HR & People

headcount request → job description → sourcing list → screening notes → interview packet → scorecards → offer letter

• Draft
  Job description drafting — AI converts a hiring manager's bullet list into a leveled, bias-checked JD and generates the interview plan from it · highly verifiable against the role brief, low stakes until posted, human owns final language
• Automate
  Candidate sourcing — AI builds boolean searches, drafts personalized outreach sequences, and enriches prospect lists · high volume, easily reversible, each message is cheap to spot-check
• Assist
  Resume screening — AI summarizes and structures applications against the JD requirements, but does not rank, score, or reject candidates · this is where AI-hiring law bites — automated screening triggers NYC LL144 bias audits and EU AI Act high-risk obligations, so keep the model as a reading aid, not a filter
• Draft
  Interviewing & scorecards — AI transcribes interviews and drafts structured scorecard entries from what was actually said; interviewers verify and own the ratings · verifiable against the transcript, but the rating itself is an employment decision with protected-class exposure
• Avoid
  Candidate evaluation & selection — the hire/no-hire call from the assembled evidence · low verifiability, high stakes, irreversible for the candidate, and the single most regulated decision in this playbook
• Automate
  Offer & close — AI drafts the offer letter from comp bands and generates candidate-specific closing materials · template-driven, fully verifiable against the approved comp band, recruiter spot-checks

offer acceptance → onboarding plan → provisioning checklist → 30/60/90 plan → first-week schedule → ramp check-ins

• Automate
  Provisioning & paperwork — agent triggers accounts, equipment, payroll enrollment, and compliance forms from the signed offer · pure rules-following, every step verifiable, high repetition across hires
• Draft
  Role-specific 30/60/90 plan — AI drafts the ramp plan from the JD, team docs, and manager input; the manager edits and owns it · context is available in writing, stakes are moderate, and the manager's edit is where accountability lives
• Automate
  First-week schedule & intro sequencing — AI assembles the calendar of intros, trainings, and setup blocks · scheduling logistics with instantly visible errors
• Automate
  New-hire Q&A — an AI assistant grounded in the handbook and internal wiki answers "how do I..." questions on demand · answers cite sources so they're verifiable, volume is high, and wrong answers surface fast; route benefits-legal questions to a human
• Assist
  Ramp check-ins & early signals — reading how the new hire is actually doing and intervening · AI can summarize check-in notes, but this is a relationship read with real stakes if a struggling hire is mishandled

goals → running feedback log → self-review → peer feedback → manager review → calibration deck → rating & message

• Automate
  Evidence assembly — AI compiles each person's shipped work, goal progress, and logged feedback into a pre-review evidence file · pure retrieval against existing records, fully checkable, painful and repetitive by hand
• Draft
  Self-review drafting — employees use AI to turn their evidence file into a first-draft narrative · verifiable against the evidence, and the employee owns the claims
• Draft
  Peer feedback synthesis — AI clusters and summarizes peer input into themes with supporting quotes · summarization is checkable against sources, but tone and emphasis need human judgment
• Draft
  Manager review writing — AI drafts the review narrative from evidence and the manager's rating rationale, never the reverse · the rating is the manager's; a review the manager can't defend in their own words is a legal and trust liability
• Avoid
  Calibration & rating decisions — comparing people and setting ratings across a team · low verifiability, high stakes, disparate-impact exposure if a model's patterns leak into ratings; AI's only safe role is formatting the calibration deck
• Assist
  Delivery conversation — the actual performance conversation · AI can help a manager rehearse hard messages, but the conversation is pure relationship work

policy question or complaint → intake record → investigation file → findings memo → outcome decision → policy update

• Draft
  Policy drafting & maintenance — AI drafts policy updates against new legislation and flags internal contradictions across the handbook · verifiable against statute and existing policy, but employment counsel owns final text
• Automate
  Tier-1 policy Q&A — grounded assistant answers routine "what's the policy on X" questions with citations · high volume, source-linked answers, cheap to audit; anything touching a complaint escalates to a human
• Draft
  ER case intake & documentation — AI structures intake notes into a consistent case record and timeline · verifiability is high for the record itself, but wording in an investigation file is discoverable in litigation, so a human owns every line
• Assist
  Investigation analysis — assembling interview notes into a chronology and mapping claims to evidence · AI can organize, but credibility assessment is relationships-not-rules and the stakes include wrongful-termination suits
• Avoid
  Outcome decision & disciplinary action — what happens to the people involved · maximally high stakes, irreversible, protected-class exposure on every dimension of the rubric

market data → comp bands → headcount plan → merit matrix → manager proposals → approved changes → comp letters

• Draft
  Market benchmarking — AI reconciles survey data across sources and drafts band updates with flagged outliers · numbers are verifiable but source-mapping judgment calls need a comp analyst's ownership
• Draft
  Headcount scenario modeling — AI builds and stress-tests workforce cost scenarios from the planning model · arithmetic is fully verifiable; assumptions are the human's job
• Assist
  Pay-equity analysis — AI runs the statistical passes and drafts the findings memo · the analysis is verifiable but the memo is legally sensitive — many companies run it under privilege, which a casual AI workflow can destroy
• Avoid
  Merit-increase recommendations — proposed raises per person · individual pay decisions carry direct disparate-impact liability; AI-suggested raises that correlate with a protected class are a lawsuit with a paper trail
• Automate
  Comp communication — AI generates individualized comp letters and manager talking points from approved numbers · mail-merge with narrative, fully verifiable against the approved file

skills-gap analysis → curriculum plan → course content → delivery → assessments → completion analytics

• Draft
  Skills-gap analysis — AI mines performance themes, support tickets, and manager surveys into a ranked gap list · synthesis is checkable against sources; deciding what the business actually needs trained is strategy
• Automate
  Course content production — AI drafts modules, quizzes, scenarios, and video scripts from SME outlines · content is cheap to regenerate, errors are visible and low-stakes, and volume is the whole bottleneck; SMEs spot-check
• Automate
  Localization & role-adaptation — one course becomes versions per region, role, and language · mechanical transformation with verifiable fidelity to the source
• Draft
  Compliance training — content and tracking for legally mandated training · content must match current law exactly and completion records get audited, so legal reviews every claim
• Assist
  Coaching & practice — AI role-plays difficult conversations, sales calls, and interviews with learners · genuinely effective as a sparring partner, but it supplements rather than replaces manager coaching
• Assist
  Effectiveness measurement — AI correlates training completion with performance outcomes · the correlation math is easy; the causal story it tempts you to tell is usually wrong