AI for HR: practical wins, legal landmines, and the bias problem.

HR is one of the highest-stakes areas for AI deployment. The work is sensitive. The legal scrutiny is intense. The bias risk is real. Done well, AI is a meaningful productivity lift. Done carelessly, AI in HR will create a discrimination lawsuit.

Where AI is helping HR teams right now

Job-description writing and editing. AI drafts job descriptions, suggests inclusive language, and flags potentially exclusionary phrasing. Used by recruiters as a first draft, then edited by a hiring manager.

Resume screening — narrow assist, never sole decision. AI can identify resumes that match objective criteria (years of experience, specific certifications, geographic eligibility). Use it to expand the pool by surfacing matches the keyword search missed, not to filter the pool by ranking and rejecting.

Drafting policy text, employee communications, and internal FAQs. Routine HR communications take less time to produce. Same review standard applies: HR owns the final text.

Summarizing exit interviews, engagement surveys, and pulse data. AI reads the open-ended text and produces themes. Quantitative dashboards on top.

Drafting performance-review feedback (manager-assisted only). AI reads the manager's notes and produces a structured draft. Manager edits and owns. The risk is managers shipping AI text without editing — train against this.

Where AI creates real legal exposure

Automated hiring decisions without human review. Multiple jurisdictions now require human review of any consequential employment decision. NYC Local Law 144 (effective July 2023) requires bias audits of automated employment decision tools. EU AI Act classifies employment AI as high-risk. EEOC has issued guidance. If your AI makes the call, you own the liability.

Resume-ranking tools that have learned bias from historical data. Amazon's well-known case (2018) — an internal resume scorer that downweighted resumes containing "women's" because the training data reflected historical male-dominant hiring. The tool worked exactly as built. The training data was the problem. This pattern recurs.

Sentiment analysis on employee monitoring data. Reading Slack messages, email, or call data to predict attrition or flag "disengagement" is legally and culturally a minefield. Some jurisdictions classify it as wiretapping. NLRB has weighed in.

Predictive performance ratings for compensation or termination. Same risk profile as automated hiring decisions, with the same legal exposure.

The questions to ask any HR AI vendor

1. What data was the model trained on? Specifically — does training data include protected-class signals? 2. Has the tool been bias-audited per NYC Local Law 144 (or analogous jurisdictional requirement)? Show the audit. 3. What is your data retention policy? Where does employee data live? 4. What is your training-on-our-data policy? (Should be "we do not.") 5. Who else can access this data? 6. What is the appeal path for an employee flagged by the system?

A vendor that cannot answer these confidently is not ready for HR data.

What to do this quarter

Inventory every AI tool your recruiters and HRBPs are using, including the ones bought "just for me" by individual managers. Build a one-page policy. Get legal to ack it. Train managers in 15-minute sessions on the difference between AI-assisted and AI-decided.

The LearnTrainAI curriculum dedicates a module to HR-specific AI deployment patterns and legal posture.