Machine learning is not a new concept. Ever since software became a model for building experiences for humans, engineers have been exploring the potential of machine learning. To quickly review, machine learning is the framework of building software that learns and builds upon its own model, ideally through intuitive design and contextual understanding. We are not quite in the futuristic world of human-like AI, but with each experiment, technology is moving toward increasingly independent solutions. At Zendesk, we too have jumped into the beginning phases of exploring this newer world of business-applicable machine learning.
Our most recently released product using machine learning is Answer Bot; a knowledge base solution that automatically serves up article suggestions to customers when they submit a ticket to Zendesk Support. We were recently recognized for this innovation at AIconics 2017 for our use of natural language processing, which is a pillar in our learning model. Customers are already seeing the real time and immediate successes of Answer Bot, but with the product being so new, how is that possible? Let’s dive into this answer for a moment. Answer Bot differentiates itself from basic machine learning by using a deep learning model. But what exactly does that mean?
How we built Answer Bot
By utilizing a global data set built on 10 years of historical ticket data, we’ve built a model that can be instantly successful on any size account. The model is able to learn from nonlinear data, building relevancy-based clusters for itself. We started with the English language. Feeding language context to the model has created an environment where not only does Answer Bot make the connection between exact word matches, but it seeks out relative words and phrases that might be associated.On top of that, relevance feedback from agents will help the model better understand interpretations of context and phrases.
How Zendesk uses Answer Bot
Like so many products we develop, our own Advocacy team implemented Answer Bot in their Support instance. As a complicated B2B support structure, there were important strategies to consider. For example – we need Answer Bot only to fire on English language tickets, only for customers, and only via certain channels. Our knowledge base content has very in depth and lengthy documentation, so it is also important to ask Answer Bot to suggest shorter articles (like ones created via our Knowledge Capture App) that are more to the point of a requester’s question.
Currently, Answer Bot helps out with a small percentage of Zendesk Advocacy tickets. What’s cool is that it’s working, even in complexity, and the nicest thing about implementation is that when we get to satisfactory iteration, we can set it and forget it. For now, our team continues to track the data from Answer Bot reporting to better understand how we can continue to make our documentation and products more accessible to our customers through self service.
Something that we have learned through Answer Bot’s release is that there is no single way to utilize a deep learning, or other, kind of machine learning model in order to fully solve for agent needs. While Dollar Shave Club sees a 25% solve rate with Answer Bot, Zendesk Advocacy will need more and different tools to see similar solve assistance levels. Eventually, we may see agent-assisting models for content suggestion to help boost self service in Guide.
Perhaps there is a way we can help predict the needs of agents by suggesting internal documentation to agents using the Answer Bot model. Or, in the future, ticket routing, macro suggestions and Answer Bot integrations could be utilized to help agents move into a world where repetitive actions and one-touch-tickets become relics, making way for more meaty problem solving and relationship building.
This idea aligns perfectly with Zendesk’s overall ambitions of changing the conversation around customer service’s value in an organisation. Leveraging more self service models that feed into customer-facing content, and making tools that help agents connect with customers in human and time-considerate ways, provide a solid foundational argument for how customer service can be a highly valuable selling point in any company’s playbook.