11. Industry Deep Dives
An Aliz data engineer walks into a telco engagement and a retail engagement in the same quarter. Both clients want "anonymized analytics." Both have BigQuery, Dataflow, and Cloud KMS. The four-zone architecture from Module 06 fits both. And yet the engagements feel completely different — because the data shapes, regulatory layers, failure modes, and the definition of "good enough" diverge the moment you open the first schema.
The preceding modules built a toolkit: techniques (Module 03), privacy models (Module 04), architecture patterns (Module 06), key management (Module 07), governance (Module 09), and operational machinery (Module 10). This chapter shows how that toolkit bends to fit four verticals and closes with the patterns that transfer across all of them.
The four verticals at a glance
Telco
The anchor vertical and the hardest anonymization problem. 4 spatio-temporal data points identify 95% of individuals in mobility datasets — identifier substitution alone is not enough. The ePrivacy Directive layers strict requirements on top of GDPR: traffic data must be erased or anonymized after transmission, location data requires explicit consent with no legitimate-interest fallback. Legacy dictionary-based pseudonymization — still common — offers no structural privacy and is vulnerable to dictionary attacks on finite-domain identifiers. The modernization path: three-layer architecture separating identifier pseudonymization, structural privacy, and access control.
Retail and e-commerce
The defining challenge is multi-domain identity: a single customer appears under 5+ identifier types, and identity resolution must happen in the raw zone before pseudonymization. Purchase sequences are quasi-identifiers — 4 spatiotemporal transaction points identify 90% of individuals. Privacy Sandbox APIs were retired in October 2025; regulatory consent pressure (not browser changes) now drives the shift to first-party data and clean rooms. PCI-DSS v4.0.1 tokenization provides prescriptive scope reduction for payment data. Consent management across touchpoints requires consent-aware views in the pseudonymized zone.
Financial services
The only vertical where regulation prohibits anonymization of core operational data. AML directives mandate 5–10 year retention of identifiable KYC records, transaction logs, and suspicious activity reports — premature anonymization is itself a compliance violation. The reconciliation: operational data stays identifiable in Zones 0–1; analytical derivatives flow to Zones 2–3. SWIFT's experimental federated learning trial (13 banks, Google Cloud, synthetic transaction data) showed 2x fraud detection effectiveness while keeping raw data local. Credit scoring faces dual compliance: GDPR Article 22 + EU AI Act both require explainability.
Energy and utilities
The youngest anonymization vertical. Smart meters generate enough data for Non-Intrusive Load Monitoring — 5 consecutive half-hourly readings re-identify over 90% of households. Germany's Smart Meter Gateway enforces data minimization at the hardware level: different granularities for different recipients. EV charging data adds a mobility dimension with CDR-like re-identification characteristics. Differential privacy is the natural fit — the primary analytical uses (grid planning, demand forecasting) all work on aggregate patterns.
Cross-vertical patterns
The four-zone architecture adapts to every vertical — what changes is zone retention, anchor identifiers, and the hardest problem. Three architecture decisions (pipeline vs. query-time, key architecture, DP budget) transfer directly. A comparison table and the transferable patterns are in the cross-vertical sub-module.
Next: Module 12 examines what happens when anonymization fails — real-world re-identification attacks, implementation mistakes, and the testing methodology that catches them.
Created by Claude Opus with ❤️ for the Aliz Data Engineering team.