Cut support costs with a window into contact center dark data

Key points to unlocking your customer service dark data: 

• Contact center is the heart of your customer conversations where brand loyalty relationships can be built or broken costing $9.7 million/year due to bad data 
• True cost of bad data: customer churn, agent attrition, & operational efficiency & time 
• Simplify the complexity to understanding your conversational data customer service topics and reasons to improve customer service 

What does customer service have to do with dark data? 

• Language data exists in vast quantities across the digital world but hasn’t traditionally been a source of insight 
• Customer service unstructured dark data examples: call-center transcripts, text messages, survey responses, customer reviews, emails, documents, and more
• The challenge: bringing order to the chaos of dark data problems—too much data, inaccurate data, & poor data quality
• Example of challenge: manual review of transcript data that costs you time and resources
• Example of solution: simplified and automated process to understand your contact center conversations and customer service topics across channels with your agents and bots 
• The result: data sources become functional & insightful to improve customer service vs. dead weight 

‍

Simplifying Data Complexity

Unstructured dark data source example chat transcript

‍

Chat transcript -> customer service topics  

Chat transcripts -> customer service topics -> reason

‍

Balancing NLP & Prompt Engineering

Constraints of solely relying on prompt engineering are cost & precision:

• Bound by API constraints (e.g. limited adjustable parameters, guardrails imposed by creators, compute time - e.g. 20 seconds per prompt)
• Prompt engineering isn't a solved problem and may not be fully optimizable in principle due to the natural language format
• Traditional NLP methods offer precise control over pipeline/output + are more optimal in some cases (e.g. topic clustering for vectors)

Harmonizing prompt engineering + traditional NLP methods based on what is optimal. Example workflow:

1. Vectorize language data
2. Cluster into topics
3. Feed topic data into prompts for label creation

‍

‍

‍