Before the Prompt: A Fail-Closed Architecture for PII Protection and Business Privacy Policy Enforcement in Enterprise AI

This technical white paper introduces Pleco, a locally deployable privacy protection solution designed to help organisations reduce the exposure of personally identifiable information (PII) when using external AI assistants. The paper explores the growing privacy challenges associated with enterprise AI adoption and presents Pleco's approach to identifying sensitive information, validating detected text spans, and replacing PII with realistic synthetic data while preserving the original message's meaning and context. It explains Pleco's technical architecture, including its fine-tuned language model, Ollama-based local inference, deterministic sanitisation process, synthetic alias management, and policy knowledge ingestion approach. The paper also discusses the fail-closed design philosophy, human review, model evaluation, deployment considerations, security boundaries, current limitations, and potential future enhancements. The publication aims to contribute to discussions around privacy engineering, responsible enterprise AI adoption, and protecting personal information before it enters AI systems.