Enhance Your Call Center Using AWS Connect and GenAI

The adoption of AI across organizations has never been faster or more aggressive, and engineering teams are increasingly building information products that process diverse data streams in production environments. A question always arises: How can we implement AI quickly in our organization? 

In this blog, we focus on human interaction experiences, specifically call centers, and how Amazon Connect can enable organizations to build end-to-end call flows, from agent configurations to a wide range of automations. However, a critical gap remains: how do we leverage the data generated by these interactions to meaningfully improve customer service? And how can GenAI minimize the delay between a customer call and the direct actions taken to improve the customer experience? 

AI Everywhere

Since the popularization of LLMs following the launch of ChatGPT 3.5 by OpenAI in 2022, companies, organizations, and governments have started incorporating AI into their products and services, with varying degrees of purpose and effectiveness.

AI has demonstrated clear value in enriching user experiences and in building information products from interaction data. This is particularly important when part of these experiences involve human-to-human interaction. By analyzing these interactions with AI, organizations can extract diverse information. By operationalizing those insights, they can develop information products with the potential to dramatically improve customer experiences.

How to Effectively Use AI in Call Centers 

Call centers require continuous human-to-human interaction. Although there are automated paths to resolutions and voicebots, there will always be scenarios that require a live agent. This is where we can use information products to improve these interactions or maintain them, taking timely action when we find something worth addressing.

What matters when analyzing a customer service call depends on the organization's core objectives, but in this case, we focus on four key areas: 

1. Understanding the call's sentiment, 
2. Identifying improvement points
3. Communicating those findings promptly
4. Flagging calls that take an unprofessional or inappropriate turn.

From Concept to Implementation

With AWS, businesses that depend heavily on human-to-human interaction, such as call centers, can leverage Amazon Connect to implement their call flow configurations and virtual agents. User experiences can be improved by adding recording functionality and storing call audio in Amazon S3. From there, the question becomes: How can we enrich the metrics and build better information products from that audio source? Let's start exploring the AI services that can support these needs and how they can work together within an event-driven architecture.

On-demand AI in AWS

The pipeline begins with Amazon Transcribe, a service that converts audio files to text. This first step can analyze the conversation, extract information, and retrieve more reference points with help from other AWS services

The next step after processing the audio transcription is to detect the call’s sentiment using Amazon Comprehend to determine if it was positive or negative. Next, Amazon Bedrock, using on-demand Large Language Models (LLMs), will generate a summary, a business improvement action, and classify the call as personal or business-related.

To deliver a unified result from processing and analyzing different data points with unique formats, Amazon Bedrock will synthesize the results, which are then converted into audio using Amazon Polly. The results and corresponding text are sent via email using Amazon SES to the designated recipients so they can take the required actions and improve service quality.

Observability and Persistence

Observability is very important, so we'll create custom metrics in CloudWatch not only to monitor infrastructure but also to capture quantifiable call information and display it in a dashboard.