PII Redaction

PII redaction is the removal or masking of personally identifiable information — names, email addresses, phone numbers, account identifiers, national ID and card numbers — from text before it reaches a model, a log, or a third-party service. In an AI stack it usually sits at one of three points, and which one a vendor offers changes what the control is actually worth. Client-side redaction strips the data before it leaves your network, so the vendor never receives it. Gateway redaction happens at a proxy you run in front of the model API, which is the common enterprise pattern because it applies uniformly to every application without each team implementing it. Vendor-side redaction happens after the provider has already received the raw text, which protects the logs but not the transmission — a distinction that matters if your obligation is about disclosure rather than storage. Redaction is a recall-versus-precision problem, not a solved one. Pattern matching catches structured identifiers like card numbers and emails reliably; free-form names, addresses, and indirect identifiers require a model or a named-entity recognizer, and both miss cases and over-redact. Over-redaction is the failure mode teams underestimate: strip too aggressively and the support ticket you sent to the model no longer contains the account context needed to answer it, so quality drops for reasons nobody attributes to the redaction layer. Practical questions for a vendor: what is redacted by default versus configurable, can you add custom patterns for your own identifier formats, is the mapping reversible so answers can be re-hydrated with the real values before display, and are redaction decisions themselves logged for audit. Redaction reduces exposure; it does not remove the need for a data processing agreement or retention terms.