ChatGPT for HR Professionals: Job Ads, Reviews and Warning Letters

An HR professional can use ChatGPT to draft job ads, performance-review wording and warning letters without exposing staff data by pasting only anonymised inputs (role labels instead of names, behaviours instead of identifiers), using a business plan or training opt-out, and sending any warning letter for a legal review before it is issued.

The hard line in a small team is simple: ChatGPT never receives anything that identifies a person together with something sensitive. It helps you find words. It does not decide who deserves a warning, what a rating should be, or whether a dismissal is fair. Those judgements stay with you and the manager, and anything with legal weight is checked by someone qualified.

Follow me on Instagram@sagnikteaches

The staff-data rules to set before the first prompt

In a team of 15 to 60 people there is often no data-protection specialist, so the HR person writes the rules. Here are the rules, in the order of importance.

Connect on LinkedInSagnik Bhattacharya

1. Anonymise at the source. Replace names with role labels ("Employee A, warehouse operative"). Do not paste dates of birth, home addresses, payroll numbers, bank details, health information, grievance details or union membership.
2. Describe behaviour, not the person. "Arrived more than 15 minutes late on six of ten shifts between 1 and 28 October" is a usable input. "[Name] is always late and unreliable" is neither anonymous nor factual.
3. Use the right account. A company ChatGPT Business, Enterprise or Edu workspace is excluded from model training by default, according to OpenAI's business-data page. A personal account on Free, Go, Plus or Pro shares content for training until you opt out, so an HR person on one of them clears the "Improve the model for everyone" setting in Data controls before pasting anything about work.
4. Keep a log. Note the date, the document and the person who reviewed it. If a letter is challenged, you can show how it was produced and checked.
5. Never paste the finished letter with the real name back in. Add the name and details in your own document after the draft is final.

Check your own employment contracts and any works-council or union arrangements. They may limit how personal data is processed, and some employers require staff to be told when automated tools are used. If you have customers or staff in the EU, data-protection law such as the GDPR affects what you can paste and where it goes; ask a qualified adviser what applies to you.

Subscribe on YouTube@sagnikteaches

Job ads: from a vague request to a clear, fair advert

Job ads are the safest starting point because they contain no staff data at all. The risk here is bias and inflated requirements, not leakage.

A tutoring centre's part-time administrator advert

Say a tutoring centre with 12 staff needs a part-time centre administrator, 24 hours a week. The manager's brief is a messy paragraph: "need someone organised who can handle parents, the timetable, payments, and ideally some social media, good with computers, young and energetic would be ideal". Run this prompt (illustrative):

Write a job advert for a part-time centre administrator (24 hours a week, afternoons and Saturday mornings) at a small tutoring centre.
Use these facts only: parent enquiries by phone and email; managing a weekly timetable of about 60 sessions; taking and recording payments; light social media posting; salary range [add your range]; start date [date].
Structure: a 2-sentence summary; "What you will do" (6 bullets); "What we are looking for" (5 bullets, only things the job genuinely needs, split into essential and helpful); "What we offer"; how to apply.
Avoid age-coded, gender-coded or culture-coded language. Do not invent benefits. Under 350 words.

Notice what is missing: "young and energetic". That phrase can exclude people by age and has no connection to the job. The model, given the instruction to avoid age-coded language, will produce "organised, friendly and reliable" instead.

An illustrative first draft of the opening and requirements might read:

Part-time centre administrator, 24 hours a week. You will be the first voice parents hear and the person who keeps about 60 weekly sessions running on time. Essential: experience of handling customer enquiries by phone and email; confidence keeping a timetable or rota accurate; care with recording payments. Helpful: some experience of posting on a business social media account; familiarity with a booking or timetabling system.

Two things to fix before it goes live. Check the draft has not upgraded "light social media posting" into "managing our social media presence", which describes a bigger job and puts off the people you want. And confirm the hours line matches the contract exactly, including whether Saturday mornings are every week or on a rota, because a mismatch here is the first thing a new starter raises.

