Free AI for HR Professionals: Are Free Plans Safe With Staff Data?

No, not for identifiable staff data. ChatGPT Free, Claude Free and Gemini are consumer services that can keep your chats and, unless you change settings, may use them to improve models or show samples to reviewers. Free tools can still draft job adverts, policies, interview guides, onboarding plans and survey questions, because those jobs need no personal data at all.

Whether a person can be identified decides everything else. A name is only the obvious part: a job title in a six-person team, an absence date and a manager's name together identify someone as surely as a payroll number. That is why "I removed the name" is not the same as "I made it anonymous", and it is the reason the safest free use of AI in HR is the work that never touches a person's record.

Follow me on Instagram@sagnikteaches

What each free plan does with a pasted staff record

These are the vendors' own statements, read on 1 November 2026. Labels and defaults change, so open the setting itself before you rely on a name from this table. The pages behind it are OpenAI's data controls FAQ, Anthropic's article on model training and Google's Gemini Apps Privacy Hub.

Connect on LinkedInSagnik Bhattacharya

Read the retention column as carefully as the training column. Switching training off doesn't make the vendor forget the chat: a pasted grievance still sits on a server for weeks, and that copy is something you would have to account for if a member of staff asked where their data went. Ads add a second reason to stay away from a consumer chat window in HR. OpenAI says advertisers get only aggregated performance data and that Temporary Chats show no ads, but a screen with adverts on it is not what you want visible in a one-to-one meeting.

Subscribe on YouTube@sagnikteaches

Four bands of HR data and where each may go

Sorting the work first saves arguing about each prompt. Every HR task fits one of four bands, and the band, not the tool, sets the rule.

If you employ people in the EU, the GDPR treats the categories in band 4 as special category data with a higher bar for processing, and it defines personal data broadly: anything relating to an identifiable person, which includes details that identify someone only when combined. Other data-protection regimes use different names for similar ideas. The practical result is the same everywhere: the more identifiable and sensitive the record, the less it belongs in a consumer chat.

Two common slips sit on the borders between bands. Band 2 material becomes band 3 the moment you add a date, a team name or a job title that only one person holds. And a band 1 prompt becomes band 3 when someone pastes the real exceptions log "for context". Put the band on every prompt you save in the team's prompt library, so the next person sees it before they paste.

HR work free tools handle well: a hybrid policy, start to finish

Take an illustrative 30-person professional services company with no written hybrid-working policy. The people lead budgets four hours to draft one from scratch. With a free assistant the first draft takes minutes, and the rest of the time goes where it belongs: on decisions only the company can make. No staff data is involved, so this is band 1 work and a consumer plan is fine.

The prompt carries the structure and leaves every company-specific fact as a placeholder.

Draft a hybrid working policy for a 30-person professional services
company. Cover: eligibility, expected office days (two a week), core
hours (10:00 to 15:00), how to ask for a change, equipment and
expenses, health and safety at home, and a review date.
Write in plain English at a reading age of about 14.
Use [square brackets] for anything company-specific, such as the
expense limit and the name of the people lead.
Do not invent legal obligations. End with a list of points that need
a legal or adviser check.

Expect eight short headed sections, a table of which roles qualify and a closing list titled "Points to check" (an illustration of the shape, not a recorded output). The closing list earns its place. It will usually flag that the policy must match each contract, that home-working equipment and expenses can have tax or insurance consequences, and that roles tied to a fixed location need a different arrangement.

What you fix is predictable. The tool may add a rule you never asked for, such as a minimum service period before anyone can work from home. Delete it unless you want it. It may write core hours as a rule for everyone, which may not suit shift workers. It can make confident-sounding commitments about health and safety that belong to your adviser. And it leaves placeholders, so search the draft for "[" before it leaves your desk. With the checking, the illustrative budget drops from four hours to about two: roughly one hour with the tool and one hour on the points only you can settle.

The same pattern works for the other band 1 tasks. Each needs a prompt that names the structure and a person who owns the decisions.

