Zendesk Enables Faster Customer Service Using Deep Learning on AWS

Zendesk is a customer-service platform designed for companies that want to create customer relationships that are more meaningful, personal, and productive. Headquartered in San Francisco, California, Zendesk helps companies provide great support and then mature with self-service and proactive engagement.

Building New Deep-Learning Customer Applications

Zendesk, which provides software-as-a-service (SaaS) customer-support platforms, is always working to create new and better solutions for its customers. Zendesk needed to respond to a growing trend: customers wanting to quickly find answers to questions on their own, without having to talk to a support agent. “We wanted to give customers more relevant answers as fast as possible, and we wanted to drive a self-service customer support model,” says Soon-Ee Cheah, a data scientist at Zendesk. Companies such as online retailers and other large enterprises use Zendesk to deliver great customer support.

Zendesk met this challenge by using deep learning—an increasingly popular branch of artificial intelligence (AI). Deep-learning frameworks use neural networks modeled on the human brain to enable computers to learn independently, based on the data they are fed, and perform tasks with little supervision.

Zendesk’s most recent deep-learning project is Answer Bot—a virtual customer assistant that automatically answers customer questions using content from the Zendesk Guide knowledge base. For example, if a customer sends an email to a shoe retailer asking for help finding sizes, Answer Bot sends the customer relevant articles about available sizes. “For Answer Bot, we liked the idea that a deep-learning model could help the application continually fine-tune itself to give customers the best possible answers,” Cheah says. Answer Bot has helped lead the charge to provide a member-centric experience for hundreds of companies, including Dollar Shave Club. “Answer Bot has been great for us to offer a simple way for our members to find the answers they need,” says Brian Crumpley, analytics manager of member services at Dollar Shave Club. “It’s never about stopping a member from contacting us, but rather equipping the member with the right knowledge and giving them a faster response—it’s a win-win.”

Using TensorFlow on AWS to Get Better Answers

Zendesk relies on TensorFlow—an open-source software library for machine learning—to develop its deep-learning applications. As Zendesk prepared to create Answer Bot, it needed an underlying technology that would enable fast development and easy scalability. “Training algorithms takes a lot of time, and we really wanted to accelerate that process to get a new solution to customers faster,” says Cheah. “We knew the cloud would help us do that.”

The company had already been running its primary platform and an internal data-logging application on the Amazon Web Services (AWS) Cloud, and it knew AWS would also be the right choice for deep learning. Developers using TensorFlow can run the environment on AWS by launching AWS GPU instances. “We already had an AWS foundation throughout the company, and the fact that TensorFlow is bundled in AWS GPU instances was perfect for our needs,” says Arwen Griffioen, a data scientist at Zendesk.

Zendesk uses Amazon Simple Storage Service (Amazon S3) to store initialization files for training models. The company also takes advantage of Amazon Elastic Compute Cloud (Amazon EC2) P2 instances for GPU-based parallel compute capabilities. “The Amazon EC2 P2 instances are very powerful, and using them really helped speed our research capabilities,” says Cheah. Zendesk also uses the Amazon Aurora relational database engine to capture changes made to knowledge-center articles, which are fed back to the Answer Bot training model in near-real time. “We used deep-learning algorithms to perform the process of matching customer queries with articles,” says Cheah.

The company is also excited to make use of newly released Amazon Sagemaker, a fully managed service that enables developers and data scientists to quickly and easily build, train, and deploy machine-learning models at any scale.

"We are excited about the recent announcement of Amazon SageMaker," says David Bernstein, director of strategic technology at Zendesk. "Amazon SageMaker will lower our costs and increase velocity for our use of machine learning. With Amazon SageMaker, we can transition from our existing self-managed TensorFlow deployment to a fully-managed service. Amazon SageMaker also gives us easier access to other popular deep-learning frameworks, while managing the infrastructure for authoring, training and serving our models."

Speeding Development through Faster Deep-Learning Modeling

Zendesk utilizes AWS to easily ingest large datasets used to train deep-learning algorithms. As a result, Zendesk built Answer Bot in a significantly shorter time frame than would have been possible using an on-premises solution. “Our existing predictive-modeling stack was already on the AWS Cloud, which made it faster to develop Answer Bot on AWS,” says Wai Chee Yau, a data engineer for Zendesk. “Instead of purchasing and installing our own hardware, we used the flexibility of AWS to quickly add the GPUs and CPUs we needed.”

Zendesk data scientists can improve the speed of research by relying on AWS. “AWS enables us to try a lot of ideas at once, and that helps us do our research much faster,” says Griffioen. “We can spin up Amazon EC2 instances very quickly as we need them and perform different permutations of our models on those instances without having to wait. We wouldn’t have been able to develop Answer Bot without this capability.”

The company is now exceeding its customers’ expectations for new and innovative customer-service solutions. “AWS enables us to develop and deliver capabilities our customers didn’t have before,” says Cheah. “With Answer Bot, for example, our customers can automatically provide more targeted and accurate answers to their customers’ questions. And because Answer Bot can reply directly to customers with answers in a few seconds, it can resolve support tickets before they reach agents. That can really transform the customer-service experience.”

Zendesk can now scale its deep-learning development environment on demand to meet developers’ requirements for more compute or storage resources. “We can scale our deep-learning models very efficiently using GPU-processing power on AWS, and that will benefit us while we grow our applications to accommodate more customers,” says Cheah. “AWS is a powerful deep-learning ideation platform which we use to conduct the majority of our research,” adds Griffioen. “The flexibility and power we get from AWS have helped Zendesk push past the cutting edge of deep-learning technology in the customer-service space. We’re not just creating different approaches—we’re inventing new algorithmic approaches, thanks to AWS.”