Handbook · AI for Professionals

AI for HR.

AI will not replace your judgment about people, your read of a team's health, or your accountability for decisions that change someone's livelihood. It will absolutely replace the hours you spend writing job descriptions, drafting the same policy for the fifth time, building an interview question bank, and turning a pile of survey responses into a readable summary. This handbook is the practical middle ground: where AI genuinely helps HR, the three rules that keep it fair and legal, and five workflows you can use this week — with anti-discrimination and the no-automated-decision line front and center, because that's where the real risk lives.

~18 min readno coding5 workflow recipes1 judgment exercise
This is a guide to working with AI tools, written by an engineer — it is not legal or employment-law advice, and it is not a substitute for anti-discrimination law (EEOC, Title VII and equivalents), local AI-hiring regulations (e.g. NYC Local Law 144), data-privacy law, or your legal and compliance team. When they conflict with anything here, they win.
01

Where AI actually helps — and where it doesn't

The single most useful mental model: treat AI like a fast, tireless HR coordinator who drafts well, has no judgment about people, can quietly encode bias, and must never make a decision that affects someone's job. You'd happily hand that coordinator a job-description draft or a survey summary. You would never let it decide who gets hired, promoted, or let go.

TaskVerdictWhy
Job descriptions & postings (bias-checked)ExcellentFast drafts; you review for inclusive, non-discriminatory language
First-draft policies, handbooks, process docsExcellentStructure and starting points you verify against law and your org
Interview question sets & structured rubricsStrongConsistent, job-relevant questions — which improve fairness
Internal comms & announcement draftsStrongTone and clarity; you add facts and human sensitivity
Summarizing anonymized survey / feedback dataStrongThemes and signals from data you provide — you sanity-check
Screening, ranking, or scoring candidates — unauditedNever unauditedEncodes and scales bias; legal exposure — see Rules 1 & 2
Hire, fire, promote, or discipline decisionsNever automatedA human decides and is accountable — see Rule 1
02

The three non-negotiable rules

Rule 1 — AI never makes a consequential decision about a person. A human decides.

Hiring, firing, promotion, discipline, and pay are decisions that change people's lives, carry legal and ethical weight, and require context and accountability an AI cannot hold. AI may help you draft and organize, but it must never make — or effectively make, by producing a score or ranking you rubber-stamp — a consequential employment decision. A human evaluates, decides, and owns it. If a tool's output becomes the decision, you've handed a life-altering judgment to software that doesn't understand the person, the context, or the law.

Rule 2 — Actively guard against discrimination and bias. This is the big legal risk.

AI learns from historical data, and historical hiring data encodes historical bias — a well-known recruiting tool learned to penalize resumes containing "women's," and screening models can produce disparate impact on protected groups. The employer stays legally responsible regardless of the tool, and regulation is arriving fast: NYC Local Law 144 requires bias audits and disclosure for automated hiring tools, the EU AI Act classifies hiring AI as high-risk, and the EEOC has issued guidance. The discipline: any AI touching candidate selection must be validated, bias-audited, transparent, and human-reviewed — and job descriptions and rubrics get checked for biased or exclusionary language before use.

Rule 3 — Candidate and employee data is sensitive and protected.

HR holds some of the most sensitive data in the company — applications, salaries, health and accommodation information, performance and disciplinary records, complaints. Free and consumer AI tools may retain what you paste and train on it, and this data is protected by privacy and employment laws (and often consent obligations). The discipline: use enterprise tools with contractual no-training and retention commitments, anonymize wherever possible (summarize feedback without names), and never paste identifiable employee records — especially health, disciplinary, or complaint data — into a consumer chatbot.

03

Five workflows you can use this week

Each recipe: what to give the AI, what to ask, and what you must verify by hand. The prompts are starting points — keep the bias-check and data-privacy discipline.

1 · Job description (inclusive & bias-checked)

  1. Give the role, must-haves, and an explicit inclusivity instruction:
Write a job description for [role]. Requirements: [genuine must-haves only]. Use inclusive, neutral language — avoid gendered terms, "culture fit," age-coded phrases ("digital native," "recent grad"), and anything that could deter or screen out protected groups. Focus on the actual work and required skills. Then flag any requirement that might not be truly essential.
You verify: review for biased or exclusionary language yourself (Rule 2) — the model can still slip it in — and confirm every "requirement" is job-related and consistent with business necessity, not an unintended barrier.

