Intact Financial Corporation

Intact Financial Corporation, Canada's leading property and casualty insurer, operates contact centers fielding up to 20,000 customer support calls per day. In P&C insurance, contact center quality directly affects claims outcomes, customer retention, and regulatory standing — making consistent call auditing a business-critical function, not a back-office formality. Yet Intact's quality control process was entirely manual: auditors hand-selected individual calls to review, meaning only a small fraction of interactions were ever examined. At that volume, systemic coaching gaps, script deviations, and early signals of customer dissatisfaction went largely undetected. The company set itself a hard internal deadline — a workable solution within six months, a production-grade platform within one year.

Intact built an automated Call Quality (CQ) platform on AWS, with Amazon Transcribe as the speech-to-text foundation for both English and Canadian French — a bilingual requirement central to serving the Canadian market. Audio recordings from on-premises and cloud contact centers are ingested automatically via Amazon EventBridge, which triggers an AWS Step Functions workflow when files land in Amazon S3. Transcripts are indexed in Amazon OpenSearch Service and enriched by custom NLP models running on Amazon Fargate and EC2. These models perform named entity recognition, sentiment analysis, speaker role identification, PII redaction, script adherence scoring, reason-for-call classification, and call outcome tagging. A self-service MLOps pipeline — built on Step Functions, Lambda, and S3 — enables data scientists to conduct shadow deployments and push new models independently, compressing release cycles from multiple days to mere hours.

The CQ platform delivered a 1,500% increase in both auditing speed and total calls reviewed — equivalent to 15 times more calls per auditor without adding headcount. MLOps improvements drove a 65% gain in auditor efficiency by eliminating manual model deployment overhead. Targeted coaching from the platform's analytics reduced agent handle time by 10% and average hold time by 10%. Qualitative outcomes compounded the headline numbers:

• Near-zero downtime since the platform's 2020 launch
• New model releases now deploy in under an hour, down from multiple days
• Near-zero deployment failure rate due to the platform's modular, decoupled architecture
• Auditors shifted time from call selection to coaching strategy and script refinement

• Manual sampling at P&C scale is structurally blind: only full-coverage automation can surface the coaching gaps, script failures, and sentiment patterns hidden in the long tail of calls.
• Bilingual ASR accuracy is a non-negotiable requirement for Canadian insurers — validate speech-to-text performance in both English and French before committing to a platform.
• Building an MLOps pipeline in parallel with the core solution pays compounding dividends; cutting model deployment from days to hours directly multiplied auditor throughput.
• Modular, event-driven architecture enables near-zero-downtime releases — design components to deploy independently from day one to avoid costly release freezes as the platform scales.