The science behind satisfaction prediction

How great would it be to know a customer will be satisfied before their ticket is resolved? That’s the idea that drives Satisfaction Prediction, Zendesk’s tool for recognizing potentially positive or negative customer experiences in real time.

Since its release in early 2016, Satisfaction Prediction has helped hundreds of companies quickly identify and escalate tickets that need special attention. If a ticket is flagged as “at risk” for customer dissatisfaction, it can be reprioritized for a faster resolution or passed off to a more specialized agent. Satisfaction Prediction works in tandem with a company’s customer satisfaction (CSAT) score, that crucial metric for knowing the actual effectiveness of your customer service.

We’re going to dive a bit deeper into the guts of Satisfaction Prediction to explain how its machine learning model can help raise your company’s overall CSAT score and maintain happy customers.

The science behind Satisfaction Prediction

Satisfaction Prediction looks at three main metrics of a ticket to tell if a customer will be satisfied. Those metrics are:

• Time metrics: When the first reply time was, how long the customer waited for a reply, and how long the ticket has been open
• Ticket text: If there have been similar conversations before (with certain prevalent words), and if the customer was ultimately satisfied or dissatisfied with the results
• Effort metrics: How much effort is being put into the ticket, noting details like re-opens, time between replies, or reassignments to other agents

Here’s a visual representation:

Now, you’re probably wondering “what happens in that predictive model?”. It takes into account all of the positive and negative CSAT ratings, along with the time, text, and effort metrics. It then uses this data to develop a “standard” for future tickets to be compared to. The machine learning model can then predict if a customer will be left satisfied by the provided service.

Looking to the past to predict the future (and gain better CSAT)

Here’s where things can get a little bit tricky—but may also lead to revelations of what’s working (and what isn’t) in your efforts to raise your overall CSAT score.

A Satisfaction Prediction score is given as a number between 0 and 100. The higher the score, the more likely the customer will be satisfied. If something occurs in a ticket that led to a lower CSAT score before (like a belated reply or poorly communicated directions), the Satisfaction Prediction score will drop. Those factors are what made a customer say they were ultimately unsatisfied with the service they received.

Customer satisfaction makes a big difference

Customer satisfaction (CSAT) can have a major effect on your business. Read our research report to learn the full details and insights.

Learn more

Take a look at this performance from a real customer using Satisfaction Prediction. This score follows a similar trend for all of our customers:

The red line is what we call the “alert threshold”. It’s calculated by taking the long term CSAT average (in this example, 63%) and calculating the difference to 100 (i.e. 100 – 63 = 37). The Satisfaction Prediction scores below 37 often led to an unsatisfied customer, while the scores higher than 37 generally resulted in satisfaction.

Over time, as Satisfaction Prediction helps average CSAT improve, the threshold decreases. If 95% of your customers reported satisfying service, your alert threshold would only be 5 and scores between 6-100 would generally lead to customer satisfaction. Happy customers mean (a lot) more wiggle room with Satisfaction Prediction scores.

But bear in mind: If a Satisfaction Prediction score is lower than the threshold, it doesn’t automatically mean the customer will be unsatisfied. Here are a few things to note when looking for insights in Satisfaction Prediction:

Nobody’s perfect: Getting a high Satisfaction Prediction score all the time is simply unfeasible, but keeping a majority of your customers is always possible. If the score is consistently above the threshold, you’re doing good work.

Note when the changes occur: Saw your score change after a reply? Review the details, like the length of time between replies or how much attention the ticket gets from other agents. Paying close attention to detail will expose trends that can be capitalized on, both in terms of what’s working and what isn’t.

Red alert: If a Satisfaction Prediction score is below the alert threshold, it could ultimately have a negative impact on your average CSAT score. These are the tickets that deserve special attention; they’re indicative of areas for improvement. Be sure that they’re escalated to mitigate the customer’s frustration and (ideally) leave them satisfied when all is resolved.

Making better agents through better predictions

CSAT is key, but it conclusively answers only one question: “Was the customer satisfied with their service?” Satisfaction Prediction gives agents an extra angle toward understanding why the customers were left happy or unhappy through the real-time changes, while also giving them a chance to swing things back toward satisfaction. It’s another way that machine learning is making humans beings better-equipped to provide excellent customer service.