Zendesk Is Applying Machine Learning to Move Beyond Customer Service

Zendesk Is Applying Machine Learning to Move Beyond Customer Service

Zendesk, the $2.6 billion company that provides customer service software to over 90,000 businesses has been around for nearly 10 years. In that time they have expanded their family of products to support end customers on a plethora of platforms and devices for both enterprise and small to medium businesses. Now the company is poised to add new machine learning based products and features to help clients impact customer relationships beyond support.

MarTech Series spoke to Sam Boonin, VP Product Strategy to discuss how Zendesk approaches applying machine learning to customer service interactions and what they have learned on their journey to incorporate AI/ML since 2013.

“Support is no longer at the end of the customer relationship, and Zendesk data is becoming increasingly relevant for marketers and sales people.”

Boonin spoke to us about how Zendesk has been working for the three years on applying AI to predict customer satisfaction levels – which can help customer support agents accurately handle tickets and to help consumers get the right information without the need for an agent.

Last year they introduced features like Satisfaction Prediction and Auto Answers and they have another product called Connect although these is still in early access mode as of now. While Satisfaction Prediction is currently available to enterprise customers, Zendesk is ready to make both Auto Answers and Connect generally available in the coming months.


Zendesk clients include many companies in areas like e-commerce, gaming, and other digital services who are sitting on a lot of behavioral intent data that they can use to drive a higher lifetime vale from their customer interactions. The Connect service is more focused on helping marketers by using data gathered from other systems to drive marketing objectives. For example to onboard and nurture customers through a sign-up.

Boonin makes it clear that Zendesk doesn’t plan to imitate companies like SalesForce, in building a data science platform. They are strong believers in using public cloud services built for AI/ML applications from the likes of Google and Microsoft, where they can take advantage of advanced machine learning frameworks like Tensor Flow. This approach, he says, allows Zendesk to build products that are based on using available ML data models, and then training these models towards customer service. Data Science teams at Zendesk work on using these available cloud services and ML capabilities developed by the likes of Google, and customizing the AI towards customer service interactions.


But, training artificial intelligence to really work well in for specific clients is not that simple. The benefit of having 90,000 Zendesk customers with a lot of similarities in the way they interact with customers meant that they could train a data set on the entire corpus of Zendesk data. That gave them the ability to really specialize in use cases in the customer service domain. The big breakthrough, said Boonin, came when they were launching the Satisfaction Prediction service, and they realized the need to build a separate model to train the AI for every client. The work they did on training models showed them that a good AI prediction is different for industries of a different nature. In B2B companies, for example, predicting the likelihood of a good or bad satisfaction level is very different from an on-demand business like Uber.


Figuring out a way to help the thousands of small businesses with relatively smaller amounts of customers and information was another learning curve. However, with the Automatic Answers feature, he believes they have found a way to help all kinds of businesses with a single model. Ticket deflection is a key metric for Zendesk – delivering customer service through automated answers taken from content sitting in a client’s help center.

“This was the big technical challenge which they were able to solve with the development of Auto Answers.” Being able to deliver these full featured ML capabilities that even small businesses can use is an important milestone, and the timing feels Boonin, is perfect in their mission to democratize AI/ML software for businesses of all sizes.

Boonin talked about their experience with seeing marketers effectively using Zendesk and how AI / ML will transform that for the better. Businesses capture a tremendous amount of behavioral intent information, he said, which allows them to reach out proactively and retarget potential customers using Zendesk information about that customer. For marketers, the real challenge today is how to deal with an overwhelming amount of customer information that they can use to target and market to customers.

When marketing teams get buried with data, customer segmentation becomes crucial. Application of ML can help marketers by automatically showing up segments likely to respond well, or likely to churn, and use them for a number of marketing objectives. By understanding performance in terms of customer satisfaction, the Connect product is aimed at giving the ability to segment customers based on actions they take on Zendesk’s customer support tools. Marketers are therefore able to identify and work with customers that really need help and show intent.

So what are the next moves for Zendesk on the path to AI-powered applications? Boonin said he does not think AI will become another category of software but all software products will be dramatically improved by AI. AI is already delivering on the promise of automating low-level tasks, which is a big moment in software. “We are seeing already with chat bots a new class of products that will apply AI as a core technology,” adding that, Zendesk has products for both live chat support and for messaging apps.

He cautions that with conversational AI, the technology is not quite there yet – bots remain a little clunky and un-intuitive, although some uses of bots have already proved to be effective, and he thinks that the technology is improving really fast. Zendesk is working with bot providers but as a customer service company, they focus on the bot to human handover. The handover is critical for clients who use Zendesk to forge relationships which still needs a human touch.

Zendesk founders
Founders Mikel Svane, Morten Primdahl & A. Aghassipour

AI may not yet have had a massive impact on the workforce but Boonin has no doubt that technology will massively impact the workforce in coming years, for example with driverless cars. Whether it takes ten years or two years it will impact – but technology has been impacting jobs forever, he added. “At Zendesk, we are excited about the ability to apply new technologies, and with AI building new features and products is made possible at a much faster rate over time, which is great.”

One of the lessons learned on their machine learning journey was that clients don’t want ML “sprinkled on top”. “Taking the time to channel it into our products is critical as is actually working with companies to integrate it into their products.” Another thing software providers need to do is to help customers through the AI/ML hype cycle – making them understand how it will change the way they interact with customers. And whether your technology is ML powered or not – you need to focus on where you can impact the customer.

In response to a question on the possibility of acquiring AI startups, Boonin said that Zendesk has a pretty robust roadmap for putting ML into our core products, and we can expect to see similar features come up in more products. Zendesk is partnering with many great AI startups and companies he said, naming three key focus areas – conversational bots, applying data science to workflow automation, and natural language processing.

“We are looking at the landscape closely for sure,” he concluded.

With strong Q4 2016 earnings just released, we may just see a lot more news from Zendesk in their plans to move beyond customer service and turn into an AI-powered MarTech giant.

Previous ArticleNext Article

Leave a Reply

Your email address will not be published. Required fields are marked *