AXA Direct Assurance boosts new customer growth 33% and cuts model update time 30% with MLOps automation

AXA Direct Assurance, the French online insurance subsidiary of AXA Group, relies on machine-learning models to calculate competitive pricing for home insurance products. In P&C insurance, pricing accuracy is existential: models calibrated even slightly below market risk erode margins, while over-pricing drives customers to competitors. The data science team was responsible for maintaining multiple pricing models that required frequent retraining to reflect shifts in market and behavioral trends. The existing process was largely manual — long iterative cycles in Jupyter notebooks, error-prone handovers between team members, and no systematic versioning or reproducibility. The result was slow responsiveness to market changes and compounding technical debt that constrained the team's capacity for higher-value work.

Xebia consultants embedded with Direct Assurance's technical direction team to implement MLOps best practices on the Databricks platform, targeting three areas: code quality, execution automation, and experiment tracking. Jupyter notebooks were replaced with structured, testable Python pipelines governed by a CI/CD pipeline that enforces unit tests, integration tests, and linting. The full retraining workflow — from data extraction through production deployment — was automated as Databricks Jobs running on dedicated job clusters, with pipeline definitions stored as versioned JSON for reproducibility and disaster recovery. MLflow was integrated for experiment tracking, data versioning, and model registry management: each job receives a unique identifier propagated across data sources, MLflow experiments, and registered model versions, enabling the team to compare model performance systematically and promote the best candidate to production with confidence.

The MLOps overhaul delivered measurable gains across both business and operational dimensions. Faster model retraining cycles allow AXA Direct Assurance to refresh pricing more frequently, improving competitiveness in the French home insurance market.

• +33% projected increase in new customer growth rate vs. the previous system
• −30% reduction in model update time for senior data scientists
• −28% reduction in model update time for junior data scientists

Beyond the headline numbers, the standardized pipeline reduced handover friction across the technical direction teams and eliminated the error-identification cycles that previously consumed hours per retraining run. The reproducible, traceable workflow positions the team to extend the same MLOps foundation to other insurance product lines ahead of AXA's broader 2026 automation target.

• Notebook-to-pipeline migration is a forcing function for quality: moving from ad-hoc notebooks to testable Python with CI/CD catches errors earlier and reduces the knowledge-transfer burden when team composition changes.
• Versioned pipeline definitions pay dividends beyond reproducibility: storing Databricks Job configurations as JSON enables environment replication and rapid recovery, not just auditability.
• MLflow model registry creates a decision layer, not just a log: linking data versions, experiment runs, and production models in one registry lets teams make promotion decisions based on evidence rather than convention.
• Productivity gains compound across seniority levels: both junior and senior roles improved at similar rates, suggesting process bottlenecks — not skill gaps — were the primary drag.
• Baseline business-impact measurement should be scoped at project start: Direct Assurance explicitly modeled expected customer growth to quantify ROI, giving leadership a concrete case for broader rollout.