Anyone else noticed how broken enterprise AI + PII handling actually is?

We've been building an AI gateway for the past few months and hit a problem we didn't expect. Most enterprises we talked to either banned LLM tools completely or are quietly using them and hoping compliance doesn't notice. When we dug into why, it kept coming back to the same thing — they can't send raw customer or patient data to an external LLM, and the tools that claim to solve this only do half the job. They redact before sending. Fine. But the LLM response comes back with placeholders and now someone has to manually fix it before it's usable. A doctor's notes system, an HR tool, a finance report the output is broken without the original values. We spent a long time on this and built something that rehydrates the response on the way back. The data never leaves your infra in raw form but the output is still usable end-to-end. Still stress testing it. Found gaps. Fixing them. Curious if anyone here has actually run into this specific problem not the general "AI and data privacy" anxiety, but specifically the part where redaction breaks your workflow. What did you do about it?