• Job advert. Paste the role profile (duties, essential skills, working pattern, a salary range if you publish one) and ask for an advert with essential and desirable requirements separated, no more than six essentials and inclusive wording. Then test each essential: is it really essential, or a habit from the last hire? More on wording sits in ChatGPT for HR professionals: job ads, reviews and warning letters.
• Interview scorecard. Ask for five competencies, the same four questions per competency for every candidate, and a 1 to 5 scale with a described example at each point. The scale must exist before anyone is interviewed, which is the fairness point. The tool drafts; the panel agrees the anchors.
• Onboarding plan. Give the role type and ask for a first-week schedule, a 30-60-90 day goal list and a checklist of accounts and equipment. It is generic by design: add real names and dates in your own document afterwards.
• Survey wording. Ask for twelve questions on workload, management and growth with a neutral five-point scale. Free tools are good at spotting a leading question. Keep the results, which are staff data, away from the tool unless they are aggregated.
• Plain-English rewrite of a policy paragraph. Company-wide policy text carries no personal data. Ask for a rewrite at a lower reading age, then compare with the original for drift in meaning, which is where legal weight hides.

Anonymising a staff case: a before-and-after you can copy

Band 2 work is where free tools earn their keep and where most leaks happen, so the method matters more than the intention. Take an illustrative note an HR lead might jot after talking to a line manager. In its raw form it holds the employee's name, job title and team, a start date, the date absence began, the named condition, the manager's name, the project that slipped, an occupational health suggestion with a return date, and a remark that the employee is the only part-time person in the team.

Every element in that note either identifies the person or adds sensitivity, and most add nothing to the task. Here is how each one is handled.

The re-identification test closes the loop. If a colleague read the cleaned note, could they say who it is about? Here the answer should be no. If it is yes, cut another detail. Then write the prompt, asking for something useful from the cleaned version.

Employee A has been absent for several weeks. Manager B is thinking
about a formal capability process. The HR lead wants a supportive
route first. Draft:
1. An agenda for a welfare conversation (30 minutes).
2. Six open questions that avoid assumptions about the cause.
3. A list of points to check with an adviser before any formal
   process starts.
Do not suggest diagnoses or outcomes.

Picture the result as a five-part agenda (how are you, what support would help, what you need to know about the role, practical adjustments to discuss, next contact), six neutral questions and a checklist reminding you to review the sickness absence policy, adjustments, documentation and timing. Nothing in it depends on the condition, which is the point: a diagnosis adds no value to a meeting agenda, so leaving it out costs nothing.

Edits needed: the tool may propose a timeline for the capability process that doesn't match your procedure, and it may use wording about "performance" that pre-judges the case. Rewrite those parts. Then do the last step offline. Open your own document, replace Employee A and Manager B with the real names, and add the dates. Keep the key (who A and B are) in your HR system, never in the chat. More prompts written this way, for probation and absence conversations, sit in AI prompts for HR managers: probation, absence and policy updates.

Pseudonymised is not anonymous: the small-team trap

Swapping names for labels is pseudonymisation. Under the GDPR, pseudonymised data is still personal data if a key linking labels to people exists anywhere, so the label hides someone from the tool, not from the law. Anonymisation, in the sense that takes data out of scope, means nobody can identify the person even by combining details, and that is hard to reach with HR records because HR groups are small.

Small groups are where it fails. Suppose a team has four people and a survey comment says "since the new desk layout I can't concentrate on calls". If only one person takes many calls, the comment names them. The same holds for pay bands, absence counts and tenure: a "department average" over three people is three people's data. My working rule, which you can adjust with your adviser, is a minimum group of five before any breakdown goes into a tool or a report, and paraphrasing comments instead of pasting them word for word.

Hidden information travels with files too. Before you paste text copied from a document, check what came along.

• Word and Excel properties. Author, last-modified-by and company fields, tracked changes and comments, hidden sheets and rows.
• File names and headers. "Warning_surname_March.docx" is a staff record in the filename alone.
• Screenshots and scans. Names in a corner, a badge in a photo, an email address in a signature.
• Email chains. The quoted history at the bottom holds names and dates that the top message doesn't.

Paste plain text you have typed or cleaned yourself, not whole files. A cleaned paragraph is easy to check; a forwarded email thread isn't.

Jobs to keep out of free chatbots altogether

Some HR work isn't a matter of anonymising better. It is work where an AI tool shouldn't be in the chain on a consumer plan at all, or where the law adds conditions on top of the vendor's terms.

