Employee relations takes nothing
Most of the never-list is about what the text contains. The ER rule is stricter: it's about where the work lives, and the answer is nowhere near a prompt.
Conflicts, investigations, grievances, complaints against a named person — this cluster is human-only, start to finish. Not "AI drafts, human reviews." Not "AI summarizes the notes to save an hour." Nothing. Three reasons, each sufficient on its own:
- The material is maximally protected. Investigation records touch medical facts, allegations against named people, and statements given in confidence. This is also the material most likely to end up in litigation, where "who saw this and when" gets asked under oath. "A vendor's API, briefly, for summarization" is an answer nobody wants to give.
- The loopholes are fake. "Anonymize it first, then summarize" — the raw material entered the prompt to be anonymized; the breach already happened. "It's just for structure, not content" — the content went along for the ride. Every workaround someone proposes on a deadline has the same shape: the protected material crosses the wire, and the justification travels separately.
- The judgment is the job. An investigation summary isn't a text-compression task. What to include, what weight a statement carries, what pattern the incidents form — those calls are the employee-relations function. Draft them with a model and you haven't saved time on the work; you've outsourced the work and kept the liability.
This is why the gate you're about to write checks provenance, not just phrases. A payload sourced from er-notes blocks even if its text happens to scan clean — because the phrase list can't anticipate every way ER material gets described, but the source label is right there. Phrase scan catches leaks; source check catches categories. ER needs both, plus a team that doesn't argue with the gate.
One more boundary worth naming while we're drawing hard lines: everything in this chapter keeps AI on the drafting side of employment decisions. The moment a tool scores, ranks, or rejects a candidate or employee, you've left drafting and entered regulated territory — NYC's Local Law 144 has enforced independent bias audits and 10-business-day candidate notice for automated employment decision tools since July 2023, at $500 to $1,500 per day for getting it wrong. Lesson two walks that line properly. For now: the firewall governs what goes into prompts; the decision boundary governs what comes out. You need both walls, and neither lives in memory.