A small business can safely hand AI the HR jobs that involve drafting, sorting or answering from documents you control: job adverts, interview question banks, onboarding plans, policy first drafts, standard letters, handbook questions, rota drafts and survey themes. What AI shouldn't do is decide anything about a person: who is hired, disciplined, paid, promoted or let go.
The dividing line fits on a sticky note: AI drafts, a named person decides, and nothing reaches an employee unchecked. The second line is data. A job advert contains no personal information, so almost any tool will do. A letter about someone's sickness absence contains health information, so it belongs only in a business account that doesn't train on your content, with the detail cut to what the letter needs.
Three questions every task on this list had to pass
The twelve tasks below made the list because each one passes the same three tests. Run any other HR idea through them before you try it.
- Is the AI deciding about an individual, or helping a person decide? Only the second kind is here. "Draft three rejection templates" passes; "choose which applicants get the rejection" does not.
- What personal data goes in? None (a blank advert), a little (a name and start date) or sensitive (health, disciplinary history, pay). The more sensitive it is, the narrower your choice of tool and the less detail the prompt should carry.
- Who checks the output, and would they spot an error? If the person reviewing a draft letter wouldn't notice a wrong holiday entitlement, the task isn't ready to hand over yet.
Risk rating for all twelve tasks in one table
| Task | Personal data in the prompt | Who checks | Risk if unchecked |
|---|---|---|---|
| 1. Job adverts and person specs | None | Hiring manager | Low to medium (biased wording) |
| 2. Interview questions and scoring guides | None | Hiring manager | Medium |
| 3. Candidate letter templates | None in the template | Whoever sends them | Low |
| 4. Onboarding plans | Name, role, start date | Line manager | Low |
| 5. Induction quizzes | None | Policy owner | Medium (wrong rules taught) |
| 6. Policy first drafts | None | Owner plus adviser | High (legal wording) |
| 7. Handbook question answers | The question asked | Spot checks, escalation | Medium |
| 8. Standard letters | Name, dates, terms | Sender, every time | Medium to high |
| 9. Absence summaries | Counts, no reasons | Manager | Medium |
| 10. Rota drafts | Names, availability | Rota owner | Medium (broken rest rules) |
| 11. Survey and exit themes | Anonymised comments | Owner | Medium (identifying people) |
| 12. Review drafts from notes | Manager's own notes | Manager, then employee | High if the rating is left to AI |
Hiring paperwork: tasks 1 to 3
1. Job adverts and person specifications
Most small-business adverts start as a manager's scribbled list. AI is good at turning that list into a readable advert plus a person specification that separates must-haves from nice-to-haves, and that split matters later because it becomes the yardstick you score applicants against.
A courier firm hiring a multi-drop van driver might start with notes like "clean licence, 6am starts, handheld scanner, lifting up to 25kg, know the area, Saturdays on rota". Given those notes and the instruction "don't add any requirement, benefit or pay figure I haven't listed", a chat assistant will produce a tidy advert in a minute. Read it for three things. First, invented perks: models like to add "competitive pension" or "free parking" that you never offered. Second, coded wording such as "young, energetic team" or "native speaker", which puts off or excludes people for reasons unrelated to the job. Third, inflated requirements: "two years' experience" appearing from nowhere. Our tutorial on writing job descriptions with AI that attract good hires covers the full workflow, including how to test an advert on a colleague.
2. Interview questions and scoring guides
Structured interviews, where every candidate gets the same questions and answers are scored against written anchors, are fairer than a chat over coffee. They are also tedious to build, which is exactly why AI helps.
Role: laboratory technician, sample preparation and logging.
Must-haves from the person spec: accurate record keeping, following written
methods exactly, working safely with chemicals, flagging problems early.
Write one behavioural question per must-have. For each, describe what a
1 (weak), 3 (acceptable) and 5 (strong) answer contains. Base the anchors
on the job tasks above only. No trick questions.An illustrative answer for "flagging problems early" might set the 5 anchor as "describes noticing an out-of-range result, stopping, recording it and telling a supervisor before continuing". That's usable. What you'd fix: models often reward polished vocabulary in the top anchor ("demonstrates proactive stakeholder communication"), which favours confident talkers over careful workers. Rewrite anchors in terms of what the person did. The tutorial on building interview questions and scorecards with AI shows a full scorecard.
3. Candidate letter templates
Acknowledgements, interview invitations, "you haven't been shortlisted" and offer-stage holding emails are the letters that small firms most often forget to send. A wholesaler that receives 60 applications for a warehouse role can ask AI for four templates, then send them with mail merge from a spreadsheet.