• Ranking or filtering candidates. CVs are personal data, and ranking them is a decision about people.
• Scoring staff performance or summarising monitoring data.
• Grievance, disciplinary and whistleblowing evidence. Witness statements hold names, allegations and sometimes special category data.
• Fit notes, adjustment records and occupational health reports. Health data.
• Pay, bonus and equity data per person. One spreadsheet row identifies someone.
• Immigration, background-check and vetting results.

The EU's AI Act lists several of these as high-risk uses. Its Annex III covers AI intended to recruit or select people, "in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates". It also covers AI used to make decisions on the terms of work relationships, on promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance and behaviour. The implementation timeline on artificialintelligenceact.eu, last updated on 31 August 2026, shows 2 December 2027 as the date from which the high-risk obligations for Annex III systems start to apply. Dates have moved before, so confirm the current position with an adviser if you employ people in the EU, and treat the classification as a signal of how seriously regulators view these uses whether or not it binds you.

The GDPR adds a separate protection: a person has the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects on them. A shortlist produced by a chatbot and acted on without a person genuinely weighing it is the situation that rule has in mind. The safe design keeps a human in the loop in fact, not just on paper: the tool may draft or summarise, and a named person decides and can explain why.

Settings that help on a free plan, and what they leave open

Settings don't make a consumer plan safe for band 3 data, but they make band 2 work safer. Do them once, on the account you will use, and check the label because vendors rename things.

1. ChatGPT. Open Settings, then Data controls, and switch off "Improve the model for everyone". Use Temporary Chat for anything near the border, choosing "Unpersonalized" when it asks, because a personalised temporary chat can still draw on saved memories. It can also be kept up to 30 days for safety. Check the memory setting and switch it off for an HR account, so a detail from one chat can't surface in another.
2. Claude. In Privacy settings, switch off Model Improvement. Use Incognito chats for anything near the border. Remember that chats flagged under the usage policy can be kept for up to 2 years.
3. Gemini. Turn off Keep Activity or shorten auto-delete to 3 months, and use Temporary Chat. Chats that were reviewed stay for up to three years and aren't removed when you delete activity.
4. Copilot Chat. Sign in with the work account, not a personal Microsoft account, and ask your admin whether web search is on. Microsoft says web queries are generated from your prompt as a few words and sent with user and tenant identifiers removed, but the Bing service has separate data-handling terms, so keep personal details out of prompts even here.

What the settings leave open: the vendor's own safety and legal retention, the browser (extensions can read pages), shared devices and screens left logged in, and the habit of pasting. A written rule on the team wall addresses the habit, and the tutorial on stopping staff pasting client data into free tools covers how to make it stick.

When a work account is cheaper than the worry: a sum for two seats

If your organisation already has Microsoft 365 or Google Workspace, ask IT before buying anything. Microsoft says Copilot Chat is included at no extra cost with eligible Microsoft 365 plans, with enterprise data protection for work accounts. Google says Gemini in Workspace and the Gemini app aren't reviewed by humans or used to train models outside your domain without permission. That can turn "free and unsafe" into "free and covered" for band 2 and some band 3 work, subject to your admin's settings and your own policy.

If you have neither, the smallest paid step is two seats. ChatGPT Business Standard lists at $20 per user per month billed annually or $25 monthly, for teams of 2 to 200, with no training on business data by default. Claude Team Standard is $20 annual or $25 monthly for 2 to 150 seats, with training none by default. For an HR lead and a coordinator that is $40 a month at the annual price ($480 a year) or $50 monthly ($600 a year).

Set that against time. Suppose the two of them save six hours a month between them on drafting and summarising, valued at $35 an hour: $210 a month, about five times the $40 seat cost. The sum is illustrative, but its structure is what carries over to your own numbers. A work plan is paid for by a few hours, so the real question isn't price but whether the vendor's data terms and your own policy fit band 3 work. Read the plan's data processing terms before you move any band 3 work onto it, and keep band 4 out regardless.

Eight rules for the team wall

Write these on one page and ask every manager who touches people data to read them. Adapt the wording to your policies.

