What is the worst outcome for a retailer? When a disappointed customer decides to shop someplace else. Zendesk wants to avoid that outcome.
The San Francisco-based SaaS company just came out with its Satisfaction Prediction feature last week, now in beta with its enterprise customers. Satisfaction Prediction relies on a deep pool of existing data, where previous calls can be rated according to feedback given in previous transaction surveys, explained Jason Maynard, director of data and analytics at Zendesk.
Customer calls and e-mails are rated, flagging the complaints that raise the risk of a negative outcome. That warning calls for special attention by a call center representative, or his or her supervisor.
Natural language processing (NLP) zeroes in on key phrases and words during the conversation between the call center and the customer, looking to identify any snippet of chat that indicates customer dissatisfaction. Steps could then be taken to salvage the situation and turn the potential loss into a win.
Satisfaction Prediction will generate a score ranging from 0 (bad) to 100 (good) during the customer call. The higher the number, the greater is likelihood of a good result for the customer and the retailer.
Machine learning also comes into play, as caller history is compiled and organized, again flagging any high-risk of a negative outcome, based on the customer history. “The more consistent consumer behavior…the more the model recognizes those patterns,” said Maynard.
Satisfaction Prediction builds a model for each customer, based on previous call history. The model then refreshes every week, based on new data that may come from the next call from the same customer. If a customer had a series of bad experiences with a retailer, that will be noted and flagged as a “high risk”.
“The agent wants to head off that bad experience before it escalates,” Maynard said.
He gave an example of one transaction. A customer receives a pair of shoes through the mail. The packaging and the merchandise are damaged. He sends an e-mail complaining about the damaged goods. Satsifaction Prediction would compare the incoming e-mail with the customer history, flagging it as a high risk because the complaint risks a bad score in subsequent surveys. The agent can try to turn the bad outcome to a good one by offering free shipping to return the damaged shoes and replace them with a good pair.
For Satisfaction Prediction to work, an enterprise needs at least 1,000 pieces of feedback, Maynard said. It will take some time for the beta version to run before “we see if we can shrink the number down.”
Zendesk hopes to offer Satisfaction Prediction as a product in early 2016 for enterprise-level customers.
