Best practices for using suggested cases

Keyword-based suggestions are rudimentary; suggestions should apply only to Case Types where the CSRs utter the keywords for that specific interaction. Overlapping the keywords in a conversation for another Case Type can lead to an inaccurate Voice AI suggestion (false positive).

Apply the following methods to gather data and identify a list of words or phrases related to a Case Type:

• Use subject matter knowledge: Existing subject matter expert knowledge can help you identify the keywords and phrases that trigger the Case selection. This method is not scientific and can result in many missing words and phrases in the interaction.
• Review call recordings: A Business Architect can review call recordings that apply to each Case Type in which they try to identify the keywords. This method is a scientific approach, but it might be difficult to correlate the Case Types to interactions. Additionally, it is the most time extensive form of research.
• Review messaging transcripts: If the customer is using Pega Digital Messaging, the transcripts for those conversations can provide insight into the keywords to use in the Case suggestions. Messaging conversations often have commonality with the speech found in a voice interaction.
• Enable Voice AI to gather only transcripts: This research method takes longer than the typical product deployment but yields the most accurate results. As a best practice, enable Voice AI for several CSRs over a short period (two weeks) to do nothing but gather transcripts for review. Then, the system can create a report for phone interactions that handle specific Case Types. You can then review the transcripts for those interactions.

Using this approach can change the project plan, which requires prompt completion of the development and lower-level environment integration and the deployment of Voice AI to production. A transcript review step can occur after the go-live to identify the interactions that did not receive Case suggestions. Still, you use a Case Type to review the transcript for improvements.

To create a machine learning model, Voice AI requires a list of utterances that led to the Case suggestion. With Voice AI machine learning Case suggestions, you want to trigger what a CSR might hear that causes them to click Add Task and select a Case Type.

To populate the model, the customer can help define the Case Types by providing sample utterances from each type of participant in the call. For example, the following table contains samples for a health insurance member calling about various Case Types:

If NLP analysis is active for both the customer and the CSR, you also want to collect the utterances that a CSR makes that might lead to a Case suggestion, as shown in the following table:

Sometimes, the conversation might include another "Persona" on the customer channel. For example, in health care, a customer service department might receive calls from health care members and health care providers. You can generate utterances that the other Persona might make, as shown in the following table:

You can use the Case Suggestion Feedback Form to gather information for implementing suggested Cases:

You can also use generative AI to create additional suggested phrases. To generate additional utterances, you can issue requests to the generative AI engine; for example:

• Provide 20 utterances for a health insurance member who inquires about the status of a claim.
• Provide 20 utterances for a health insurance CSR who asks a member what type of health care provider they can help find.
• Provide 20 utterances for a health care provider contacting a health insurer indicating the reason for their call was due to a claim denial.

You can add the utterances to a spreadsheet and then review them with the customer for validity. If the customer believes that a CSR might create a Case Type based on the spoken utterance, then it is a valid use case. You can add the utterance to the Case suggestion spreadsheet.

Each topic needs at least 30 suggested utterances. The larger the volume of utterances, the better the model performs. In some Cases, you might want to include hundreds of phrases.

After you have generated the phrases and utterances, you use this data to train the topic model. You can use Prediction Studio to review and update your topic models. For example, for a topic named Manage claims, you will find a topic model for Manage claims in Prediction Studio. The Manage claims model may initially contain only the keywords associated with the topic. You can import your phrases/utterances into the model, then rebuild the model.

For information about importing the data and rebuilding a topic model, see Importing topics to text predictions.