1. Free consumer AI accounts are for band 1 and band 2 work only.
2. No names, no job titles that identify one person, no dates and no health information in any prompt on a free plan.
3. Switch off model-improvement settings and use Temporary or Incognito chats for band 2 work.
4. Never upload whole files. Paste text you have cleaned yourself.
5. Candidate ranking, performance scoring and monitoring don't go through a chatbot.
6. Every AI draft is checked by a named person before it reaches an employee.
7. Keep the label key (Employee A is...) in the HR system, never in a chat.
8. If in doubt, treat it as band 3 and ask.

A rule sheet works because it ends the debate before it starts. If you'd rather build yours with someone and map which of your HR tasks sit in which band, that is what my AI implementation consultation covers. For policy wording to attach to the sheet, an AI policy for HR covering hiring, staff data and reviews is the natural next read, and ChatGPT and Copilot for a small people team compares the two tools most HR teams end up choosing between.

Free AI and staff data: HR questions

Is Temporary Chat or Incognito enough for staff data?

No. Temporary Chat in ChatGPT isn't used to improve models but may be kept up to 30 days for safety, Claude's Incognito chats aren't used for improvement but flagged chats can still be retained, and Gemini keeps Temporary Chats for 72 hours. These modes reduce training risk, not the fact that a copy of staff data sits with a vendor. Anonymise first, then use them.

Can free AI write a warning letter?

It can draft the structure, with placeholders such as [employee], [date] and [policy clause] in place of facts. It cannot know your procedure, the contract terms or what actually happened, and a letter that misstates a process can undermine a disciplinary case. Fill the facts in yourself, offline, and have the final text checked by whoever owns employment advice for the business.

Do I have to tell staff if I use AI on their data?

If personal data about staff goes into a tool, the notice you give staff about how their data is handled should cover the vendor and the purpose, and a free consumer plan makes that hard to defend. Using only anonymised or no data avoids the question for most tasks. For anything else, ask a qualified adviser what your notice must say.

What about AI note-takers in HR meetings?

A transcript of a grievance or disciplinary meeting is a record full of personal and sometimes special category data, held by a vendor. You need the consent of everyone present, a plan whose data terms you have checked, and a policy on retention. Free plans add practical limits too: Otter Basic stops each conversation at 30 minutes, which cuts off a long hearing.

Further reads

• AI Tools for HR Professionals: Policies, Onboarding and Staff Questions — Policies, onboarding and staff questions with AI tools.
• Best AI for Human Resources: Can It Draft a Grievance Report? — Whether AI can safely help draft a grievance report.
• AI Survey Analysis for HR Consultants: Coding Employee Comments Safely — Coding employee survey comments safely.
• Free AI Tool for Recruitment: What a Small Agency Gets for Nothing — What free AI plans and notetakers cover for a small recruitment agency, where free limits and candidate-data risks bite, and when paying makes sense.
• AI Salary Benchmarking for HR Consultants: Sources, Bands and Checks — A repeatable method for turning survey and job-advert data into pay bands with AI, plus the checks that catch invented figures.
• 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 Audit Checklist: Run a Client Audit With AI and Check the Gaps — A consultant's document intake, gap matrix and sampling method for an AI-assisted HR audit.
• 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.
• HR Consultant Proposal Template: Scope, Fees and Exclusions With AI — Copyable scope and fee clauses, plus a filled proposal for a 25-person client and a scope-change check.
• HR Consulting AI: Sell an AI Policy and Handbook Update Package — A defined consultant offer for AI policy and handbook updates, including a filled fee example and client handover.
• Free AI for Tax Professionals: Where the Free Tier Stops at Client Data — Which tax-practice jobs suit a free AI tier, which never should touch return information, and what a paid business plan changes and does not change.
• AI HR Software for Startups: Which Tool Fits Which Job — Which AI HR software suits your startup stage? Compare Rippling, Deel, and Gusto across onboarding, policies, scheduling, and payroll admin.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help articles on data controls and ads, Anthropic privacy centre articles and pricing page, Google Gemini Apps Privacy Hub and Workspace AI privacy page, Microsoft Learn enterprise data protection and Copilot Chat pages, Otter pricing, the EU AI Act Annex III text and implementation timeline, read on 1 November 2026 and checked again on 4 November 2026; GDPR Articles 9 and 22 and Recital 26.