Human Resources, Management and Training

HR and management are full of writing tasks that AI can speed up: job ads, onboarding guides, training materials, feedback preparation. They are also a field where mistakes affect people's lives, careers and rights. In this lesson you will separate the safe uses from the sensitive ones, learn to prepare without delegating judgment, and understand why using AI to sort or evaluate people calls for great care.

This lesson follows Nadia, the HR officer, and also speaks to any manager who writes feedback, trains a team or recruits. The key message is clear from the start: AI can help you write about work; it should not decide about people.

Nadia's busy week

This week, Nadia must publish two job ads, update the welcome booklet, prepare a short training session on a new software tool and help a manager write a difficult piece of feedback. At the same time, her director asks whether the company could use AI to rank the 120 applications received for a single position. The same assistant could help with every one of these tasks, but not every task has the same level of risk. Let us look at each.

Safe and useful: job ads, onboarding, training

Job ads

A good job ad is clear about the role, realistic about requirements and welcoming. Give the assistant the real content: the mission, the main tasks, the requirements that are truly necessary, the team, the conditions you can state. Then ask for a clear, structured ad and check it for two things: accuracy (no promise you cannot keep) and inclusive wording (avoid unnecessary criteria and wording that could discourage certain groups). Remember that rules on discrimination in recruitment apply to ads, and that some information, such as salary ranges, may be subject to transparency rules in your country.

A ready-to-use prompt: inclusive job ad

Write a job ad for the position of [job title] in [type of organization, location, work arrangement].
Use only this information:
- Mission: [mission]
- Main tasks: [3 to 6 tasks]
- Essential requirements (really necessary): [list]
- Nice to have (not required): [list]
- What we offer: [only what is true and confirmed]
Structure: 2-sentence introduction, tasks, profile, what we offer, how to apply. Tone: [e.g. clear and welcoming]. Length: about 300 words.
Use plain, inclusive language: avoid gendered wording where the language allows, unnecessary jargon, superlatives and requirements that are not essential. Do not invent benefits or salary details. At the end, list any phrase that could be read as excluding some candidates.

Onboarding and training materials

Onboarding guides, welcome emails, FAQs and training outlines are high-value, low-risk. Nadia gives the assistant her real process (the checklist from chapter 3) and asks for a friendly welcome booklet. For a training session, she specifies the audience, the learning objectives, the time available and the level, and asks for an outline, exercises and a short quiz. She then checks the content for accuracy and adapts examples to her company.

For training, a useful pattern is to ask for learning objectives first ("at the end, participants can..."), then activities that support them, then ways to check understanding. An assistant can also generate realistic practice scenarios (with fictional people and companies).

A ready-to-use prompt: a 45-minute training session

Design a 45-minute training session on [topic] for [audience, level, number of people].
Learning objectives: at the end, participants can [3 measurable objectives].
Constraints: [room or online, tools available, no preparation required from participants].
Provide: a minute-by-minute outline, one short exercise with a fictional scenario, 5 quiz questions with answers, and a one-page handout.
Do not make claims about regulations or technical details that I have not provided; mark items I must verify with [VERIFY].

Feedback and difficult conversations

Many managers struggle to put feedback into words. An assistant can help you structure it (situation, behavior, impact, request), propose neutral wording and prepare for the conversation. But feedback is a human act: the person receives it from you, with the facts you observed.

• Describe observed behaviors, not personality ("missed three deadlines this month", not "is unreliable").
• Remove names and sensitive details from the prompt: use "an employee" and the role.
• Use the output as a draft or a rehearsal, never as a script to read aloud.
• Never use AI to write a disciplinary document or decide on a sanction; this requires procedures, human judgment and often legal advice.

The sensitive part: sorting and evaluating people

Back to the question from Nadia's director: can AI rank the 120 applications? The honest answer is: she should not hand it over, and she should be very cautious about any use that screens, ranks, evaluates or monitors people.

Why this is risky

• Bias: models learn from human-produced data, which contains historical biases. A tool can reproduce discrimination (based on gender, origin, age, disability and others) in ways that are hard to detect, even when those characteristics are never mentioned, because other details (a school, a hobby, a career gap) can act as proxies.
• Lack of explainability: if a person is rejected, can you explain why? Can the person challenge the decision?
• Equity: a score creates an appearance of objectivity that makes it harder to question.
• Data protection: CVs and evaluations are personal data. The GDPR sets principles (lawfulness, transparency, data minimization, purpose limitation) and gives people rights, including specific protections where decisions are based solely on automated processing and have significant effects.
• Regulation: the European regulation on artificial intelligence (the "AI Act") treats certain AI systems used in employment and worker management, such as recruitment, selection, evaluation, promotion or termination decisions, as sensitive "high-risk" uses subject to strict obligations, and some practices in the workplace are restricted or prohibited. The rules apply in stages, so check the official texts and the dates that apply to you.

This is general, informative content, not legal advice. For official sources, see the CNIL (the French data protection authority) and its guidance on AI and recruitment, and EUR-Lex for the text of the GDPR and the AI Act, and consult your organization's legal or data protection contact before using any AI tool on candidate or employee data.

What you can do instead

• Use AI to prepare the process: write clear selection criteria based on the job (not on a tool's suggestion), build a structured interview guide with the same questions for everyone, draft a fair evaluation grid.
• Keep people reading applications and deciding, with documented criteria.
• Run swap tests on your own prompts: change a name or a gender in a sample text and see if the output changes in tone.
• Involve your HR and data protection contacts when considering a recruitment tool.

Bad use versus good use

Bad use	Good use
Uploading 120 CVs to a chatbot and asking for the top 10.	Defining job-based criteria, using a structured grid, and having people review all applications.
Writing a performance review from a prompt such as "Write a review for Alex: average performance".	Writing your own observations with examples, then using the assistant to structure and neutralize the wording.
Using a generic job ad copied from a chatbot, including benefits you do not offer.	Using verified information and an inclusion check.

Your turn: sort the HR tasks

Exercise. For each of these tasks, decide whether AI use is (A) reasonable with review, (B) possible only with strong safeguards and approval, or (C) not appropriate, and explain why: (1) drafting a welcome email for a new employee; (2) summarizing a candidate's CV for the hiring manager; (3) creating a quiz for a safety training; (4) ranking employees by "potential" using their performance data; (5) rewriting a manager's feedback to make the wording more neutral.

See a solution

1. A. Low risk, no personal data beyond a first name, easy to check.
2. B or C. It processes personal data in a recruitment context and influences a decision. It requires an approved tool, a legal basis, information for candidates and human review of the original CV; many organizations will choose not to do it.
3. A, with validation. The content must be checked against the real safety rules by a competent person.
4. C. Evaluating and ranking employees is a sensitive use with risks of bias, unfairness and legal obligations. Not delegated to AI.
5. A, with anonymization. Remove names and details; the manager keeps the facts and the responsibility.