2 · Structured interview questions and rubric

  1. Give the role and the competencies you're assessing:
Create a structured interview guide for [role], assessing [these job-related competencies]. For each competency: 2 behavioral questions and a simple 1–4 scoring rubric with anchored descriptions. Keep every question job-relevant and avoid anything touching protected characteristics (age, family, health, religion, origin).
You verify: no question strays into protected territory (Rule 2), and remember the rubric supports a human's judgment — the interviewer scores and decides, the AI just helped structure a fairer, more consistent process.

3 · First-draft policy or handbook section

  1. Give the topic, your org context, and jurisdiction:
Draft a [remote-work / PTO] policy section for a [size, industry] company in [jurisdiction]. Clear, plain language, employee-friendly tone. Cover [the points I list]. Note where a specific rule depends on local law that I need to confirm, rather than stating a legal requirement as fact.
You verify: every legal or compliance statement against actual law and with your legal team (the model states requirements confidently and sometimes wrongly), and adapt it to your real org — a policy is a legal document, not just prose.

4 · Summarize anonymized survey or feedback data

  1. Provide the data with names and identifiers removed (Rule 3):
Summarize this anonymized [engagement survey / exit interview] data: the main themes, strongest positive and negative signals, and 3 areas worth acting on. Quote only text present in the data — do not invent responses. Note where the sample is too small to draw conclusions, especially for any subgroup.
You verify: every "quote" is real (the model can fabricate representative comments), the data was truly anonymized before pasting, and small-subgroup conclusions aren't over-read — thin data can imply bias signals that aren't real, in either direction.

5 · Draft internal communications

  1. Give the message, audience, and tone — no sensitive individual data:
Draft an internal announcement about [a policy change / org update]. Audience: all employees. Tone: clear, warm, transparent — acknowledge impact honestly without over-promising. Under 250 words. Leave placeholders for specifics, effective dates, and the contact for questions.
You verify: you add the real specifics, read it for tone and sensitivity (people react to how change is communicated), and route anything about specific individuals through private, appropriate channels — never a chatbot.
04

The judgment exercise: spot the danger

Three scenarios from real practice patterns. Pick what you'd do — the point is calibrating when the rules bite.

1. A vendor's AI tool scores 400 applicants and hands you a ranked shortlist of 20. It'll save days. Do you interview the top 20 and reject the rest?

2. To get a second opinion on a termination, you paste an employee's full performance file — with complaints and a medical note — into a free chatbot. OK?

3. The AI-drafted job posting asks for a "young, energetic digital native who's a great culture fit." It reads upbeat. Post it?

05

Choosing tools: the questions that matter

You don't need to understand the technology to procure it well. You (or your legal team) need answers, in writing, to five questions:

QuestionAnswer you want
Is our candidate/employee data used to train your models?No, contractually — not "you can opt out somewhere in settings"
For any selection tool: is it bias-audited and validated?Yes, with documentation you can rely on and disclose
How long is our data retained, and can it be deleted?Defined, short, deletable on request
Compliance with hiring-AI and privacy law?Yes, with support for your audit/disclosure obligations
Security posture and access controls?SOC 2 Type II or equivalent, encryption, admin/audit controls

General-purpose enterprise AI (Claude, ChatGPT's business tiers) covers the drafting and summarizing recipes here — none of which make decisions about people. Recruiting/HRIS AI that screens or scores candidates is a different, higher-risk category requiring bias audits, validation, transparency, and legal review before deployment. And if your org has an AI or hiring policy, it precedes everything here; if it doesn't, these three rules are a reasonable seed for one.

06

Quick answers

Can I use AI to screen resumes?

Only with a validated, bias-audited, transparent, legally-reviewed tool and a human in the loop — never as an unchecked filter. Employers are liable for disparate impact regardless of the tool, and regulation (NYC LL144, EU AI Act) is tightening.

Can AI write a performance review?

It can draft or structure feedback from documented notes you provide, but the assessment and any decision are the manager's — AI must not score, rank, or decide who is managed out.

What HR data can go into AI tools?

Keep identifiable employee/candidate data — especially health, disciplinary, and complaint data — out of consumer tools. Use enterprise tools with no-training terms and anonymize wherever possible.

Where should a skeptical HR pro start?

Recipe 1 — drafting an inclusive job description, then reviewing it for biased language. Immediately useful, and it builds the exact muscle (spotting bias in AI output) the rest of the work depends on.

Finished this one? 0 / 99 Handbooks done

Explore the topic

See this alongside everything else on the same subject — handbooks, system designs, challenges and tools, in one place.