AI-Powered Quality Assurance for Insurance Customer Support
A multi-agent QA system that reviews support tickets against multi-page policies, with built-in PII redaction for GDPR.
~$1.5M
Annual cost reduction
~85%
QA manhours automated
Meet our client
Heymondo is a European travel insurance provider operating across multiple markets and support channels. As their customer base grew, so did the daily volume of support tickets requiring quality review under strict GDPR and regulatory constraints.
Context
Heymondo was scaling support across markets and channels faster than it could check quality. A 25-person QA team could review only about 5 percent of daily tickets, and every review had to respect strict GDPR and PII rules. They needed far more coverage without adding headcount or risk.
Challenge
The constraints were as much regulatory as operational:
- High daily ticket volumes across distributed support teams
- A 25-reviewer QA team able to evaluate only about 5 percent of tickets
- Manual reviews exposed to fatigue and inconsistency
- Complex, multi-page policies raising the risk of uneven evaluation
- Strict GDPR and PII requirements limiting how customer data could be processed
What we did
VentureSEA built an AI-powered QA platform, known internally as Mondoguardian, around a multi-agent evaluation framework, with privacy designed in from the first step.
- Processed resolved tickets in scheduled batch runs, reviewing full conversation history in context
- Embedded customised PII redaction with Microsoft Presidio, the open-source de-identification framework, directly in the pipeline, so sensitive data is removed before any model, log, or feature store ever sees it
- Aligned configurable redaction rules to GDPR and internal compliance policy, with the redaction layer treated as a tested, governed component, not a one-line import
- Designed the full pipeline as an auditable compliance architecture: every processing step logged, every model decision traceable, every redaction decision reviewable
- Used agentic reasoning with Gemini 3 to cross-check responses against multi-page policy documentation, applying consistent evaluation logic across every ticket
- Calibrated the evaluators against a blind set of human-reviewed tickets until model-human agreement passed 90 percent, and only then switched on automated scoring
Outcome and impact
- Review quality comparable to or better than manual review
- Privacy and compliance architecture designed in from day one, not layered on top
- A fully auditable QA system a regulator or auditor can follow end to end
Business value
Reaching 30% QA coverage with people would mean several times the current 25-person review team. In practice it removes the need to add an estimated 30 to 50 reviewers as volumes grow, worth roughly $1.2M to $2M a year in salary. The automated pipeline handles roughly 85 percent of the review manhours that would otherwise fall to human reviewers, redirecting the QA team onto the contested cases where their judgment matters most. The consistency gain also lowers the rate of uneven handling, where a single GDPR or regulatory misstep can cost more than the whole QA function.
More case studies
AI Suite Enablement for an HVAC BMS Platform
A modular AI suite for a building management platform: fault detection, remaining-useful-life prediction, and energy optimisation.
AI Agent-Powered Travel Concierge with End-to-End Booking
A hierarchical multi-agent travel concierge: conversational itinerary generation plus direct flight booking through a GDS.
Pricing, ERP and Data Architecture Modernisation
Redesigned pricing and data architecture for a last-mile fuel distributor: governed workflows, role-based approvals, and full auditability.
Have a project like this?
Tell us the outcome you need. We will map the path and the team to ship it.