should regularly service your bike to ensure it is running efficiently and that there are no worn or damaged components. The more you ride the bike the more frequent you should be servicing it, especially after riding in dirt, sand, mud and in the rain.
That’s all it takes to start getting valuable and relevant answer prompts to various kinds of questions that visitors of a bike shop may ask! Here is a real-life example of the conversation completely built from prompts offered by GPT-3, with zero editing by an agent. The only action taken by the agent was choosing which one of the three prompts to use as a response.
Clicking any of these hints will copy the answer to the message window, and your agents can send it immediately or edit it, adding specific links to your product pages or other resources. Based on our research, this can reduce the time needed to find and type the answer by 70-80%!
If you’re not satisfied with what you see, just hit the “Refresh” icon and a new set of hints will appear. These hints are powered by all knowledge that the Internet has (GPT-3 model is trained on Common Crawl data, Wikipedia, and multiple other data sources, including a huge amount of books). This makes suggested answers sound human-like and in most cases, they will be relevant, useful, and valuable to your customers.