AI Tools for HR Professionals: Policies, Onboarding and Staff Questions

An HR professional in a small business can use AI for three jobs: drafting policies from a facts sheet you control, building onboarding checklists from your own process notes, and answering routine questions from the approved handbook. Keep staff records out of public tools and have an employment-law adviser review anything binding.

The common failure is asking a chatbot to write a policy from nothing. The result reads well, quotes rules from nowhere in particular and may describe a law that has since changed. The workable pattern is the reverse: you supply the facts and the rules, the tool drafts and cross-checks, and someone qualified signs off. Everything else here rests on that one habit.

Follow me on Instagram@sagnikteaches

Policies: write the facts sheet first, then let AI draft

A facts sheet is one page listing the decisions your business has already made about a topic. It contains no legal claims, only your choices. The AI drafts from it and from nothing else, and where it lacks a fact it must say so instead of filling the gap with a plausible clause.

Connect on LinkedInSagnik Bhattacharya

For a sickness absence policy at an illustrative 25-person roofing company, the sheet might read:

Subscribe on YouTube@sagnikteaches

SICKNESS ABSENCE: FACTS SHEET
- Reporting: phone your line manager before 7:30 on the
  first day (site start is 8:00). If no answer, text, then
  keep trying until 9:00.
- Self-certification covers days 1-7. Medical evidence
  from day 8.
- Return-to-work conversation after every absence, within
  2 working days, manager plus a short form.
- Welfare meeting after 3 absences in 6 months or 8 days
  in 12 months (company decision).
- Pay: statutory sick pay only, no enhanced scheme
  [adviser to confirm wording].
- Safety rule: nobody returns to work at height until the
  manager has confirmed fitness in the return-to-work
  conversation.
- Contact: office manager, [phone].
- Status: non-contractual. Reviewed every year.

The prompt that goes with it does three things: restricts the source, forces gaps into the open, and asks for a review sheet.

Draft a sickness absence policy in plain English (reading
age about 12) for the company described in the facts sheet.
Use ONLY the facts sheet. Sections: purpose, who it covers,
reporting, evidence, return to work, welfare meetings, pay,
confidentiality, review.
Wherever you need a fact that is not on the sheet, write
[DECISION NEEDED: what].
Do not state any legal entitlement, minimum or notice period.
After the policy, add a REVIEW SHEET table with columns:
sentence | why it depends on law or on another document |
question for the employment adviser.

A sickness policy refresh, start to finish

Here is how the job runs with numbers. Before AI, the office manager would spend about five hours: finding the old policy, reading two internet templates, rewriting, checking the rest of the handbook for conflicts and sending a draft to the adviser with a note asking them to "have a look".

With the method, the facts sheet takes about 45 minutes, because the decisions mostly exist already in managers' heads. The draft and the review sheet take 20 minutes of prompting and reading. The office manager spends another 60 minutes reading the draft line by line against the facts sheet. Total: 2 hours 5 minutes, against five. Across a twelve-policy handbook that is about 35 hours a year back. The adviser's part gets smaller too, because they receive a clean draft and a list of exactly which sentences depend on law.

The review sheet is the useful part. A typical output has rows like these:

Expect the draft itself to contain one sentence you did not ask for. In this case the model wrote "statutory sick pay is paid from the fourth day of absence", a sentence that came from its training, not from the facts sheet, and that conflicts with the government's current sick pay guidance (more on that below). Because the review sheet lists every legal claim, it appears there, and the adviser can confirm or correct it. Without the sheet, it would have read as settled fact.

A quick example of the plain-English pass, which is where AI is quietly excellent. The old wording: "Employees shall notify their line manager of any absence due to incapacity for work no later than 7.30am on the first day of absence, failing which the Company reserves the right to treat the absence as unauthorised." The rewrite: "If you cannot work because you are ill, phone your line manager before 7:30 on your first day off. If you cannot reach them, send a text and keep trying until 9:00. If we hear nothing, we may record the day as unauthorised absence, so please get in touch." Same obligations, read in ten seconds. The longer method for a whole handbook is in how to write an employee handbook with AI and what to check.

Which policies to start with, and which to leave to the adviser

Not every policy suits the method. The cost of a wrong sentence differs enormously between an expenses limit and a redundancy procedure, so sort the handbook into three bands before drafting anything.