What to check in the draft

Illustrative problems in a first draft, and the fixes:

• "Fast-paced environment" and "digital native" appear. Remove both; replace with "comfortable using a timetable system and email".
• A benefit you never mentioned ("generous holiday") has crept in. Delete anything you did not supply.
• Essential requirements have grown to nine. Cut to those without which the person could not do the job. Long lists discourage some applicants more than others, and they filter out people who could do the work.
• No salary. Add your range: stating pay lets candidates judge fit before they apply, and some jurisdictions now require employers to give applicants the pay or pay range before interview.

Then ask a second question: "Read this advert as a 55-year-old returning to work, and as a 19-year-old with no experience. What would put each off?" The answer is a useful prompt for editing, not a legal test.

Screening questions and interview scorecards

The same approach works for the next steps. Ask for five application questions that test the essential requirements, and a scorecard with a 1 to 4 scale and written behavioural descriptions for each level. A scorecard (illustrative) for "handles parent queries" might describe 1 as "unable to give an example", 4 as "describes calming a difficult call and recording the outcome". Use the same scorecard for every candidate and have two interviewers score separately. ChatGPT drafts the criteria; the panel decides who is hired.

Performance reviews: shaping evidence into fair wording

Review writing suffers from vague praise, recency bias and awkward criticism. ChatGPT helps most when you give it evidence and ask for structure and balance. See also this tutorial on fairer performance reviews, which covers the process around the writing.

Example: from bullet evidence to a review section

Say a manager supplies, anonymised: "Employee B, yoga studio front-of-house. Evidence: handled 3 class cancellations in one week without complaints; booking errors fell from about 8 a month to 2 after the new checklist; two late arrivals in the quarter; asked to cover Saturday classes 4 times and agreed each time." Prompt:

Turn this evidence into a review section for a front-of-house team member. Use only the evidence listed. Write three short paragraphs: strengths, areas to improve, and goals for next quarter. State each point with its evidence. Do not use words like "always", "never", "attitude" or "potential". Do not rate or score. Neutral, specific, kind.

A good output states "booking errors fell from about 8 to 2 a month after the checklist was introduced" and "two late arrivals this quarter; we agreed on a 15-minute window and will review". The manager adds judgement and the rating. The "do not use" list keeps out vague loaded words that make reviews unfair.

Example: balancing a review written in a hurry

Paste a manager's own draft (anonymised) and ask for an audit: "List every sentence that lacks evidence, uses a personality label, refers to the last month only, or differs in tone from the other sections. Quote each." Managers often write four detailed paragraphs on the last week and one line on the previous eleven months. The audit highlights that, and the manager goes back to the records.

Example: a difficult conversation script

For a review meeting where the news is hard, ask for three opening lines that name the issue, give the evidence and invite the employee's view. Practising these with the manager takes five minutes and reduces defensive reactions. No personal data is needed: describe the situation in general terms.

Warning letters: the careful path

This is the highest-risk use. A warning letter can end up before a tribunal or court, so the stakes for accuracy are high. ChatGPT can help structure and phrase a letter; it must never be the source of facts or the decision-maker, and a legal review step is not optional.

Inputs, anonymised

Gather the facts from your records: the policy breached (by its number or title), dates and times, what was said or done, any previous informal conversations and their dates, the employee's response in the meeting, the sanction proposed and the right of appeal. Remove the employee's name, replacing it with "[Employee]". Write facts as observed and recorded, not as impressions.

Draft a first written warning letter using only the facts below. Use a professional, calm, neutral tone. Include: the allegation or conduct issue stated factually with dates; reference to the policy; summary of the meeting and the employee's response; the outcome; what improvement is expected and by when; how long the warning stays on file; the right to appeal and to whom. Placeholders in square brackets for anything I have not supplied. Do not add facts, do not speculate about motives, do not quote law.

FACTS: [paste anonymised facts]

What to fix in the output

Illustrative issues and fixes: the draft says the employee "deliberately ignored" the rule, although your notes say only that they did not follow it. Delete "deliberately". It states a timeline for improvement ("within 30 days") that you did not supply. Replace with the one your policy specifies. It cites a legal obligation in a confident tone. Remove every legal claim. The model's version may omit the appeal right or the length of time the warning remains live, both of which are usually required by your procedure.

