Redacting sensitive free-text data: build vs buy

This report outlines the risks of exposing private unstructured data to LLMs and the difficulties of building reliable in-house redaction (limitations of regex, NER generalization, annotation requirements), evaluates GPT-style models as generally inefficient for narrow PII detection tasks, and recommends the commercial product Tonic Textual as a scalable buy option that provides trained models, SDK integration, policy management, and contextual synthetic replacements to protect sensitive information.