Learn the loop on a process policy first. An expenses policy gives you a complete facts-sheet, draft, check and issue cycle in an afternoon, with almost nothing riding on it. Move to the rule-heavy band once the review sheet feels routine. For the third band, a wrong sentence is the kind that turns up in a tribunal bundle, so the adviser writes the text and AI only helps you read it back to staff in plain words.

When the law moves under your handbook

A model drafts from what it learned in training, not from today's rules. Employment law is a fast mover: the Employment Rights Act 2025 changes employment rules in stages, with some measures in force from April 2026 and others planned for 2027. A policy drafted from a model's memory, or from a template written last year, can describe a rule that has since moved. For instance, the government's own sick pay guidance now says statutory sick pay is paid for every full day off sick that the employee would normally have worked, while older handbooks and templates may still describe unpaid waiting days before it starts. The same guidance asks for a fit note after more than seven days off in a row, which is why day 8 appears on the facts sheet above. Check current official guidance before relying on any date in this paragraph.

The safe pattern is law in, never law out. Paste the official text yourself, whether a government guidance page or your adviser's update, and ask the tool to compare two documents you supplied.

Below are (A) the official guidance text I have pasted and
(B) our current sickness absence policy. List each section
of B that A may affect, quote the lines from each side, and
say what kind of change might be needed. Do not tell me what
the law is; only compare the two texts.

Keep a one-page policy calendar: policy, owner, date last checked by the adviser, next review date and the events that trigger an early review. It is the dullest document in HR and the one that stops stale policies surviving for years. For the deadlines side, see the probation, renewal and right-to-work deadline checklist.

Onboarding checklists built from your own process

Onboarding fails in small businesses because the process lives in one person's memory. AI is good at turning a messy description into a checklist with owners and dates, which makes it a fair first step towards writing the process down. It needs role-level information only: no candidate's CV, no identity documents, no personal details. Use "[new starter]" throughout.

Feed it the role, the systems the person will need, the mandatory training, who does what, and the order things must happen in, then ask for one checklist per role arranged by timing and owner. A plumbing firm hiring an office administrator and a van-based apprentice shows why one master checklist is not enough. The administrator needs email, the job-management app and phone set-up. The apprentice needs a safety induction before touching a site, PPE and tool issue, a driving-licence check if they will drive, and a supervisor pairing. An excerpt of the apprentice's sheet looks like this:

Read the draft for steps that are quietly unsafe. In one realistic output the tool included "ask the new starter to email their bank details to payroll". Replace it with the secure form in your payroll system. Also strike any step the model has invented, such as a formal "probation sign-off" you do not run. The AI can draft the welcome email, the day-one logistics message and the manager's first-week plan from the same checklist in a minute or two. For a two-week example with the messages included, see planning a new starter's first two weeks.

The welcome email and the day-one message are where an eager tool does most damage, because it likes to be reassuring. Left alone it promises a laptop on the first morning, a parking space, a team lunch and a mentor you have not chosen. Give it a confirmed-logistics box, as you gave the policy tool a facts sheet, and tell it that anything outside the box must read "we will confirm".

Write a welcome email for [new starter], joining as [role]
on [date]. Friendly, under 150 words.
Use ONLY these confirmed details: start 8:00 at [address];
ask for [supervisor] at reception; bring photo ID; return
the payroll form (not by email); safety boots are issued on
day one.
Anything not listed (parking, lunch, equipment, seating) must
say "we will confirm this before your start date".
Do not mention pay or probation.

A good output ends with something like "We will confirm parking and your equipment before you start", which is honest and costs nothing. A bad one says "your laptop will be waiting at your desk", which you must then deliver or apologise for. The same box works for the manager's first-week plan.

Time it once. A first checklist for a new role takes about three hours by hand, including formatting. With AI it takes 40 minutes including your corrections, and every later hire in the same role takes ten minutes to adapt. If you hire six people a year across three roles, three templates take two hours instead of nine.

A staff question assistant that knows when to stop

Most HR inbox traffic is routine: how to book leave, who to tell about sickness, what the expenses limit is, where the policy is. An assistant that answers those from the approved handbook saves time. Its more important feature is knowing which questions it must not answer.