The legal-review step

Before issuing, send the final draft and your procedure notes to someone qualified: an employment solicitor, your HR adviser or an employer-organisation helpline that your business has access to. Ask three specific questions: does the letter follow our procedure, are the facts supported by the record, and does the sanction match our policy? Keep their reply on file. A small team will pay for this by the letter or through a membership; it costs less than a mistake. This tutorial is general guidance and not legal advice, and rules differ by place, so your adviser decides what applies to you.

Finally, put the real name and details back into your own document, not into ChatGPT, and have the manager read the letter aloud before the meeting to check it sounds like something a person would say.

Other HR drafts that suit the same method

Once the three main documents are set up, several routine items follow the same pattern.

• Offer letters: a template with placeholders, never a pasted candidate record.
• Probation review forms: a form with structured prompts and evidence boxes.
• Return-to-work conversation notes: a generic template; never paste health information.
• Policy summaries: paste the policy text (not staff data) and ask for a one-page plain-English version for the handbook.
• Staff FAQs: turn policy text into questions and answers; see this tutorial on an AI helpdesk.

For a wider view of what else is safe to hand over, 12 HR tasks you can hand over safely gives a broader list.

Example: rewriting a job ad that is already live

Often the task is not a blank page but an old advert that attracts the wrong applicants. Say a landscaping company has run the same ad for years: "Strong lads wanted for busy season, must have own transport, experience essential, long hours." Paste it and ask: "Audit this advert. List every phrase that could exclude a group of people without being necessary for the job, every requirement that might not be essential, and every important fact that is missing (pay, hours, location type, training). Then rewrite it." The audit returns "lads" (gendered), "strong" (vague; replace with the actual lifting requirement, such as "able to lift 20 kg repeatedly"), "experience essential" (could training cover it?) and no pay range. The rewrite gives the same job to a wider pool without lowering standards, because the physical requirement is now stated instead of implied.

Example: a rejection email that is kind and consistent

Candidate communication is an HR task that takes less judgement and more care. Ask for three versions of a rejection email for applicants who were not shortlisted, interviewed once, and interviewed twice, each under 90 words, with placeholders for the name and role. The post-interview versions should offer brief, honest feedback only if you decide to give it, and must not mention other candidates. Keep one template per stage so that every applicant is treated alike, which helps if a decision is later questioned. Do not paste candidates' CVs; if you want feedback phrasing for one person, describe the gap in a line ("limited experience with the timetable software") and nothing else.

Example: onboarding checklist for the first 30 days

A new-starter checklist contains no staff data, so it is a safe early win. Prompt: "Create a 30-day onboarding plan for a part-time centre administrator in a 12-person tutoring centre. Week-by-week, with who is responsible, what the new starter should be able to do by the end of each week, and three check-in questions for the manager." The first draft will be generic. Make it specific by adding your own systems (the timetable tool, the payments app), your safeguarding or health-and-safety induction, and the names of roles rather than people. Keep the finished plan in your shared drive and update it after each hire.

A prompt library for HR that holds the rules

Once your prompts work, store them in a ChatGPT Project (a shared workspace that keeps instructions and files in one place). Put the rules in the Project instructions so every chat starts under them: "Never ask for names or personal identifiers. Use role labels. Use only the facts provided. Never give legal advice or cite law. Mark gaps with [NEED]." Add a short file of your organisation's tone and your standard letter structure. OpenAI lists project sharing on every plan, but sharing with colleagues is only appropriate where each person is covered by the same rules. Review the Project instructions every quarter and after any incident.

One more safeguard: add a line at the top of every output template saying "DRAFT: not for issue until reviewed". A watermark like this sounds trivial, but it prevents a common accident, which is a draft with a placeholder still in it going to an employee. Before issuing any document, search it for square brackets and the word NEED.

Who is responsible when the draft is wrong

A model's draft carries no responsibility. The HR professional who issues the letter, and the employer, carry it. That is why the checking steps above are not bureaucracy: they are the point at which a person takes ownership. If an employee later disputes a review or a warning, "ChatGPT wrote it" is not a defence, and it may weaken your position if it suggests nobody read the document carefully. Keep the facts in your own records, the reasoning in your own notes, and the model in the role of a drafting assistant.

