Insurance Data — Governance & Analytics
Four years inside a Belgian insurance group — two as a data steward on data quality, governance and GDPR compliance, two as an analyst on data extraction, KPI data preparation and process automation. Currently coordinating a process consolidation project across fifteen people. Specifics anonymised; methodology and outcomes described in detail.
2022 — 2024 · Data Steward
The data steward years were spent on the cases that didn't fit. Records that wouldn't merge across systems. Fields that disagreed with each other. Edge cases the model handed back with low confidence. The work was unglamorous and slow, and it taught me more about how insurance data actually behaves than any course could have.
What I learned: data quality is not a problem you fix once. It's a continuous translation between what the business says it wants, what the systems actually capture, and what humans actually do. AI handles the patterns. What stays on a human desk is the texture — and the texture is where the real decisions live.
2024 — present · Data Analyst
The analyst years turn the better-quality data the stewardship years produced into business decisions: pulling customer data out of Salesforce, preparing the datasets behind the KPIs, and SAS extraction of the information behind commission and mass payment runs — work where a wrong field does not produce a wrong chart, it produces a wrong payment. I also automated the team's email workflow in Python, replacing a repetitive manual process.
Current · Coordinating a process consolidation project
My main work now is pulling procedures that were never standardised into one simple workflow, and preparing the data that feeds the KPIs behind it. The coordination runs across fifteen people in different functions, which is most of the difficulty: the technical part is defining a field once, and the real part is getting fifteen people who each had a working local definition to accept a single one.
A KPI is only as good as the agreement underneath it. If two departments count the same thing differently, the dashboard is not reporting reality, it is reporting the disagreement. Most of what I do is making that agreement explicit and then holding it.
Selected case studies
The following case studies are described at the methodology level, with company-specific details and identifying information anonymized in line with confidentiality obligations.
Case 1 · Improving merge accuracy on legacy customer records
A category of customer records had been routinely flagged as uncertain merges by the deduplication pipeline. Manual review revealed a pattern the model had not been trained on. By documenting the pattern and adjusting the human-review queue, merge accuracy on this category improved meaningfully and the review burden dropped.
Methodology — pattern identification through manual review, documentation, queue restructuring.
Case 2 · Reducing false-positive escalations
A specific class of claims was being escalated to human review at a higher-than-expected rate. Analysis showed the model was treating routine variations in input as anomalies. Refined input pre-processing reduced false-positive escalations substantially without affecting true-positive detection.
Methodology — escalation pattern analysis, input normalization, controlled rollout.
Case 3 · Cross-system reconciliation for regulatory reporting
A reporting requirement involved reconciling fields across three internal systems whose definitions had drifted over time. The reconciliation pipeline was rebuilt with explicit mapping documentation, reducing manual reconciliation work and producing a clearer audit trail for the regulator.
Methodology — system-by-system field audit, definition reconciliation, documented mapping.
What stays on a human desk is the texture — the cases that don't fit. That texture is where most real decisions live.
What this work taught me
- The interesting cases are the ones AI hands back, not the ones it confidently classifies.
- Good data quality looks like a non-event in production. It's noticeable only when it fails.
- Domain knowledge compounds. Four years of working with insurance data is qualitatively different from four months — the patterns you've seen before are what let you spot the new ones.
- Documentation is not optional. The cleanest systems are the ones whose definitions are reconciled and written down.
For role-specific questions about methodology, tools, or anonymized examples, I'm happy to discuss in interview.