Set it up as a project in a tool the business already has. ChatGPT Projects support sharing on every plan and, on the Business plan, up to 40 files per project; Claude Projects can be shared on Team and Enterprise plans; a Gemini Gem works for Google Workspace teams, though Google is replacing Gems with skills (on Workspace business accounts from March 2027, with existing Gems moved across automatically). Upload only documents staff are allowed to read, such as the handbook, leave and expenses policies, IT rules and health and safety procedures. Never upload salary bands, disciplinary records or anything about individuals, and check what collaborators can see: Google's help pages say people a Gem is shared with can view its instructions and uploaded files, so assume the same until you have checked another tool.

You answer staff questions for [Company] using ONLY the
documents in this project (updated [date]).
1. Quote the policy name and section you rely on.
2. If the documents do not answer it, say: "I cannot find
   that in our policies. Please ask [HR contact]."
3. Never give legal advice or guess.
4. If the question involves pay disputes, discipline,
   grievances, health, disability, pregnancy or family
   leave, redundancy, immigration status or harassment,
   reply ONLY: "This needs a conversation with [HR contact].
   They will reply by [time]." Do not draft an answer.

Test before launch with 20 real questions from your inbox, anonymised: ten the handbook can answer and ten that must escalate. A good answer to "How much notice do I need to book a week off?" reads, from an illustrative leave policy: "Policy section 3.2 asks for at least twice as much notice as the leave you want, so two weeks for a week off. Your manager approves it. Request it in [system]." A good answer to "I'm pregnant and want to talk about my duties on the roof" is the escalation line and nothing else. The launch rule is simple: every one of the ten must-escalate questions escalates. Nine out of ten is a fail, because the tenth is the one that matters.

Here is the shape of the test sheet, with illustrative questions. Notice that the third row looks routine and is not.

Keep a log of questions it could not answer. That list tells you which handbook pages are missing, and it is arguably the assistant's best output. The fuller design is in how to answer staff HR questions with an AI helpdesk. As a quick sum for a company receiving about 25 routine questions a week at six minutes each, even answering half instantly gives back about 75 minutes a week, which is useful, though modest. The value is more consistent answers.

Staff data: where UK GDPR draws the line

Staff information is personal data, and UK GDPR applies to it as it does to customer data. In outline, you need a lawful basis for each use, you must tell staff how their data is used, you should collect only what you need, and you must keep it secure. Health, disability, ethnicity, religion and trade-union membership are special-category data, which carry stricter conditions. If an AI vendor handles staff data for you, you need written data-protection terms with it, and business plans normally publish a data-processing addendum. A data protection impact assessment is expected where processing is likely to be high risk, which an AI tool that scores or profiles staff probably is. If you employ people elsewhere, substitute the equivalent rules and take advice.

Business plans from ChatGPT, Claude, Microsoft 365 Copilot and Gemini in Workspace do not train on your business content by default, while consumer plans carry a training switch that staff may or may not have turned off. That difference drives a three-zone rule for HR use:

One more rule belongs in the policy. Data-protection law restricts decisions made solely by automated means that significantly affect people, so no AI tool decides who is shortlisted, disciplined or dismissed. It may help you draft the letter after a person has decided, and nothing more. If you are unsure about free plans and staff data, free AI for HR professionals and staff data goes through them one by one.

What to send your employment adviser

A review request that says "can you have a look at this handbook" gets a vague answer or a large invoice. Send a packet:

• The final draft, with the facts sheet it came from.
• The AI-produced review sheet, with your own answers filled in wherever you can supply them.
• Where your staff work and what roles they hold, since rules differ by location and by role.
• The relevant contract templates, so the adviser can check the policy does not conflict with them.
• Three specific questions you most want answered, in order of worry.

Ask for a fixed fee up front if you can, and explain that the review sheet narrows the scope. Ask whether each policy should be contractual or non-contractual, whether the wording matches the contracts, and whether anything has changed recently that your template might not reflect. Record the date of the review and the adviser's name on the policy calendar.

Four numbers that show the workflows pay

Keep a simple tally from the first month. First, the hours: time the next policy refresh against your five-hour baseline. Second, adviser edits: count the changes the adviser makes to each AI-assisted draft. The count should fall as your facts sheets improve; if it does not, your sheets are too vague. Third, assistant accuracy: re-run the 20-question test every quarter, and after every handbook change. Fourth, new-starter feedback: ask each person at day 30 what was missing from their first week and add it to the checklist.

