This one carries real legal weight
Employment decisions are regulated in most jurisdictions, and several have specific rules about automated decision-making in hiring, including notice and audit obligations. This is not a category where you can experiment and adjust afterwards.
Bias is inherited from your own history
A system learning from past hiring decisions learns whatever patterns were in them, including ones you would not defend. The output looks objective because it is numeric, which makes it more persuasive and no more correct.
Automating a biased process makes it faster, more consistent, and much harder to argue with.
What it can do safely
Administrative work: acknowledging applications, scheduling interviews, chasing missing documents, answering candidates’ questions about the process. That is operational chasing applied to recruitment, and it improves candidate experience measurably.
Structured extraction, with care
Pulling stated qualifications into a consistent format is defensible when it extracts rather than infers, and when a human reviews. Extraction is checkable; inference presented as a score is not.
What it should not do
Rank, score, or reject candidates. Rejection is an irreversible decision about a person, which puts it squarely in the category an AI employee may never do alone.
Whatever you do, be able to explain it
Candidates may ask, and regulators increasingly can too. Record what was used, what it produced, and which human decided — and take advice for your jurisdiction before deploying anything that touches the decision itself.