The before-and-after is worth seeing. Before: "Unfortunately you have not been successful on this occasion. We wish you luck." After an AI rewrite with the instruction "warm, plain, under 80 words, tell them what happens to their data": "Thank you for applying for the warehouse operative role. We've now chosen the people we'll interview and, on this occasion, you're not among them. We'll keep your application for six months in case a similar role opens, then delete it. If you'd prefer we delete it now, just reply to this email." Check the retention period matches what you actually do. The watch-out: never ask the model to write personalised rejection reasons from a CV. Reasons, if you give them, come from the person who made the decision.
New starters: tasks 4 and 5
4. Onboarding plans and first-fortnight checklists
A packaging supplier taking on a machine operator needs the new starter to meet the right people, get the right training in the right order and have kit ready on day one. Give AI the role, the shift pattern, the list of mandatory training and who delivers each part, and ask for a day-by-day plan for the first ten working days with an owner against every item.
The draft will be tidy and slightly wrong. Typical errors in an illustrative first attempt: scheduling forklift familiarisation on day one before the operator's certificate has been checked, booking the health and safety lead on a day they're off, and forgetting payroll forms. Those are exactly the errors a line manager spots in two minutes, which is why this is a safe task. Save the corrected plan as a template per role. Planning a new starter's first two weeks with AI has a complete example.
5. Induction quizzes and refresher material
Paste your own manual handling or chemical storage procedure into a business-plan assistant and ask for eight multiple-choice questions, each with the correct answer and the clause it comes from. The citation requirement is the safeguard: models trained on the whole internet sometimes swap your rule for a general one, such as a different lifting limit or a different spill procedure. If a question can't point to your document, delete it.
An illustrative question for a laboratory: "A sample container arrives with a cracked lid. What do you do first? (a) Log it and continue (b) Place it in the secondary containment tray and tell the supervisor (c) Transfer the contents to a new container (d) Return it to the courier." Correct: (b), per section 4.2. That last reference is what lets a supervisor check twenty questions in ten minutes.
Policies, letters and staff questions: tasks 6 to 8
6. First drafts of policies and handbook sections
AI is a fast way to get from a blank page to a structured draft of a lone-working policy, a phone-use-while-driving rule or a hybrid-working section. Tell it what you actually do today ("drivers check in by app at each drop; if no check-in for 90 minutes, the controller calls"), and ask it to write that down clearly, not to invent a best-practice policy you won't follow.
Two watch-outs. Models cite laws, sometimes from the wrong country or out of date, so strip legal references from the draft and let your employment adviser add the ones that apply. And a policy is a promise: if the AI writes "the company will provide annual refresher training" and you don't, that sentence can be used against you later. The tutorial on writing an employee handbook with AI lists which sections need professional review.
7. Answering staff questions from your handbook
Many small-business HR queries are the same dozen questions: how much notice for holiday, what counts as overtime, who to call when sick, where the pension form is. A shared assistant built as a ChatGPT or Claude Project, or a Gemini Gem (becoming a skill from November 2026), loaded with the current handbook can answer these and quote the section it used. (Custom GPTs are being retired by OpenAI, so don't start one now.)
The instruction that matters most: "If the handbook doesn't answer the question, say so and tell the person to ask the office manager. Never guess." An illustrative exchange: a driver asks "Can I swap a Saturday with someone?" and gets "Yes, if both of you agree and the controller approves by Thursday noon (Handbook 6.3). Send the request through the rota app." Spot-check ten answers a week at first. The fuller build, with escalation rules, is in answering staff HR questions with an AI helpdesk.
8. Standard HR letters from checked templates
Letters confirming a change of hours, successful completion of probation or the factual details for a reference (job title and dates) are repetitive and low on judgement. Keep an approved template for each, and use AI only to fill it and adjust tone, never to change the operative wording.
Fill this approved template. Change nothing in the [FIXED] sections.
Employee: [FIRST NAME] Role: warehouse supervisor
Change: hours from 40 to 32 per week, Monday to Thursday
Effective: [DATE] Requested by: employee, approved by: operations manager
Tone: plain and friendly. Output the letter only.Placeholders in capitals keep names out until the final step if you prefer to add them by hand. The watch-out is scope creep: a letter that confirms a contract change, issues a warning or ends employment has legal effect, and the decision and the key wording come from a person with advice where needed. AI can fix the grammar of that letter; it shouldn't write its substance.
Records and rotas: tasks 9 and 10
9. Monthly absence and holiday summaries
Export last month's absence records with names replaced by staff numbers and the reasons column removed, then ask AI for a summary by team: days lost, number of separate absences, holiday booked against allowance, and anyone with more than half their allowance still unbooked by the ninth month of the leave year. For a three-person laboratory team that summary might read: "Tech team: 4 absence days across 3 occasions; 38 of 75 holiday days booked; staff 104 has 17 of 25 days remaining."
