Chatbots suck! I can hear the screams of customer service chatbots’ victims in my dreams. Being the CEO of a startup automating coaching, I had doubts about chatbots myself. I am still surprised that there are users who report getting value from working with our Chatbot Coach.
Yet, the positive evidence is increasing each day that we are serving our clients. I would like to share four secrets that contributed to our preliminary success in building a conversational chatbot.
1. Evergreen content
First of all, the nature of the conversations you are automating is a big factor. Unlike mentoring, which focuses on providing advice, coaching is about helping people to set goals, asking insightful questions, and providing the space for reflection. Questions are easier to automate than advice because questions are more likely to be relevant. Here is an example:
A) To succeed, you should focus on working smart rather than working hard. B) How are you going to measure your success?
A question is typically relevant to a wider audience than advice. Therefore, when building your own conversational chatbot, consider if you can provide value by asking questions. It would be easier to build! As a side note, here; is a short, beautiful comic about the value of questions.
2. NLP to cluster challenges
The core of our innovation is using AI to understand the problem the user is facing, not to generate content or provide the solution to the user. NLP (natural language processing) is advanced enough to take the user input and categorize it.
In our case, we would ask a simple question: “What is the challenge you would like to work on next time?” If the user says “I am running from meeting to meeting and feeling very stressed,” the bot could classify it as a “prioritization/time management” challenge.
3. Modular design
Next, we would match the user challenge to the content. In our case that would be “Prioritization frameworks” and “Project-and-time-management systems”. All the content is created by ICF (International Coach Federation) certified coaches and arranged into a content library. Here is an overview of all the categories we defined: A content library with well-defined categories is essential for training a successful NLP model and AI recommendation engine.
4. GPT-3 to vary the filler content phrasing
Finally, there is one case in which AI is great for content generation: filler content. In our case, we check with the user on their progress with their goals in every session. Currently, every time they achieve their goal the bot says “Wow, that’s quite something!” You can imagine that it starts sounding quite repetitive after a while.
This is where we plan to use GPT-3; to vary the bot’s responses. GPT-3 is developed by OpenAI and uses deep learning to produce human-like text. The goal is for the bot to say the same “Good job!” point in 20 different ways, just like a human would. Not only does this make the bot more humanlike, but it also helps increase the retention rate of users. Down the road, we could also create variations on the expert content.
Conclusion
There are ways to make conversational chatbots suck less, and perhaps even to provide value at scale. I hope that our learnings inspire you to give your ideas a shot, too. If you want to find out more about our Chatbot Coach and try it out, you can find here. I am also happy to answer questions at dima.syrotkin@pandatron.ai. image credit: Domingo Alvarez E on Unsplash