Chaos Gears helps insurancemarket analyze 100% of calls using AWS

Insurancemarket has a mission to make insurance simpler, more transparent, and more accessible for consumers. As a leading insurance aggregator, the company helps people make more-informed decisions through its digital platform. It serves more than 500,000 customers and handles 1 million contracts per year.

Opportunity Supporting agents and customers for Insurancemarket

Every month, Insurancemarket receives more than 10,000 calls, reaching up to 40,000 during high-demand periods. Each carries details about coverage confusion, complaints, and unmet customer needs. Because manual reviews required roughly 20–25 minutes per call and two full-time employees, the company could check only 3–4 percent of calls each month. The rest went unheard and unreviewed. This approach gave managers an incomplete view of service quality, customer needs, and agent development requirements. It also made it difficult to spot less common but important issues.

This limited view slowed Insurancemarket's response to recurring concerns. For example, confusion about a specific policy clause or frustration with a delayed document might surface across multiple calls before the pattern becomes visible and the issue escalates. To identify service issues earlier and support agents with more relevant coaching, the company needed a consistent way to determine whether agents followed the required steps. Equally important, Insurancemarket wanted to respond to customers, especially those who needed additional information or quick follow-ups, before dissatisfaction affected the broader business relationship. So the insurance aggregator needed to move from statistical sampling and guesswork to a complete, timely view of every interaction.

Solution Building an AI-powered call analytics system on AWS

Already running most of its infrastructure on AWS, Insurancemarket turned to the cloud provider for assistance in automating call analysis. AWS connected the insurance aggregator with Chaos Gears, which has deep experience in AI engineering and voice applications. Working with Insurancemarket through several calls each week, the partner translated the company's business and quality assurance requirements into a technical design for Greek-language calls. After 2–3 weeks of scoping, the team produced a working prototype for the system, Voxflow, within 1 month and completed the production solution 2–3 weeks later.

At the heart of Voxflow is Amazon Bedrock, a solution for building generative AI applications and agents at production scale. With access to Anthropic's Claude Sonnet in Amazon Bedrock, the company can handle transcription analysis and quality evaluation for each call. The system stores recordings in Amazon Simple Storage Service (Amazon S3), object storage that's built to retrieve virtually any amount of data from anywhere. To process each recording as it arrives, the team uses AWS Lambda, a service for running code without thinking about servers or clusters. Amazon EventBridge, an event-driven integration service, schedules the workflow. By adopting this fully serverless architecture, the team removed the need for manual intervention.

Voxflow does more than transcribe calls. It separates speakers, identifies the call's subject, assesses the agent performance, and detects the customer's tone and sentiment throughout the conversation. The system does not only check if an agent said hello, says Tomasz Dudek, Head of AI at Chaos Gears. It evaluates the agent's empathy, tone, and success in making the customer feel heard. That's a qualitative leap. Without listening to a single recording, supervisors automatically receive a structured report for every call. This approach applies the same evaluation criteria to every call, producing consistent outputs for review.

Outcome Analyzing almost every call and responding to issues in NRT

Where Insurancemarket once needed months to work through a meaningful sample of calls, Voxflow now analyzes virtually 100 percent of calls. Processing 400–500 calls in parallel per hour, the system can scale up during peak times. Insurancemarket has redeployed the two employees who were previously dedicated to repetitive manual call review to critical and meaningful tasks, enhancing operational efficiency. Voxflow also supports internal discussions about performance improvement, agent coaching, and employee development. By gaining call-level insights, managers and team leads can have more-informed conversations with agents to identify areas where additional support or guidance might help. Initial estimates show that this increased visibility can help focus coaching and training efforts more effectively, boosting overall performance by about 10–15 percent.

By implementing Voxflow, Insurancemarket can now act more quickly to resolve customer issues. We can examine a typical 10-minute call in about 2 minutes instead of the 20–25 minutes that a manual review required, says Gongolidis. When a customer leaves a call confused or dissatisfied, the system flags the conversation and initiates a follow-up by a more experienced agent with the context already available. Equipped with NRT response, Insurancemarket can explain complex requirements, address misunderstandings, and act on customer concerns before they develop into larger service issues. We moved from assumptions to a clear picture of what needs to be done with our customers and with our agents, says Gongolidis.