Regardless of how good a bot’s UX flow is, people sometimes send inputs that the developer hasn’t accounted for. Unexpected behavior can potentially break the flow of a bot’s experience and lead to customer frustration when the bot responds incorrectly or doesn’t respond at all.
The Messenger Platform’s built-in natural language processing (NLP) feature can be an effective way to handle these types of off-script user behaviors. To show you how this can work, let’s take a look at PayPal’s Messenger bot.
The fintech company recently integrated built-in NLP into its Messenger bot to take advantage of Wit.ai. This allows their bot to process and understand questions and commands from their customers as free-form input, including questions about their accounts, password resets, declined payments, issues with purchases, and requesting refunds.
PAYPAL BOT DESIGN APPROACH
PayPal designed their bot to have a conversational flow while combining features from the Messenger/Wit.ai platforms. Their team used Wit.ai to create a conversational bot to interact with their customers and provide help with their accounts in real time.
“We looked at a number of NLP tools before choosing Wit.ai. The fact that Wit.ai was integrated into Messenger was a big plus, but we also found that Wit.ai provided the right level of customization with an easy to use interface that allowed both technical and non-technical team members to directly contribute to the bot’s success.”
— Principal Product Manager, PayPal
In the example below, the user wants to ‘dispute a charge,’ which is recognized by built-in NLP and passed to the bot as an intent. The bot then recognizes that it is not able to handle this request with an automated response and uses the Messenger Platform’s handover protocol to seamlessly transition the conversation to a PayPal customer support agent. This design helps the team manage handoffs and conversation control without needing to store the handoff/control state for each conversation.
USING NATURAL LANGUAGE PROCESSING
For companies in industries like fintech, it can be difficult to handle free-form user inputs, since there is often not a common vernacular, and there can often be a lot of overlap in how certain terms are used. For example, if a customer send the input “payment” are they referring to the money they sent to a friend or a payment toward an account balance? Traditionally the solution for these cases is to ask a series of questions to narrow down the user’s intent. This can create additional effort and frustration for the customer, so PayPal decided to focus on simple questions and answers that provide a lot of value, while recognizing when it’s best to pass the conversation to a human.
When they built the bot, PayPal’s developers integrated built-in NLP and customized their entities via Wit.ai. This eliminated the need for separate API calls to other NLP services and reduced latency. They also used both the Wit console and the POST/sample API to provide examples of questions the bot might receive from customers, and because the Wit console is user-friendly and doesn’t require technology-specific input, they were able to quickly scale their bot. In addition, PayPal is a global company with users interacting with their bot from many different countries, so having an NLP framework that supports many languages was important for them (click here to see the languages currently supported by built-in NLP).
KEY TAKEAWAYS
PayPal’s developer team focused on simplicity and creating a bot that could respond quickly to customer needs. NLP combined with other Messenger platform tools gives their bot the ability to handle varied customer inquiries while minimizing the effort needed by the customer to complete their request. In addition, utilizing the Wit.ai platform allowed the PayPal product, design, customer support, and engineering teams to work together seamlessly to understand common questions and needs from their audience base so their bot would deliver the desired results.
Messenger Developer Blog