Four slips to plan for

These are illustrations of how the work goes wrong, not reports of real events.

• The helpful clause: a draft includes a "probation extension" right nobody chose, because most templates contain one. The review sheet catches it only if the instruction says to list every sentence not traceable to the facts sheet, so say so every time.
• The generic answer: the staff assistant answers a question about time off for a family emergency with a cheerful summary of the leave policy. Add "family emergency" to the escalate list, then retest all twenty questions.
• The bank-details step: the onboarding checklist suggests emailing payment details. Replace it with the secure form and add a line to the prompt: "no step may ask for personal or financial data by email".
• The spreadsheet paste: someone drops a full sickness absence export into a free chatbot to look for patterns. That is red-zone data in the wrong tool. The safe route is counts by team and month, with no names, run in a business account.

Each of these costs minutes to prevent. If you would rather build the facts-sheet routine and the staff assistant with someone alongside you, that is what my AI implementation consultation covers, and the contact page is where to start.

Questions HR leads ask before using AI with staff documents

Can I put an employee's name into ChatGPT or Claude?

On a business plan with a data-processing agreement, a name in a routine letter is a modest risk, provided your staff privacy notice covers it. Use initials where the task allows, and keep health details, disability, pregnancy, grievance and performance information out of general AI tools altogether. A free personal account is the wrong place for any staff data.

Do I have to tell staff we use AI?

If staff personal data goes into an AI tool, data-protection law expects you to tell people how their data is used, usually through your staff privacy notice, so update it and ask your adviser to check the wording. Policies drafted with AI need no label, but a person you name should own and sign them.

Should the staff assistant answer questions about pay or discipline?

No. Limit it to questions the handbook answers in plain words, such as how to book leave or report sickness. Anything about pay disputes, discipline, grievances, health, pregnancy, family leave, redundancy or immigration status should produce a short message sending the person to a named HR contact.

How often should AI-drafted policies be reviewed?

At least once a year and whenever the rules change. Put each policy on a calendar with an owner, the date it was last checked by your adviser and a trigger for early review, such as a change in the rules, a new contract template or a site or role change. The calendar matters more than the drafting tool.

Which plan is enough for a small HR team?

A business plan with two seats is enough to start: ChatGPT Business and Claude Team both list $20 a user a month on annual billing or $25 monthly, with a two-seat minimum. If your business is on a Microsoft 365 business plan, Copilot Chat is included at no extra cost and may cover drafting. Check data terms before staff records go in.

Further reads

• AI for HR in Small Businesses: 12 Tasks You Can Hand Over Safely — Twelve other HR jobs, sorted by how safe they are to hand over.
• ChatGPT and Copilot for HR: Which Suits a Small People Team? — Helps choose between ChatGPT and Copilot for a small people team.
• ChatGPT for HR Professionals: Job Ads, Reviews and Warning Letters — Covers job adverts, reviews and warning letters, which this tutorial leaves out.
• HR Automation AI Tools: What a Small Business Can Automate Safely — What can be automated in HR without a person in the loop.
• AI HR Software for Small Business: What to Buy and What to Skip — Useful if you are weighing HR software with AI built in.
• Can AI Draft Disciplinary Letters? What HR Consultants Must Check — Shows the checks a disciplinary letter needs before it is sent.
• 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.
• Best AI for Business Training: Onboarding, Compliance or Skills Gaps? — Which AI to use for onboarding, compliance training or skills gaps in a small business, with a decision rule, prices and worked examples.
• Care Staff Supervision Template: Notes and Action Logs With AI — A care-specific supervision agenda, record format and action log, with AI drafting the actions from your notes while staff data stays inside approved tools.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business pricing and OpenAI Projects help pages, Claude pricing page, Microsoft 365 Copilot business page, Google Gemini help pages on sharing Gems and on the move from Gems to skills (checked November 2026); government guidance on statutory sick pay and the Employment Rights Act implementation timeline; UK GDPR articles 9, 28 and 35 and the regulator's guidance on automated decisions. Worked figures and sample documents are illustrative.