Key takeaways
Ask anyone who has run a review cycle in a small business where the three weeks went and the answer is not writing reviews. It is scheduling them, chasing self-assessments, collecting peer input, formatting inconsistent documents into something comparable, and filing the results.
All of that is coordination. Automating it does not touch a single judgment and typically returns the cycle to something that fits in a week.
A manager who has kept notes through the year has the content of a review already; what they lack is an hour to turn it into readable prose. Handing those notes to a drafting tool and editing the output is a legitimate use, and it usually produces a better review than the one written at eleven at night.
The line is whether the assessment existed before the draft. If the manager formed a view and the tool wrote it down, that is assistance. If the tool formed a view from activity data and the manager signed it, that is something else, and the manager will find out how different when they have to explain it.
A rating produced by a model reflects whatever the model learned, and the shape of that risk is familiar: a factor that looks neutral can have an adverse effect on a group identified by a prohibited ground, which is discrimination under the human rights codes whether or not anyone intended it. A rating you cannot explain is a rating you cannot defend.
There is a second exposure in Quebec, where Law 25 requires an organisation to tell someone when a decision about them was based exclusively on automated processing and to let them have it reviewed by a person. An employer whose ratings are genuinely human decisions, informed by data, is not in that territory. One that has quietly outsourced the judgment is.
Progressive discipline and termination are legal decisions. They carry notice and termination-pay obligations under provincial employment standards, they can be challenged, and the quality of the record matters enormously when they are.
Automation belongs on the record and the process: making sure the documentation exists, is consistent, and is filed. It does not belong on the decision, and no system we build produces one. If a tool ever offers to, that is the moment to close the tab and call an employment lawyer instead.
A 30-minute call is enough to tell you whether AI pays for itself here.
