In this episode of the O’Reilly Media Podcast, my colleague Shannon Cutt sat down with Josh Zheng and Tom Markiewicz, developer advocates for IBM Watson. They discussed how natural language processing (NLP) APIs, and chatbots in particular, represent just one of the ways AI is augmenting humans and boosting productivity in enterprises today.
In order to apply AI to the enterprise, Zheng and Markiewicz explain, developers first need to understand the importance of sourcing and cleaning an organization’s data, much of which is coming in unstructured formats like email, customer support chats, and PDF documents. This can be “unglamorous” work, but it’s also critical to building a successful NLP app or chatbot.
From there, Zheng and Markiewicz offer some practical tips for developers looking to build chatbots: to have context awareness, to fail gracefully, and to have patience—building a successful chatbot can take time.
Below are some highlights from the discussion:
The hype behind chatbots
Josh Zheng: I think one of the biggest propellers for [chatbots] now is the increase in availability of the NLP capabilities. So, a chatbot uses a couple NLP techniques to make the whole thing work, but these things are actually not very new. They've been around for a while. I think what's different is that they've always been kind of locked up in research labs. There have been open source tools like Python's NLTK that made them more accessible, but it's not until recently, where companies like IBM and Google have put APIs on the cloud and made them very user-friendly and easily accessible, that large enterprises—which are usually more behind on the adoption curve—are able to access them and use them.
Use cases for chatbots in tech and travel
Josh Zheng: Autodesk built a customer support virtual agent on top of IBM Watson. This need came when they first moved from a client-per-software model into more of a SaaS model. They really widened their customer reach, but with that came a lot more customer increase and the need for customer support. ... They were able to build a chatbot [Autodesk Virtual Agent] that is able to answer a lot of the questions. And it turns out, a lot of the questions people have are very similar. ... A lot of these are very simple questions that a machine can take over and let the humans focus on the complex questions or the complex requests. ... They were able to reduce the average time-to-resolution by a huge margin. After implementing the chatbot, we see that on average it takes 1.5 days to resolve questions involving humans and only 5.6 minutes to resolve chatbot-only questions.
Developer-first mentality: Prototyping your way to successful AI apps
Tom Markiewicz: You can try all of the APIs for free and just build little prototypes to see if that fits into what you're trying to do before planning a giant budget and going through the process. That's the beauty of the shift over the last couple of years, with more of a developer-first kind of mentality—the understanding is no longer, ‘Okay, we're going to start a project and we're going to set a big budget, and then we're going to push it from the top down.’ Now, developers are really enabled with the advent of the API economy to go out and find what they want; some of it’s almost Lego-like. ... It’s fast, easy, and cheap (if not free) to build and stand up a project quickly to see if it works. Then, you can take that to whatever approval process you need, but you can build something. There are no barriers anymore to get started.
This post is part of a collaboration between O'Reilly and IBM. See our statement of editorial independence.