What we learned from a chatbot failure

As many as 130 million out of India’s 153 million social media users are on mobile, and they’re spending the majority of their time on messaging apps. By leveraging messaging apps, businesses would not only be better able to offer superior customer experience and service, but they would also be exposed to a large pool of messaging platform users. Clearly, conversational commerce is the future.

Thus, we decided to build MindIQ, a bot builder platform. The aim was simple: help businesses create chatbots as easily as possible without requiring programming/coding knowledge or A.I. jargon.

After reading numerous research papers, learning about bots through online tutorials, and toiling away, we ended up creating a bot builder platform that help developers create chatbots with a conversational tone. After enough user interactions, the ecosystem would be tailored according to the user.

But just as we thought we were ready, we received our first blow. One of the first users asked the bot this question:

I weigh 120 kg, my wife weighs 80 kg, we’re travelling to hilly terrain, which bike do you recommend?

Our bot couldn’t answer this question. And as far as we know, none of the powerful artificial intelligence machines could. Also, we discovered that users had to type a lot of words to get an answer. This was definitely against our aim of helping businesses deliver superior customer experience. We therefore had two options. Either we had to educate the users on how the query according to the parameters of the bot or introduce buttons and quick replies, thereby giving options to the user. We naturally went for the second one.

With buttons, we were able to kill two birds with one stone. First, we reduced the amount of words the user had to type to get an answer. Second, the chatbot was able to drive the conversation to its final destination, recommending to the user a product or service catered to their needs.

Above: Use of buttons to deliver superior customer experience.

FreshMenu, one of India’s most popular online food delivery companies, was one of our platform’s early adopters. When you search for FreshMenu on Facebook Messenger, you get a “Hello, I’m FreshMenu bot…” message, along with the day’s menu. We were thrilled to see it working as we imagined it to be. The bot interacted seamlessly with users and was able to process orders.

After the success with FreshMenu, we decided to up the ante. We built CBPredictor, an algorithm that can predict the next user action and provide recommendations for the same. For example, our Food Ordering Bot can cue the user to track his order once it has been placed, because that’s the most likely action to follow.

Above: The bot predicts the next step user is likely to take.

How do we predict these steps? We generate a graph of all possible paths the user can take in the user journey, with the root of the graph being their first interaction with the bot. As more users go through the process, we start weighting the paths of the graph. The top-weighted paths are then returned as quick replies, with buttons on Messenger along with the response. We also give preferences to the paths the business wants the user to go through; for example, “add to cart” is preferred over “show details” and thus is shown as the first button on a response.

Through our journey, we have understood that the best way to build bots for businesses is through a hybrid approach — use of buttons and quick replies along with text-based queries. This approach helps both businesses and customers. For the business, the chatbot will drive the conversation to its final destination: the user purchasing the product or service. For customers, the amount of words the user had to type to get an answer has reduced, giving an overall superior experience.