That last line prompts a helpful conversation about booking leave. Where it goes wrong is when the summary turns into a score. Ask AI to "rank staff by absence risk" and you've built an automated judgement about individuals from health-related data. Keep it to counts; the manager decides whether any conversation is needed and how it's handled.
10. First-draft rotas and shift patterns
A courier firm with 14 drivers, early, late and weekend shifts, and a handful of standing requests ("no Mondays", "must finish by 3pm on school days") can describe the rules and ask for a four-week draft. Chat assistants are decent at this for small teams, and they fail in one predictable way: they silently break a rule to make the pattern fit.
The fix is to make the model audit itself. Add: "After the rota, list every rule and state for each whether the rota meets it, naming any driver or day that breaks it." Then check the totals with a spreadsheet formula, not by eye. In an illustrative test, the first draft gave one driver six consecutive early shifts against a five-in-a-row rule; the self-audit caught it, and the second draft didn't. Minimum rest between shifts and maximum weekly hours are legal limits in many places, so the rota owner remains responsible for them whatever the draft says.
Listening to staff: tasks 11 and 12
11. Themes from staff surveys and exit interviews
Forty free-text survey comments are hard to read fairly; you remember the loudest. AI can group them into themes with counts and a representative quote for each. An illustrative output for an import-export business: "Workload peaks at month-end (11 comments), unclear who approves overtime (7), praise for the new shipping software (6), car park safety at night (3)."
Small teams need one extra rule: don't report any theme with fewer than three comments, and strip job titles, locations and distinctive phrases before upload, because in a 15-person firm a quote can identify its author. Never ask the model who wrote something. Running staff surveys and exit interviews with AI analysis covers question design as well as the analysis.
12. Turning a manager's notes into a review draft
Managers in small firms often have a year of scattered notes and no time to write them up. Pasting those notes into an assistant and asking for a structured draft (achievements, areas to develop, agreed goals) saves an evening per review. The manager chooses the rating first and tells the model what it is; the model only organises the evidence.
Then run one extra prompt: "Flag any sentence that describes personality rather than work, or that would read differently if written about someone of a different age, sex or background." In one illustrative draft, the model had described a quiet employee as "lacking presence" while the notes only said "doesn't speak up in stand-ups". The fix was to state the behaviour and the goal: "Share progress updates in the Monday meeting." Sensitive notes about health or grievances stay out of the prompt entirely.
HR decisions that stay with a person
Everything above is drafting or summarising. These are the jobs to keep away from AI, however good the tools become:
- Rejecting applicants without a human look. A tool that filters out candidates on its own is making the decision, whatever the vendor calls it.
- Disciplinary outcomes, dismissals and redundancy selection. AI can tidy the letter afterwards; the reasoning and the decision are a manager's, usually with advice.
- Pay, bonus and promotion decisions. Models reproduce patterns in whatever data they see, including historic unfairness.
- Judgements about health or fitness for work. These need a qualified person and the employee's input.
- Monitoring and scoring staff behaviour. Productivity scoring from keystrokes, sentiment analysis of staff chats and similar tools carry the heaviest legal and trust costs.
If you employ people in the EU, two parts of the EU AI Act are relevant. Annex III lists as high-risk the AI systems used to place targeted job adverts, analyse and filter applications, evaluate candidates, make promotion or termination decisions, allocate tasks based on behaviour or personal traits, and monitor or evaluate performance; the obligations for these stand-alone systems have been deferred to 2 December 2027, but the direction is clear. Separately, using AI to infer employees' emotions from biometric data such as facial expressions or voice has been banned in the workplace since 2 February 2025, apart from medical or safety uses. Data-protection law such as the GDPR also restricts decisions made solely by automated means that significantly affect someone. None of this bites if AI only drafts and a person decides, which is one more reason to keep to that rule. For anything near the line, ask an employment or data-protection adviser before you build it.
A 26-person courier firm puts six of the twelve to work
Picture an illustrative courier firm: 20 drivers, four controllers and two office staff, with the office manager spending about a day and a half a week on HR. The owner picks tasks 1, 3, 4, 7, 8 and 10 because they're frequent and easy to check, and leaves surveys and reviews for later.
