How AI Can Flag Inconsistent Discipline Without Making the Decision
Every vendor in this category will tell you their AI "doesn't make decisions." Almost none of them will tell you exactly where that boundary lives - which prompt, which function, which line of code prevents the model from crossing it. That vagueness is usually the tell that the boundary is a talking point, not an engineering constraint.
Why this distinction is a legal question, not a UX preference
If an AI system concludes that a policy was violated, or effectively recommends a disciplinary action, that output starts to look like it's making an employment decision - and employment decisions made or substantially influenced by an opaque model are a genuine legal exposure, independent of whether the model happened to be right. The moment a tool crosses from "here is relevant information" to "here is what you should do," the human in the loop stops being a decision-maker and starts being a rubber stamp, whether anyone intended that or not.
Where the line actually sits in Rectifia
The scoring engine that processes a new case is explicitly instructed - in the system prompt that governs it, not just in how the feature is described - to treat company policy documents as reference material only. It's told directly never to conclude that a policy was violated. It produces a severity score and an evidence score, kept separate on purpose, and it can flag gaps ("this allegation lacks a specific date" or "no witnesses have been named yet") - but it does not, and structurally cannot, output "this violates section 4.2 of the harassment policy."
The Consistency & Bias Checking Engine, which compares a proposed action against how similar closed cases were handled, follows the same rule from a different angle. It surfaces a deviation: "five comparable cases exist, the typical action was a written warning, the proposed action is termination." It does not say which action is correct. A human investigator sees the comparison and makes the call, and that call - not the flag - is what gets documented as the actual decision.
What this looks like when it's wrong
Suppose the engine flags a proposed termination as "harsher than the typical pattern" and the investigator proceeds with the termination anyway, with good reason - maybe this case has an aggravating factor the historical pattern didn't capture. That's fine. The system isn't broken when a human overrides it; it's working as designed. The flag did its job by surfacing the deviation for conscious review. What would actually be broken is a system that either silently suppressed the flag, or treated an override as something requiring justification back to the AI - at that point it's stopped being an advisory tool and started being a gatekeeper.
The honest limit of "AI never decides"
This boundary protects against a specific failure mode - AI substituting its judgment for a human's on questions of guilt or consequence. It does not mean the AI has no influence at all. A severity score does shape which case gets attention first. A flagged evidence gap does shape what question gets asked next. Those are real effects, and pretending an AI system that touches a workflow has zero influence on outcomes would be its own kind of dishonesty. The claim isn't "the AI has no effect." It's narrower and more defensible: "the AI never concludes what happened or what should be done about it." That's the boundary worth asking any vendor to show you in their actual prompts, not just their marketing page.