Building Agent Co-pilots for Proactive Call Centers
Guy Lecker and Nick Schenone | May 29, 2025
Gen AI call center co-pilots can provide enterprises with operational visibility and insights while automating repetitive tasks, to improve the customer experience. In this session, we’ll show how a large health insurance provider implemented an agentic co-pilot designed scale across multiple call centers and environments.
To dive deep into the architecture and see a demo of the co-pilot, you can watch the webinar this blog is based on. In this session, Guy Lecker (ML Engineer Team Lead) and Nick Schenone (Senior ML Engineer) from Iguazio, acquired by McKinsey) walk us through how generative AI-powered agent co-pilots are revolutionizing the call center experience.
Use Case: Real-Time Call Center Co-Pilot
A major health insurance provider faced a familiar challenge: agents in their call centers were struggling to consistently guide customers in real-time through complex healthcare decisions, while following strict and data-heavy call scripts and also upselling during calls.
The Solution: Combining a Real-Time Agent Co-Pilot with a Post Call Analysis Report
The solution was designed as a multi-faceted architecture, which includes both a real-time agent co-pilot and a post-call analysis report.
The Agent Co-Pilot
The co-pilot enhances agent performance live during calls. The client-facing UI was designed with four main components:
- Live Transcription Panel – Captures real-time audio conversation in chat format and tracks relevant data as the call unfolds.
- Call Steps Panel – Ensures agents follow every step of the sales process. It presents the script, highlights skipped steps and places where agent performance could be improved, and changes dynamically according to the call.
- Agent Suggestions - Provides suggestions for the call, helping agents stay compliant. The agent will suggest answers and address objections. For example, if the client comments they are living a healthy lifestyle, the co-pilot will suggest an article about an athlete who caught a disease.
- Data & Plan Panel – Displays client-specific data (like age, health background, smoking status) and recommends suitable insurance plans as upsell opportunities based on the profile and data from the call.
The Post-Call Report
The report is a post-call analysis that helps improve human performance for future conversation. It includes:
- A short summary of the call
- Call analysis to identify areas of improvement and missed opportunities.
- Agent score based on empathy, professionalism, kindness, and additional parameters.
Under the Hood: The Workflows Architecture
To power the UI, the GenAI app was built with four workflows running under the hood:
- Transcription Workflow - Captures and converts live audio into text. It is triggered by the audio stream from the call center.
- Analysis Workflow - Performs sentiment analysis and extracts demographic data (like gender and age) to recommend suitable insurance plans. This workflow is triggered per message.
- Guidance Workflow - Provides live objection handling scripts, product facts and sales suggestions based on the conversation so far, allowing agents to follow guidelines without memorizing scripts. It is also triggered per message.
- Post-Call Workflow: Summarizes the call, evaluates script adherence, assesses soft skills and provides coaching tips. This is a batch workflow.
Let’s dive into each workflow.
The Transcription Workflow
This workflow is fed the latest transcription (from previous calls) and a new audio chunk (existing call). Voice activity is detected to reduce cost and ensure transcription only operates when someone is speaking.
Since this flow (and all of them) are modular, we can switch various components based on needs. For example, using open-source MLRun for transcription and translation, or disconnecting them and using Amazon Transcribe and Translate instead.
The output is the call transcription in text form
The Analysis Workflow
This workflow analyzes the text, extract data sentiments and recommends a plan.
It is fed past analysis and conversation history as well as unseen text (the transcription). Then, the workflow extracts properties like gender, age, etc. from existing databases. It also conducts sentiment analysis. Then, it provides a plan recommendation based on existing plans and pricing as well as a sentiment analysis and summary of extracted data.
The Guidance Workflow
This workflow tracks the conversation according to the scripts and offers suggestions based on the existing knowledge base. It is in this workflow that the copilot provides the objections info to the human agent.
The workflow is fed past guidance (to avoid duplications), unseen text, which is the new transcription, and the conversation history. It then pulls data from objection scripts, the main script and the knowledge base that have facts and data for automated classification.
In the end, it generates guidance and suggestions.
The Post Call Workflow
This workflow provides a detailed summary of the call with scores for the agent on soft skills and performance. It is a batch workflow.
It is provided with the transcription and conducts call summarization, script performance evaluation and soft skill assessment, resulting in the call summary, scores and explanations for skills and steps.
Under the Hood: MLRun AI Gateway and Monitoring
To power these workflows, Iguazio’s MLRun deploys each one as a serverless function (with Nuclio) connected to an AI gateway. This allows full modularity and monitoring across the entire project, as well as per workflow, per system and per model.
Next Steps
Building generative AI-powered co-pilots for call centers is no longer a futuristic idea. It’s a practical, scalable solution already delivering real-world results. The integration of real-time assistance with post-call analytics can dramatically elevate agent performance, ensure compliance and uncover upsell opportunities, all while delivering a better customer experience. With MLRun’s serverless architecture and AI Gateway, this solution is both adaptable and production-ready across multiple environments.
