Before you build
Before you begin creating an application, you must start with a well-defined and well-understood problem and an idea of how this use case can be measured.
For example, you may have a goal to improve Agent adherence to a script specifically designed for customer churn interactions. In this case, you may want to 1) identify interactions where customers called to cancel an account and 2) score how well the responding Agent followed the appropriate script. A well-structured application can readily surface both poor and well-performing Agents and measure their improvements after targeted-training.
You can use the following questions to help refine your own application criteria:
What problem am I trying to solve?
Is constant analysis required to understand this problem?
What types of calls do I need to identify?
What speaker phrases or metadata can be used to surface this problem?
What data can be used for measuring meaningful differences in performance?
Is it sufficient to know that a call falls into a defined category or do I need to know how closely that call matches the category?
Once you have defined your criteria, you can begin thinking about how to structure your application. You will identify any logical groupings that can be used to define application categories, and then begin to use the accumulating data to populate your categories. Application categories use a combination of metadata filters, include phrases, and exclude phrases to match and score calls.
Best Practices for Application Development
Applications should be structured so that the same score type (high or low) in every category is considered "good" or "bad". This makes is easier to understand and visualize the data.
For example, if you want to measure customer experience, we recommend separate applications for measuring negative experiences and positive experiences.
While developing your application, it is best to use a well-understood sample data set. Your data set should contain both transcripts that you know expose the problem you are trying to solve, and some that do not.
A smaller data set will allow your to develop your application more efficiently, by cutting down on reprocess times.