Understanding call center sentiment scores and how to use them in your business
- Sentiment scores provide deep insight into the customer experience during call center interactions
- Contact Lens for Amazon Connect breaks sentiment scores into overall scores and individual scores for each quarter of the call
- Have a strategy for which calls sentiment analysis will provide the greatest insights
Contact Lens for Amazon Connect provides a set of machine learning (ML) capabilities supporting real-time and post-call analysis of call center interactions. While Contact Lens for Amazon Connect can be applied to voice and chat interactions, let us use the term call to refer to both.
Start with the second part of that sentence: “supporting real-time and post-call analysis of call center interactions.” This is a powerful capability. Whether you are training new agents or monitoring call quality, Amazon Connect allows a supervisor to immediately recognize a call sentiment problem and step in to help agents at any time. Post-call analysis provides an opportunity to understand the good and bad experiences customers have had with your products and your call center.
Now let’s look at the first part: “It provides a set of machine learning (ML) capabilities.” Contact Lens allows users to create a set of categories with phrases to listen for. A simple example would be a positive category tagged “Politeness,” and a phrase in that category might be “Thank You” or “You were great.” The Contact Lens will hear or read that phrase and tag that interaction as positive. Each category is identified as positive, negative, or neutral.
Sentiment Scores are ratings assigned to a call or chat transcript indicating the call or chat was mostly positive, negative, or neutral based on ML analysis of the terms and phrases used. Scores are assigned to the call as a whole and to each quarter of it. Managers can search the sentiment scores to identify these ratings and use that data to improve call center performance.
Let’s look at how transcripts are scored and how we can use that scoring to improve the experience of our agents and customers.
Sentiment Scores and Talk Time
Sentiment scores range from +5 for Positive to -5 for Negative, with 0 as Neutral. Every conversation is a back-and-forth interaction between two parties, and each turn in the conversation is analyzed by Contact Lens to assign it a score in the positive or negative range.
The call is reported with a sentiment for each quarter determined by the frequency and proximity of the scores from each interaction. An overall score for the conversation is calculated as an average of the quarterly scores.
Sentiment scores are tilted toward the customer experience. The analysis includes the agent interaction only in an effort to determine customer satisfaction. In addition to the sentiment scores, the analysis provides statistics related to the talk time, non-talk time, customer talk time, and agent talk time.
It was important to understand what you want to get from Amazon Connect. Do you want to monitor and help new agents? Improve the experience of customers? Improve the effectiveness of your agents and maybe teach them how to handle unruly customers?
Start with the sentiment scores. Overall call scores tend to form a pattern based on the four-quarter scores that allow you to identify which calls you should focus on the improve call center performance.
Neutral calls are often customers looking for information. Focus your ML terms so you can separate information calls for non-problem analysis.
Calls that are positive from start to finish are not generally helpful because it’s hard to find lessons in them to improve upon. Statistically, if the total number of neutral and positive calls is over 90%, that’s very good.
Look closely at negative to positive trends (calls that start negative and end up more positive. These allow you to see which issues customers are calling about that make them upset, which is great information for your product team. Also look at what the agents did to turn the call around These calls are fundamentally successful, and you should celebrate those agents while learning from them.
Positive to negative and all negative calls are critical for revealing where your call center needs to improve. Look at how the agents involved were trained and the systems they used to help customers. Talk time can also be very informative.
If the agent wasn’t talking, maybe they were having problems with the system. Did you have a new release that you could have provided better training on?
If the customer is doing all the talking, maybe there was nothing the agent could do. But you can listen to the call to understand the issue the customer experienced and why it might make them so upset that they called to vent.
If the agent is doing all the talking, review the call. This is not a best practice for positive customer support experiences, and the agent may need additional training.
If the call time between the parties is balanced, but the issue was not resolved, this was still not a successful call. Review the content to understand the customer issue and why it was not resolved. Maybe the issue was resolved, but the agent was not able to turn the sentiment of the call around?
Experience and hindsight
As you continue to use Contact Lens for Amazon Connect, look for ways to configure your ML categories, terms, and phrases to capture technical issues as well as some subtle, nuanced emotional points. Customer feelings are not stated in so many words.
Agents often appreciate the grading of sentiment scores. Interactions are not just good, bad, or indifferent but graded on a scale of 1 to 5. A score of 3 might be positive, but a 5 is stellar! A -2 or -3 might be worth analyzing, but the agent has an opportunity to turn that around.
Analyzing sentiment scores and call times is just the starting place with Amazon Connect. The Amazon Connect API allows you to pull data into your own systems for alarms and notifications, as well as generate management reports. The Lex bot can help cut down on agent time by collecting customer information upfront and, in many cases, resolving issues without agent involvement.