You don't need a frontier model to redact PII

Short summary

A detailed benchmark compares six PII redaction approaches across cost and accuracy: Amazon Nova Micro matches Comprehend at 1/20th the cost, Nova Pro ties a 4GB local model at 94% German recall, and a fine-tuned German model performed worst. The article provides decision frameworks based on data residency, volume, and existing subscriptions, recommending frontier models for low-volume work and cheaper models for pipeline-scale redaction.

• •Amazon Nova Micro ties Comprehend at ~1/20th cost per document for PII redaction
• •Nova Pro matches a local mistral:7b model at 94% German exact-value recall
• •Decision framework splits on data residency, volume, and existing tooling subscriptions

Generated with AI, which can make mistakes.