Upwork, a platform that connects professional freelancers with businesses, has grown quickly over the past few years, especially since its IPO in late 2018.
The rapid growth created scaling problems, as the company couldn't hire people fast enough, according to Joe Wang, senior director of customer experience at Upwork. As a way to help handle the employee deficit, the company adopted an automation-first strategy, and began working with Ada, a 2016 startup based in Toronto that sells a customer experience chatbot.
Internally, Upwork relies heavily on Zendesk, the ubiquitous customer service platform. The platform powers its customer support capabilities and connects its hundreds of support specialists on a common system. Ada's tight integration with Zendesk enabled Upwork to add its customer experience chatbot capabilities to the already-established Zendesk platform with just a single line of code, Wang said.
Still, Upwork took its time deploying Ada, only using the bot for a limited number of self-service questions as Upwork tested its capabilities and ensured the bot could answer the questions correctly.
"We started small," Wang said. "We didn't want to go crazy."
Joe WangSenior director of customer experience, Upwork
Ada, according to Mike Murchison, CEO and co-founder of the company, builds bots specifically for customer service, and often sells directly to customer service departments.
Usually, he said, customers are able to get bots up and running within 30 days, and it's a low-code, no-code setup.
"We focus on building intent recognition specifically for customer service," he said.
Building a bot
For Upwork, getting Ada ready took about a month, but it was a month of hard work, mainly because Upwork had to create branching decision trees for question variations. Sometimes a customer will ask a question in a way that Upwork hadn't dealt with before, so new content has to be developed for the bot so it will understand the question and answer it correctly.
"Even though the bot is able to identify which answer is correct, we have to develop the answers," he said.
In some situations when a bot is unable to answer a question, it can connect with a human support employee through Zendesk, who can take over long enough to answer the question before giving control back to the chatbot, Wang said.
The customer experience chatbots are able to answer a variety of questions on their own, Wang said, but Upwork is working on enhancing the bots further by adding better integrations with various APIs used by the Upwork platform.
The integrations will enable the Ada chatbot to have some personalized customer information prefilled when a customer connects to it, saving the customer time by not having to input the information themselves and enabling the bot to provide better responses.
While Upwork may have started with only a small Ada application, the customer experience chatbot can now answer most of the self-service questions, Wang said.
To Upwork, Ada is more of a search technology than a "deflection," a tool to avoid human interaction with a customer. Customers can ask the bot questions about their account or technical problems, for example, as opposed to having to look up answers manually in a FAQ section.
"We see the chatbots as an upgraded version of search," Wang said.
The bot's accuracy has improved since Upwork first began using it because the team has manually trained it and it has learned on its own by machine learning.
Still, there are problems. Customer experience chatbots are still somewhat of "an uncharted territory," Wang said. Ada occasionally responds to questions incorrectly. Also, it's only available in English at Upwork, and has trouble understanding the grammatical mistakes that users with limited English tend to make.
"That's something we are still trying to figure out," Wang said.
Still, "customers actually like it," he said of Ada.