Tooling: two Claude Team Standard seats (the plan's minimum) at $25 a seat on monthly billing, so $50 a month, for the office manager and the operations manager. Drivers use the handbook assistant through a shared Project that the office manager maintains; if the firm wanted every driver to have their own seat, the cost would rise per seat, so it starts with a staff-facing channel where the office manager relays answers for the first month.
| Task | Hours a month before | Hours after (illustrative) | One-off setup |
|---|---|---|---|
| Adverts and candidate letters (1, 3) | 5 | 2 | 2 hours |
| Onboarding plans (4) | 3 | 1 | 2 hours |
| Staff questions (7) | 5 | 2 | 10 hours tidying the handbook |
| Standard letters (8) | 4 | 1.5 | 3 hours approving templates |
| Rota drafts (10) | 6 | 2.5 | 3 hours writing the rules down |
| Total | 23 | 9 | 20 hours |
On those assumptions, the firm saves about 14 hours a month for $50, and recovers the 20 setup hours in under two months. The first month also produced the most useful lesson. The handbook assistant told a driver that overtime was paid after 45 hours, because an old version of the handbook was still in the Project alongside the new one, which says 42. The fix took five minutes: one handbook file only, with the version date in the file name, and a monthly check that the Project holds nothing else.
Pre-send checks for any AI-drafted HR document
- Facts match your records: dates, hours, entitlements and names checked against the HR system, not the draft.
- Nothing invented: no benefits, policies, legal references or promises you didn't supply.
- Wording is about work, not personality, and would read the same whoever it was about.
- Personal data kept to the minimum the document needs, and none of it pasted into a personal or free account.
- A named person approved it, and for anything with legal effect, from a template your adviser has seen.
- Tone fits the moment: AI drafts can sound cheerful in letters that should be neutral.
- A record exists of what was AI-drafted and who decided, kept with the employee file where appropriate.
Start with two tasks from the low-risk rows of the table, run them for a month, and only then add the ones that touch individual employees' records. Handing over HR admin this way frees hours without ever handing over the judgement your staff expect from you.
Follow-up questions small employers ask about AI and HR
Do I need to tell staff we use AI for HR paperwork?
It is good practice to say so in your staff privacy notice or handbook: which tools you use, for which jobs (drafting letters, answering handbook questions, drafting rotas) and that decisions about individuals are made by people. If you ever use AI to evaluate or monitor staff, the disclosure and consultation duties get much stricter in many places, so talk to an employment adviser before starting anything like that.
Can a free ChatGPT account be used for HR drafting?
For templates that contain no personal details, such as a blank job advert or a policy outline, a free account with the model-training switch turned off is workable. As soon as a prompt names an employee or describes their situation, move to a business plan such as ChatGPT Business, Claude Team, Microsoft 365 Copilot or Gemini in Workspace, which don't train on business content by default and keep the work in a company account.
What if an employee asks whether AI was involved in a decision about them?
Answer plainly. If AI drafted the letter but a manager chose the outcome, say exactly that. Keep a short note of which parts were AI-drafted and who made the decision, so the answer is easy to give months later. If the honest answer would be that a tool scored or ranked the person and nobody reviewed it, the process has crossed the line this tutorial draws, and it needs fixing before the next case.
How soon do these tasks start saving time?
Letter templates, job adverts and onboarding checklists pay back within the first week because you reuse them straight away. A handbook question assistant takes longer: expect two to four weeks, most of it spent tidying the handbook so the assistant has one current version to answer from. Rota drafting sits in between and depends on how clearly you can write your shift rules down.
Further reads
- Can AI Screen CVs Fairly? What Small Employers Need to Know — Why CV shortlisting is left off this list, and how to do it fairly if you must.
- Employment Contracts With AI: Drafting Steps and Legal Checks — Contracts carry legal weight; here is where AI drafting stops.
- How to Write Fairer Performance Reviews With AI — A fuller method for task 12, with bias checks on the wording.
- AI Shift Scheduling vs a Rota Spreadsheet for Small Teams — When a rota tool beats prompting a chat assistant for task 10.
- How to Automate Holiday Requests and Absence Tracking — Takes task 9 further by automating the requests themselves.
- How to Write an AI Usage Policy for Your Small Business — Puts the 'AI drafts, a person decides' rule into writing for staff.
- AI Use Cases by Department for Small Businesses (With Examples) — 21 practical AI use cases across seven departments, each shown in a small insurance brokerage, with how to start and what to watch.
- AI Payroll Checks: Catch Timesheet Errors Before Payday — A pre-payroll check that runs exact rules on timesheets first, then uses AI to read notes, spot odd patterns and draft one query per approver before cut-off.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: EU AI Act Annex III point 4 (employment) and Article 5(1)(f) on emotion recognition, via the AI Act Service Desk and legal commentary; Digital Omnibus deferral dates; GDPR Article 22 text; OpenAI and Anthropic business plan pages for seat prices and training defaults.