Case study · Contact center · AI/ML and streaming

Real-time call insights on Confluent Kafka and AI/ML for a leading contact center company

A leading contact center company streams call audio through Confluent Cloud, transcribes it in real time, and gives agents and supervisors sentiment and keyword insights while the conversation is still going.

Industry
Contact center
Capability
AI/ML and streaming
Platform
Confluent Cloud, Elasticsearch
Implementation period
8 months

Context

The client is a leading contact center company with a global customer base. It wanted to improve customer experience and understand its customer interactions in more depth, and approached VerticalServe to build a real-time call insights platform using Confluent Kafka and AI/ML.

The problem

The company needed a scalable, real-time way to identify patterns and trends in customer interactions and to put actionable insight in front of contact center agents and supervisors.

What was built

VerticalServe designed and implemented the platform end to end, from the telephony system to the application agents and supervisors use.

Flow: RTP audio streams and call metadata enter Kafka topics on Confluent Cloud, are transcribed and analysed for sentiment and keywords, and land in Elasticsearch behind the Call Insights application. TELEPHONY RTP audio streams Call metadata CONFLUENT CLOUD + AI/ML Kafka topics Real-time transcription Sentiment and keywords INSIGHTS Elasticsearch Call Insights application

Flow: RTP audio streams and call metadata enter Kafka topics on Confluent Cloud, are transcribed and analysed for sentiment and keywords, and land in Elasticsearch behind the Call Insights application.

  • Confluent Cloud setup: managed Kafka clusters as a scalable, secure platform for real-time streaming.
  • RTP audio handling in Kafka: Real-time Transport Protocol audio streams from the telephony system are handled through Kafka.
  • Call metadata: caller ID, agent ID, call duration and call type are extracted and ingested into Kafka topics.
  • Audio transcription: AI/ML speech-to-text engines transcribe call audio in real time.
  • Utterance detection and speaker handling: algorithms identify the different speakers and detect specific utterances or phrases.
  • Sentiment detection: NLP and machine learning determine the sentiment of customer interactions.
  • Keyword extraction: keywords and phrases relevant to the business are pulled from the transcripts.
  • Insights in Elasticsearch: the generated insights are integrated into Elasticsearch for real-time access and visualization.
  • Call Insights application: a custom application where supervisors and agents access and visualize the insights.

Stack

Confluent Cloud Confluent Kafka AI/ML speech-to-text NLP Elasticsearch RTP

How it was delivered

The engagement ran for eight months. VerticalServe set up the Confluent Cloud environment, brought RTP audio and call metadata into Kafka, added the AI/ML stages (transcription, speaker and utterance detection, sentiment, keywords), integrated the output into Elasticsearch, and built the Call Insights application on top.

The result

20%
Increase in customer satisfaction scores
15%
Increase in agent performance metrics
30%
Reduction in escalated calls
  • Customer experience: agents could address customer concerns quickly using real-time insights, and customer satisfaction scores rose 20%.
  • Agent performance: with real-time feedback, agents adapted their approach during calls, and agent performance metrics rose 15%.
  • Operations: supervisors could see trends and patterns in call data and use them to optimize call routing and staffing decisions.
  • Proactive issue resolution: keyword extraction and sentiment analysis surfaced emerging issues before they escalated, and escalated calls fell 30%.
  • The company has since extended the platform to additional locations and plans to integrate it with other data sources.

The figures above are as reported in the original write-up.

What it means for a similar team

  • Treat call audio as a stream. Insight that arrives during the call can change its outcome; insight that arrives the next day can only report on it.
  • Carry call metadata alongside the audio. Caller, agent, duration and call type are what make transcripts searchable and comparable.
  • Put the output where agents and supervisors already work. Here that was a dedicated application on a search index.

This case study was first published on our Medium publication. Original article

Engage

Start a similar POD

Tell us the outcome you want and your constraints. We will reply with a discovery proposal.

Talk to us

Working on something similar?

An Innovation POD is a dedicated team that builds inside your environment. Tell us what you need to ship.

Start a POD