Developing a chatbot is on more digital organizations' to-do lists than ever before, but it isn't an easy task, says Lauren Kunze, CEO of Pandorabots, a platform for building and deploying chatbots. Too much disorganized data, too little data and legacy systems can all prevent organizations from realizing their chatbot strategy goals, Kunze told SearchCIO at the recent Chatbots & Virtual Assistants for the Enterprise event in San Francisco. Here, she details the challenges involved in developing a chatbot and explains why humans still need to be kept in the loop as the technology advances.
What are the biggest roadblocks IT leaders must overcome when developing a chatbot?
Lauren Kunze: There are actually a lot of different cases and it depends on the vertical. Sometimes a company has no data at all, and then it's really more of a creative project or trying to guess, to some extent, how users want to interact with the service -- and most of the time you are going to be bad at guessing. It's very important for development cycles to be one to three months, and then you put the bot out there, start collecting data on how people actually want to interact with the system and iterate from there.
On the other side of the coin, we have companies that have too much data -- they're drowning in data and they don't know how to make sense of it. And [that data] comes from all different sectors and channels. They'll have forums, they'll have live chat, they'll have email, they'll have call center logs, and they would love to leverage that data, particularly in an FAQ-type of customer service case. And there's a big misconception in the industry -- although I think people are becoming increasingly more educated on this -- that if we have all this data, we can simply put it in a blender and push a machine learning button and it will spit out this perfectly conversational chatbot. That's not the case. That data has to be pre-processed and tagged in order to be useful. Even then, machine learning techniques can only get you some of the way there. You still very much need humans in the loop to train the chatbot or be escalated to if the chatbot fails in the customer service case.
The other big roadblock when developing a chatbot is your existing infrastructure and legacy systems. One presenter spoke today about some of their documentation being on microfiche and developers not knowing how to update that. Another example might be pre-ordering at a restaurant. It seems like a great use case for a fast food restaurant to allow people to order ahead of time and skip the line, but in some cases companies just don't have the infrastructure to support fulfillment and services like that. Or they do, but it's very old and out of date. That's another challenge: How do we actually get the data that we do have into a format that we can use?