How to build a chatbot… that fixes itself
Humans misspeak all the time. Even though we think we are stating our desires and questions clearly, we often don’t form sentences well, leading others to ask, “Did you actually mean…?” It takes listening skills and intelligence to keep even basic conversations on track.
Now imagine customer care chatbots doing the same thing. In most cases, they can’t. They simply don’t have enough intelligence to detect that the conversation is going off track the way humans would. So they just repeat an answer cheerfully — angering the customer and leading to the spiral of misery and frustration with customer service.
Building on its market-leading intent detection, Watson Assistant now has the smarts to detect conversational pitfalls. With the new Suggestions feature, it can now detect that it’s not answering a customer satisfactorily, and it takes a different approach. It suggests relevant conversation topics in slightly different contexts to try to get the end user unstuck.
Let’s look at an example of how Suggestions would work in a common conversation.
Marissa, our hypothetical end user, wants to pay her bill. She phrases her request as:
- “How can I pay?”
The assistant returns a relevant answer that answers the question:
- “Payments can be made with a Visa, Mastercard, Amex, Paypal, or at the post office.”
However, Marissa doesn’t want to just learn about paying bills, she wants to actually pay them! She then asks the question again:
- “How can I pay????”
At this point, the assistant recognizes that Marissa has a different goal than what she asked. Because the assistant has now repeated itself, it knows that it’s likely not giving Marissa the answer she wants. The assistant now opens Suggestions automatically and surfaces this message:
- “Here are some other options:
- CVV info
- Pay bills
- Browser support
- Connect to an agent”
Marissa notices that Suggestions has given her relevant alternative responses, and she’s able to click “Pay bills” and enter the right flow — preventing her from asking for an agent or giving up on the interaction altogether. The Suggestion feature correctly detected the end user was stuck and successfully offered an alternative route to a successful interaction.
Let’s continue the example with Marissa.
After paying her bill, she switches topics and asks:
- “Can I order foreign currency online”
In this case, the assistant was not configured to respond to this message. Instead of returning an unhelpful “can’t help you with that” error, end-users can now use Suggestions to view relevant search results (if you are using Search skill). In this example, Marissa is able to find out the answer to her question by using Suggestions with Search.
Marissa can also use Suggestions to exit the current topic whenever she chooses and jump to alternative popular topics. After inquiring about foreign currency, she glances at the “People are also interested in” section and chooses to check her account balance. She’s able to log in and view her balance, completing this flow.
She then changes topics again and asks if she needs to inform the bank when traveling overseas. The assistant surfaces a relevant search result that tells her she does need to inform the bank. At this point, Marissa can again use Suggestions to connect with a human agent. The assistant then seamlessly hands off to a human agent to provide further support. If you don’t have an agent-handoff feature implemented, you can instead configure the “connection to support” button to surface a helpline phone number, customer support email address, ticket submission form, etc.
With Suggestions and all of the other AI intelligence of Watson Assistant, end-users avoid broken conversations. They receive intelligent alternatives, relevant search results, easy access to popular topics, as well as seamless connections to human agents. No dead ends!
Sign up for Watson Assistant to start building, and if you haven’t already signed up, join the IBM Watson Apps Community to continue the discussion with our experts and other Watson Assistant users!
