How does an AI-based tool identify and redact PII from tax documents before processing?

One vendor described training their system on roughly 60,000 tax documents so it understands what counts as PII (name, social security number, address, and similar identifying details) versus what does not. They emphasized that data security is critical because a breach could destroy the business, drawing on a background selling enterprise bank software. Their core extraction and classification models are proprietary, built on top of Gemini and secured within their own environment rather than relying on public models. When they do use outside models like OpenAI or Anthropic, it's rare, mostly for operational (not tax/compliance) tasks, and they apply strict access controls since that work happens outside their own infrastructure. Their general advice: look at vendors who specialize in this, be careful about data security, and if you can't ensure PII isn't discoverable by an outside model, don't send it. They noted E&O insurance exposure as a serious risk if something goes wrong.