Zendesk has developed Answer Bot - a chat bot that can deflect email support requests. And while they're not the first to do this, they are the only company I've found that has been developed a bot Ike this locally that can work at the scale Zendesk delivers. I spoke with Zendesk's Brett Adam, the managing director ANZ and VP of engineering for APAC and data scientist Chris Hausler.
Adam says the company wanted to take advantage of machine learning technologies to enhance Zendesk's product.
"We didn't want to build a simple pattern-match, response bot. We wanted to base it on the really rich content we have," said Adam. "It just so happened we started building some capacity around data in Melbourne - our data team - that's why we started in Melbourne.
That content comes from the billions of messages sent to support functions operating on Zendesk's cloud platform across their customer base of almost 110,000 global customers. By leveraging that data, they are able to use a massive repository of questions and answers to train their algorithms.
And while the data from the entire customer base is used to train the bot, data is kept isolated. So, if a user from Company A sends a query to their support desk, they won't see potential solutions from Company B.
Hausler says one of the reasons the team is in Melbourne is because it's easier to find staff here than in San Fransisco.
"There's a great talent pool here. We've been able to get really good people," said Hausler.
That talent pool had already worked on other machine learning products so it made sense for Zendesk to build on that capability.
Interestingly, Hausler said the market perception that there is a skills and talent deficit in the areas of data science and machine learning works in his favour. Because people think there aren't enough potential staff in the market, they aren't looking. That has created an opening in the market for Zendesk to recruit strongly. With over 160 employees already working in Melbourne, they plan to double that by 2020.
Adam is also on the Victorian goverment's Innovation Panel. He noted that while there are some tech fields where there is a pool of talent, there are other areas, particularly when it comes to skilled engineers with management experience that are much harder to recruit for.
The main use-case for the bot, said Hausler is question/answer or ticket deflection. When a user office customer sends an email to support, the bot carries out a semantic analysis of the message and identifies what the query is really about. It then marches that question with potential answers and delivers a potential solution to the user.
Customers involved in testing, such as the Dollar Shave Club in the US, have been able to successfully answer over 10% of inbound queries this way. If a user receives an email from Answer Bot that fixes their problem, they simply mark the ticket as closed.
Hausler says email is often difficult to intercept and traditionally has required support and call centre staff to manually intervene and suggest ways for the user to self-solve. The main application is to handle repetitive questions that have a known solution.
That loop is important as it helps quantify the benefits of the bot. Adam said the bot service is charged to Zendesk's clients on the basis of closed calls. With the cost of closing a reasonably simple email query running around $4, Adam says Zendesk charges their customers $1 for each support call that is successfully resolved by Answer Bot.
As well as helping to solve user queries faster and help users be more independent, Adam says support and call centre associates are freed from boring, tedious tasks so they can focus on more complex and interesting problems. He beehives this will assist with improving employee satisfaction and retention.
The first focus for Answer Bot is on email but there are plans to expand it to other communication channels such as IM and web-based chat and forms.
Hausler said email posed some interesting challenges. Unlike searching, where users enter a query based on specific keywords, email has its own semantic structures such as greetings and can use descriptive language to describe a problem. For example, a user might have, in the past, been directed to a knowledgebase and told to search for "reset password". But Answer Bot has been trained to understand terms such as "locked out of my account" or "I've been on holidays and my account needs to be reset'.
Answer Bot's machine learning algorithms have been trained by looking at the billions of emails received across Zendesk's entire customer base, to understand that those queries refer to resetting a network password. It has even been trained to deal with emojis and other informal language.
"We've built a system that gets to the semantic intent of the question, not just the patterns if the words," said Adam. "It's not brute-force pattern recognition - it's semantic magic".
It's the data that Zendesk stores, across their own data centres as well as Google's Cloud and AWS, that was a strategic asset they could build on rather than acquiring or licensing some other machine learning system. It also means the system is effectively being trained with a massive set of data rather than relying on lots of professional services to tune the system.
Adam says this means Answer Bot is ready to use for all Zendesk customers.