Leverage powerful call center analytics with the Amazon Nova foundation model

Call center analytics plays a key role in improving customer experience and operational efficiency. Fundamental models (FM) can be used to improve the quality and efficiency of call center operations and analytics. Organizations can use generative AI to assist human customer support agents and contact center team managers to gain more nuanced insights and help redefine how and what questions can be asked from call center data.

Some organizations are looking for turnkey solutions to bring generative AI to their operations, such as Amazon Connect Contact Lens, while others are building custom customer support systems using AWS services for microservices backends. This creates an opportunity to integrate FM into systems to provide AI support to human customer support agents and their managers.

One of the key decisions these organizations face is which models to use to power AI support and analytics on their platforms. To this end, the Generative AI Innovation Center has developed a demo application that features a set of use cases that leverage Amazon’s newest FM family, Amazon Nova. This post describes how Amazon Nova demonstrates capabilities in conversation analysis, call classification, and other use cases often associated with contact center solutions. We validate these features for both single-call and multi-call analysis use cases.

Large scale Amazon Nova FM

Amazon Nova FM offers excellent price performance and is suitable for large-scale generative AI. These models are pre-trained on vast amounts of data, allowing them to perform a wide range of language tasks with incredible accuracy and efficiency while effectively scaling to meet large-scale demands. In the context of call center analytics, Amazon Nova models can understand complex conversations, extract critical information, and generate valuable insights that were previously difficult or impossible to obtain at scale. Demo applications demonstrate the capabilities of Amazon Nova models for a variety of analysis tasks, including:

sentiment analysis
Identifying the topic
Weak customer rating
Protocol compliance check
Interactive Q&A

These advanced AI capabilities in Amazon Nova FM enable businesses to better understand customer interactions and make data-driven decisions to improve service quality and operational efficiency.

Solution overview

The Call Center Analytics demo application is built on a simple architecture that seamlessly integrates Amazon Bedrock and Amazon Nova, enabling end-to-end call center analytics for both single-call and multi-call analytics. The following diagram shows this architecture.

Amazon Bedrock – Provides access to Amazon Nova FM and enables powerful natural language processing capabilities.
Amazon Athena – Used to query call data stored in a structured format, enabling efficient data retrieval and analysis.
Amazon Transcribe – Fully managed automatic speech recognition (ASR) service
Amazon Simple Storage Service (Amazon S3) – An object storage service that provides industry-leading scalability, data availability, security, and performance.
Streamlit – Powers your web-based UI and provides users with an intuitive and interactive experience

The application is divided into two main components: single-call analysis and multi-call analysis. These scripts work together to provide a comprehensive solution that combines post-call analytics with historical data insights.

Single call analysis

The application’s single call analysis feature provides detailed analysis of individual customer service calls. This functionality is implemented in the Single_Call_Analytics.py script. This section describes some key features.

Sentiment analysis and vulnerable customer assessment

This solution uses Amazon Nova FM to derive insights about both customer and agent sentiment, as shown in the following screenshot.

Chatbot functionality allows users to ask for an explanation as to why a sentiment was classified as such, or get references from the transcription. This feature allows you to quickly find supporting phrases from the transcription itself to better understand sentiment classes, which can later be used for other analyses.

Vulnerable or potentially vulnerable customers are customers whose personal circumstances make them particularly vulnerable to financial harm or who require special consideration in financial services. The application evaluates whether the calling customer is considered vulnerable or potentially vulnerable by passing the call transcript of the selected call in the next prompt.

vc_prompt = f"""You are a AI Assistant for Banking Call Center. 
Your goal is to determine if the customer in the  below 
qualifies as Vulnerable Customer (VC) or Potentially Vulnerable Customer (PVC).


{speaker_texts}


If the customer qualifies as a VC or PVC, return Yes and explain why.
If the customer does not qualify as a VC or PVC, return No and explain why.
"""

isVC = invoke_llm(vc_prompt, vc_model)

For this prompt, Amazon Nova FM uses a common definition of a vulnerable or potentially vulnerable customer to perform the assessment. However, if your company has its own definition of vulnerable or potentially vulnerable customers, you can use this custom definition to prompt FM for classification. This feature helps call center managers identify potentially sensitive situations and ensure vulnerable customers receive appropriate care and attention along with an explanation of why they were so identified.

