7 Tips to make chatbot sound smarter when things go wrong
How to help your chatbot handle difficult conversations
I wish all digital assistants could work like J.A.R.V.I.S. from the Marvel Universe. But building an intelligent assistant is not as easy as Tony Stark makes it look.
Working on a Watson AI-powered chatbot for a health insurance client helped me understand the challenges and effort required to build our dream assistant. However, as Experience Designers, we can work with the team to develop solutions to ease people’s frustration when the chatbot cannot answer their questions.
When something goes wrong, the reason behind it varies. So, depending on the scenarios, we design different ways to handle errors accordingly. The goal is to help people get unstuck and find the answers they’re looking for without connecting to a live agent — so that users are willing to come back and ask the chatbot for help again.
Here are some UX tips to help your conversational design sound more intelligent:
1. Randomized prompts
When a chatbot repeatedly responds with the same verbiage, it’s a sign of being too robotic. To humanize this experience, we provide variations for answers to the same questions. This technique applies to most scenarios.
For example, instead of using the stale response, we vary responses when customers are unable to access their claims:
Option 1: “It looks like you don’t have access to claims. How can I help you next?”
Option 2: “I’m having trouble accessing your claims. Would you like to try something else?”
Option 3: “Well, it seems I can’t get you to your claim summary now. How can I help you next?
All responses deliver the same message, but rephrasing makes it sound much more natural than saying the sentence over and over again.
2. Confirm the inquiry
Just as human beings do, the chatbot uses confidence levels to determine if it feels comfortable answering a question. When the confidence is low, asking a follow-up question or simply paraphrasing the question can help validate what customers were saying. The specific threshold of low confidence can vary by product. Adding this extra buffer step not only demonstrates the bot is listening like a human, but it also helps train the data for that specific intent.
For example, in this scenario below, a customer asked about dental benefits but misspelled the word dental.
User: “What’re my dentel coverages?”
Chatbot: “Are you looking for your dental benefits?”
The original inquiry triggers the dental benefits intent with low confidence. However, once the customer confirms the paraphrased question, our chatbot learns about the misspelling. It will be able to return the dental benefits without needing to confirm the misspelling in the future.
3. Offer alternative solutions
I’m lucky to work on a digital assistant deployed on three channels: Web, App (iOS & Android), and SMS. Because of resources and technical limitations, each platform covers a specific set of questions. SMS is the most limited platform.
For example, when a customer asks a question that the SMS channel cannot answer, we suggest the customer visit our website or download the app.
User: “Show me my digital ID”
Chatbot: “I’m still working on sending your ID over SMS, but here is a link to download our award-winning app, and you’ll be able to find your digital ID anytime.”
It’s Important to highlight the benefits when asking customers to switch platforms to find answers. A good reason can help customers understand why and make the change without complaining.
4. Provide helpful context
Rather than saying that we cannot help and transferring customers to a live agent, consider providing relevant information first.
For example, when a customer wants to submit a claim, only live agents can help with this. However, before sending the customer to an agent, the chatbot provides information to help customers prepare for the conversation with the agent first.
User: “I’d like to submit a claim.”
Chatbot: “We can help connect you with the agent to do that. Please have your member ID and provider name ready so that we can work faster for you.”
Notice that I’m also emphasizing the benefits of preparing all the necessary documents because that requires extra effort from the customers.
5. Offer follow-up options
From user research, I learned that many of our customers prefer to call an agent because the agents will promise a follow-up with an estimated time to resolve the inquiry. So, what if we carry over this good practice to our chatbot conversation design?
For example, instead of simply stating that a claim statement is not ready, a good practice is to proactively let customers know that their in-process claims will be available soon and offer a follow-up when it’s ready.
User: “Can you show me the most recent claim.”
Chatbot: “Your claim is still processing and can take up X days. Would you like me to send a notification when it’s ready?”
6. Escalate to a live agent
When you must connect the customers with a live agent, make sure the transferring process is quick and effortless. When calling customer service, it’s common practice to ask the topic you’d like to discuss with the representatives first. This usually happens at the beginning of the call to connect customers to the right specialist and avoid added frustration. However, in a chat experience, when the chatbot doesn’t take this extra step, they are very disappointed. By recognizing the intent from the previous conversation, we can pre-select the topic for the customers to connect them to a live agent who knows how to help them quickly.
7. Set expectations and emphasize that the bot is still learning
This might feel contradictory while discussing how to sound “smart.” However much we would love our chatbot to answer every question, the reality is it takes time for the bot to learn and get there. Setting the right expectations upfront can help avoid unhappy customers after interacting with the chatbot.
For example, if the bot can only handle five questions, let the customers know which five they are upfront. If a customer asks an out-of-scope question, let the customer know that we’re still learning and ask for feedback to show the desire for improvement.
Chatbot: “I’m sorry to hear this. I’m still learning how to answer all your questions. If you can give me feedback, it helps our future interactions be a great experience for you.”
The best design when handling errors is to create an experience where users don’t notice any conversational issues. It’s about de-escalating the issue discretely. The foundation to make this happen is to define the voice and tone of the chatbot first so that the bot sounds more human and conversational. Natural conversation helps establish a connection with users, and when there’s a good connection with them, chatbots can foster empathy and reduce conflict.
We still have a long way to go before building a perfect chatbot. But I hope those tips can help you make your chatbot appear intelligent when dealing with difficult conversations.
Acknowledgements
I’d like to acknowledge and especially thank the Experience Design Team on this project (listed alphabetically): Amy Sharp, Brian Adams, Brittany Bell, Melissa Geissler
Xindeling Pan is a Product Designer & Design Thinking Coach at IBM iX based in Cambridge, MA. The above article is personal and does not necessarily represent IBM’s positions, strategies or opinions.
