What deterministic software approach can be used to strip PII from taxpayer data before sending it to an external LLM?

One firm's policy requires that any taxpayer data sent to an external LLM first pass through a piece of deterministic software—described as essentially a pre-programmed script (as opposed to a nondeterministic AI agent)—that extracts everything except PII. For example, this might involve an OCR read that scrapes the document data and then pulls out name and Social Security number, after which the de-identified data is served to an application that feeds it into the LLM. The output of deterministic software always follows the same fixed rubric, unlike an AI agent given a general instruction to 'solve this problem.'