maxkabakov - Fotolia
Intuit Inc.'s data science team wanted to see if AI chatbots could take some of the customer service workload off the shoulders of human agents. The answer: Yes, but in a limited way -- at least for now.
As a result, the vendor of finance and accounting software for individuals and small businesses is taking a "very constrained" approach to chatbot design, according to Andrew Mattarella-Micke, a senior data scientist at Intuit.
Initially, Intuit is building relatively simple chatbots to handle a targeted set of tasks involving information retrieval and automated responses to customers' frequently asked questions.
"The models that we can apply are pretty general," Mattarella-Micke said. "We're building the types of features that are going to represent the bulk of the common things customers are looking for."
In a session at the 2018 Strata Data Conference in San Jose, Calif., Mattarella-Micke shared tips for designing chatbots based on his team's experiences with the AI technology. Chatbots offer some clear benefits in customer service operations, he said. For example, they are available 24/7, can be quickly scaled up as customer inquiries increase and are more interactive than conventional search tools.
Chatbot development challenges
However, he also cited hurdles that can trip up chatbot design efforts. Enabling chatbots to engage in a real dialogue with people is a tall order, Mattarella-Micke said. The same goes for ensuring that the technology can meet user expectations on service and information delivery.
"Threading the needle exactly on what users want and what they get is a difficult challenge with chatbots," he warned.
That's why Intuit isn't asking its chatbots to do more at this point. They aren't being designed to carry on lengthy conversations with users of the Mountain View, Calif., company's software; interactions tend to be brief, without a lot of back-and-forth discussion.
Andrew Mattarella-Mickesenior data scientist, Intuit
"Dialogue is hard, but the best way to tackle that is to make the conversation as simple as possible," Mattarella-Micke said. "We've created very basic rules for responding to what customers are saying. It's less of a dialogue now than a sort of informed lookup that we're iterating on."
Nor are the chatbots being set up to address particular issues faced by individual customers.
"We can have customers asking about things that are very specific to their situations -- that's not something we can tackle in our [chatbot] applications," he said.
Keeping it simple on chatbot design
The chatbot's limited uses will be supported by a mix of retrieval-based chatbots and ones that base their responses on natural language understanding (NLU) rules.
Mattarella-Micke and his fellow data scientists decided against using more complex generative sequencing and reinforcement learning chatbot models, which he said require larger inputs of relevant data to work effectively and "can easily go off the rails" in responding to customers if they aren't designed properly.
The NLU piece of the chatbot design process involves building an understanding of Intuit's product vocabulary into the models -- so, for example, a chatbot knows that deleting and removing invoices mean the same thing. Embedding sentence vectors to help the chatbots predict the right responses based on what customers say "has been a very fruitful approach for us" on NLU, Mattarella-Micke said.
While the chatbot implementation strategy is narrowly focused now, Intuit does have some bigger plans in mind. For example, Mattarella-Micke said one of the data science team's future goals is to build a "chatbot orchestrator" that can assess incoming customer inquiries and route them to different chatbots based on the information or assistance people are seeking.
Why it's important to respond to all customer inquiries on social media