Protocol assistance and step completion

The application uses the Amazon Nova model to identify the protocols associated with each call and verify that agents are following the prescribed steps. Currently, protocols are defined in JSON files and are brought in locally at runtime. The following code shows an example of how this is implemented.

protocol_identification_formatted = protocol_identification_prompt.format(transcript=context, protocols=protocols)
llm_protocol_key = invoke_llm(protocol_identification_formatted, protocol_model)

step_completion_formatted = step_completion_prompt.format(protocol_steps=protocol_list, context=context)
step_check = invoke_llm(step_completion_formatted, protocol_model)

This code snippet shows how an application can use a call transcript and a list of available protocols to initially identify the relevant protocols. After the protocol is identified, the call transcript and protocol steps of the determined protocol are passed together and each step of the protocol is checked to see if it has been completed by the agent. Results are presented in an easy-to-use format, allowing administrators to quickly assess agent performance and guideline compliance.

Interactive transcription view and AI assistant

[シングルコール分析]The page provides an interactive transcription view so users can read the entire conversation between the agent and the customer. Additionally, it includes an AI assistant feature that allows users to ask specific questions about their calls.

user_message = call_prompt.format(query=prompt, context=context, chat_history=st.session_state.messages)
ans = invoke_llm(user_message, cb_model)

Powered by the Amazon Nova model, this Assistant feature allows users to gain deeper insight into specific aspects of a call without having to manually search through transcripts.

Multicall analysis

The multi-call analytics feature, implemented in the Multi_Call_Analytics.py script, provides aggregate analysis across multiple calls and enables powerful business intelligence (BI) queries.

Data visualization and flexible model selection

This feature allows users to quickly visualize trends and patterns across multiple calls and easily identify areas for improvement and success.

The “Top 5 Call Topics” visual in the previous screenshot also utilizes the Amazon Nova model. Users can classify the topic of a call by passing the call transcript and letting the model determine the main topic of the call. This feature helps users quickly categorize calls and place them into determined topic buckets to generate visuals. By identifying the main reasons customers call, companies can focus on devising strategies to reduce call volume in these topic categories. Additionally, the application provides flexible model selection options, allowing users to choose different Amazon Nova models (such as Nova Pro, Nova Lite, and Nova Micro) for different analysis tasks. This flexibility means users can choose the model that best suits their specific needs and use case.

Analysis AI assistant

[Multi-Call Analytics]One of the key features of Page is the Analytical AI Assistant, which can process complex BI queries using SQL.

The following code shows how an application uses Amazon Nova models to generate SQL queries based on natural language questions.

user_prompt = """Given the following schema:
{schema}
and a user query, generate a SQL query which can be executed in AWS Athena. 
The table name is {table_name}.

Give the SQL query as a JSON response.
"""

sql_query, chart = invoke_llm(final_prompt, cb_model, "sql")

The assistant can also understand complex queries, convert them to SQL, and suggest appropriate chart types to visualize the results. SQL queries are run against the processed data from Amazon Transcribe, queried using Athena, and displayed in Analytical AI Assistant.

implementation

The Call Analytics demo application is implemented using Streamlit UI for speed and simplicity of development. This application combines specific use cases and AI tasks to provide a sample of what Amazon Nova models can do for call center operations and analytics use cases. For more information on how to implement this demo application, see the following GitHub repository:

conclusion

In this post, we described how Amazon Nova FM powers the call center analytics demo application, which represents a major advancement in the field of call center analytics. By harnessing the power of these advanced AI models, businesses can gain unique insights into customer interactions, improve agent performance, and improve overall operational efficiency. The application’s comprehensive features include sentiment analysis, protocol compliance checks, vulnerable customer ratings, and powerful BI capabilities, giving call center managers the tools they need to make data-driven decisions and continuously improve customer service operations.

As Amazon Nova FM continues to evolve and improve, you can expect even more powerful and sophisticated analysis capabilities in the future. This demo serves as a great starting point for customers who want to explore the potential of AI-powered call center analytics and apply it to their environments. We recommend exploring the Call Center Analytics demo to learn more about how Amazon Nova models integrate into your applications.