Most insurers know what a sample of their customers think. Quality teams listen to a handful of calls per agent each month, survey programmes report scores by quarter, and complaint teams see what customers chose to escalate. Everything else — the reasons customers call back, the claim that is approved but not paid, the policy wording nobody understands — sits in transcripts and verbatims that no one reads.
That was a constraint of human review. It no longer needs to be. In our work for a Fortune 500 specialty insurer protecting more than 300 million consumers, a GenAI pipeline analyses every call transcript in about 2.3 seconds, alongside complaints and survey verbatims in six languages1.
From sampling to full coverage
Full coverage changes what the business can ask. Instead of ‘what did the sample say?’, leaders can ask which claim types drive repeat calls this week, which agents resolve on the first call, and which regions are seeing new themes. The answers are only useful if the model output is structured. In the call pipeline, each transcript returns a category, sub-category and reason, sentiment with emotional context, a complaint level, any claims mentioned, an agent quality score and a summary — as JSON, ready for the warehouse and dashboards1.
What one call returns
Fields produced for each transcript in the call-intelligence pipeline
| Field | What it captures |
|---|---|
| Intent | Category, sub-category and reason |
| Sentiment | Overall score with emotional context |
| Complaint level | L0 critical to L3 low |
| Claims | Claim numbers and types mentioned |
| Quality | Five-stage agent assessment, intro to close |
| Summary | A short narrative of the call |
Note: From the specialty-insurer case study (scale of the build).
Source: DaasLabs, “Case study: GenAI customer intelligence for a Fortune 500 specialty insurer” (2026)
Privacy comes before the model
Calls and surveys are full of policy numbers, addresses, phone numbers and email addresses. The pipeline we built masks them before any text reaches the language model, using Microsoft Presidio for detection alongside spaCy entity recognition and context-aware checks12. That ordering matters: it is far easier to evidence to a privacy regulator that personal data never reached a model than to explain what the model did with it.
Language is culture, not just words
Across five Latin American markets, generic sentiment tools misread regional Spanish and Brazilian Portuguese: sarcasm read as satisfaction, idioms lost in translation. The fix was not a bigger model but regional prompt tuning, automatic language detection before processing and themes that work across markets1.