AI in HR: augment the human, do not automate the judgment.

Every HR vendor in 2026 is selling AI. Most of it is real, some of it is dangerous, and almost none of the marketing tells you which is which. After piloting more than a dozen tools across hiring, performance, and employee relations, here is the line I've drawn — backed by what regulators, researchers, and the courts are now saying out loud.

Where AI earns its keep

• Drafting first versions — JDs, offer notes, feedback summaries — that a human edits in minutes.
• Summarizing long-form input — exit interviews, engagement comments, intake calls — into themes.
• Surfacing patterns a human would miss across hundreds of data points.
• Coaching managers in the moment with structured prompts before a hard conversation.
• Translating policy and onboarding content across languages with reviewable diffs.

Where it must not go alone

Hiring decisions. Termination decisions. Promotion ranking. Pay bands. Anywhere a human's livelihood, dignity, or legal status is on the line, AI is an input — not an authority. The EU AI Act now classifies most of these use cases as 'high-risk', requiring documented human oversight, bias testing, and the right to explanation.

“An algorithm that cannot be questioned is not a decision support tool. It is a decision laundering tool.”

Three guardrails I won't negotiate on

• Explainability: any AI output that affects a person must be reproducible and reviewable.
• Consent: candidates and employees must know when AI is in the loop, and what it is doing.
• Override: a named human owns the final call. Always. Documented.

The bias trap

Models trained on the last decade of hiring data inherit the last decade of hiring biases. The Amazon resume-screening case (2018) and the iTutorGroup EEOC settlement (2023, $365K) are not anomalies — they are warnings. Treat every AI feature in your stack as guilty until proven fair, and audit the disparate-impact metrics quarterly, not annually.

Used well, AI gives a small HR team the leverage of a much bigger one. Used badly, it scales the worst instincts of the function faster than anyone can audit them. The difference is governance — set it before you scale, not after.