If you are working through how far to use AI across the people side of the business, an outside view can save mistakes. That is what my AI implementation consultation covers, and you can reach me through the contact page.

A simple review table for every HR document

The last row is the line: for dismissals and redundancy, use professional help rather than a model draft. The cost of error is large and the facts are usually sensitive.

A quarterly audit of HR drafts and saved prompts

Three checks keep the arrangement honest. Quarterly, take five documents at random and trace every fact to a record. Check the log shows who reviewed each letter. Ask managers whether the review drafts reflect the whole period rather than the last month. If you find a real name, a date of birth or a health detail in a saved prompt, delete the chat, tell whoever needs to know in line with your policy, and tighten the anonymising step. On a business plan the admin settings help, but the habit of removing identifying details matters more than any setting.

Expect modest time savings. As a rough illustration, a job advert that took an hour of fretting might take 15 minutes plus a check, and a review section 10 minutes instead of 30 once the evidence is gathered, which you were doing anyway. Warning letters save the least time, because the checking step is the point. For a deeper look at software built for the job, see this comparison of AI HR software; and if you advise several small employers, how HR consultants use AI shows a similar approach at consultancy scale.

Questions HR professionals ask about ChatGPT and staff data

Can I paste an employee's review notes into ChatGPT?

Only anonymised, and only if your employer's policy allows it. Replace the name with a role label, remove dates of birth, health information and anything that could identify the person to a reader outside the company. When in doubt, summarise the behaviour yourself and paste that.

Is a warning letter drafted by AI valid?

Validity depends on your employment contract, your process and local employment law, not on who typed the draft. A letter must be accurate, follow your procedure and be reviewed by someone qualified before it is issued. Treat the AI output as a first draft only.

Will ChatGPT write biased job ads?

It can. Models reproduce patterns from their training text, including gendered or age-coded wording. Run every ad through a checklist of exclusionary phrases and ask a colleague to read it, then compare the requirements to what the job needs.

Should I tell employees that AI helped draft a review?

Check your policy and any rules where you operate. A safe habit is to be open that tools may help with drafting while the manager and HR remain responsible for every judgement in the review.

Further reads

• AI Consultant Red Flags: 12 Warning Signs to Walk Away From — Warning signs when you hire outside help with AI.
• Is ChatGPT Available for Commercial Use? What Owners Should Check — What owners should check on ChatGPT terms.
• AI Words to Avoid: A Banned-Phrase List for Your Brand — A banned-phrase list that also helps job-ad wording.
• Restaurant Job Descriptions: Can AI Write a Server or Line Cook Ad? — ChatGPT can draft a server or line cook ad in a minute. Here is what a hiring owner must check on shifts, pay, tips and wording before posting.
• ChatGPT and Copilot for HR: Which Suits a Small People Team? — A four-question rule, seat costs for a three-person people team, and the employee-data handling that matters more than which assistant you choose.
• Free AI for HR Professionals: Are Free Plans Safe With Staff Data? — Free AI plans can draft policies, adverts and interview guides, but staff records are another matter: data settings, anonymising steps and when to pay.
• AI Tools for HR Professionals: Policies, Onboarding and Staff Questions — How a small-business HR lead can use AI for policy drafts, onboarding checklists and a staff question assistant, with data-protection and legal review built in.
• HR Analytics and AI: Analyse Staff Turnover From a Spreadsheet Export — A privacy-first spreadsheet method for checking turnover and absence, with a worked rate and practical prompts.
• HR Automation AI Tools: What a Small Business Can Automate Safely — A task-by-task boundary for small-business HR automation, with review gates and practical examples.
• Can AI Draft Disciplinary Letters? What HR Consultants Must Check — A process-first way to draft invitation, outcome and appeal letters without inventing evidence or decisions.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business data privacy page, ChatGPT pricing page, OpenAI help-centre data controls and projects articles, EU Directive 2023/970 on pay transparency (checked 4 November 2026).