Deciding a harassment complaint is exactly the kind of call AI must never make alone.

Every harassment complaint involves unique facts, relationships, power dynamics, and contextual nuance. There is no repeatable structure that an agent can reliably apply — each case demands fresh, situated judgment.

Success criteria are deeply contested and context-dependent: what constitutes harassment, what evidence is sufficient, and what outcome is fair are all judgment calls with no crisp, computable answer. The task cannot be verified as 'done correctly' without human review.

Critical inputs — witness credibility, tone of voice, body language, organizational history, unwritten cultural context — are largely inaccessible to an agent. Even if documents are provided, the agent lacks the full evidentiary picture needed to decide fairly.

A wrong outcome can destroy careers, expose the organization to serious legal liability, cause lasting psychological harm to the parties involved, and set dangerous precedents. These consequences are largely irreversible and affect real people's lives.

This task demands ethical accountability, empathy, credibility assessment, and legal reasoning that current AI cannot reliably provide. Beyond capability, there is a fundamental accountability requirement: a human must own this decision and be answerable for it.