How to Answer Staff HR Questions With an AI Helpdesk

Collect your handbook, policies and the 30 to 50 questions staff ask into one current set of documents. Load them into a shared AI assistant your team already uses: a Gemini Gem (becoming a skill from November 2026), a ChatGPT or Claude project, or a Copilot agent. Tell it to answer only from those documents and hand anything personal to a named person.

The hard part is the documents, not the AI. An assistant will confidently quote the old expenses rate if last year's policy is still in the folder. And some questions should never be answered by a bot at all: anything about a particular person's pay, health, conduct, a grievance or their job goes to a human, and the assistant's job is to say who and how.

Follow me on Instagram@sagnikteaches

Two weeks of HR questions at a small charity

The illustrative example here is a small charity with 18 staff across two sites, plus about 30 volunteers. Its operations manager handles HR one day a week alongside everything else, and most of that day goes on answering the same questions by email and in the corridor. Before building anything, she logged every HR question for two weeks: 52 questions from 14 people. An extract, grouped:

Connect on LinkedInSagnik Bhattacharya

Two patterns stood out. About two thirds of the questions were general policy questions with answers already written down somewhere. And the personal ones were few, but they were the ones where a slow or clumsy answer does real damage.

Subscribe on YouTube@codingliquids

Three lanes: answer, answer and signpost, hand to a person

Decide the routes before choosing any tool. Every question falls into one of three lanes:

Write the third lane as a list of topics and trigger words in the assistant's instructions, and err wide. It is far better for the assistant to hand over a question it could have answered than to give a policy summary to someone who has just disclosed a health condition.

Clean the source documents first

The assistant can only be as right as its documents. The charity's clean-up took about four hours and found problems in almost every file:

• Two versions of the same policy. The time off in lieu policy existed as a 2023 version and a 2025 update, both in the shared folder. The older one allowed TOIL to be taken within three months; the newer one said one month.
• Conflicting figures. The handbook gave a mileage rate of $0.30 per kilometre; the expenses claim form, updated more recently, said $0.35. Staff had been asking because they had noticed.
• Drafts and templates. A draft hybrid working policy that was never approved sat alongside the real ones.
• Missing answers. Nothing written said who to ring when off sick at the second site.

The fixes: one folder holding only current, approved documents, each with a version date on the first page; old versions moved to an archive the assistant cannot see; the conflicts resolved and corrected; and one new document, a plain question-and-answer sheet covering the gaps the log found. An AI assistant helps with the clean-up itself: give it two policies and ask "List every point where these documents contradict each other or give different figures." If your handbook needs rewriting rather than tidying, writing an employee handbook with AI covers that, and building a company knowledge base AI can answer from explains how to structure documents so an assistant finds the right passage.

Where the helpdesk should live

The best home is usually a tool everyone already has, because an HR helpdesk that needs a new login will not get used. The main options for a small organisation, as of September 2026:

A few details matter when choosing. With a shared Gem, Google's help page on sharing Gems notes that people you share it with can view its instructions and uploaded files, which is fine for policies but means nothing confidential should go in. For Microsoft, Microsoft's pay-as-you-go overview explains the billing policy an admin needs to create. HR platforms are the only option here that can answer "how much holiday do I have left": BambooHR, for example, describes its Ask BambooHR assistant as answering from HR data, company policies and uploaded documents, on desktop, mobile or in Slack; check which plan includes it. One thing not to do: do not build a new helpdesk as a custom GPT, because OpenAI is retiring custom GPTs, which stop running on 11 December 2026.

The charity already used Google Workspace for Nonprofits, so a shared Gem cost nothing extra and every staff member already had access. That decided it.

Instructions that keep the assistant honest

The instructions are the most important few hundred words in the project. The charity's version, lightly edited:

You are the HR helpdesk for [charity name] staff. Answer questions about our HR
policies using ONLY the documents provided. They are current as of the date on each.

How to answer:
- Give the answer in 2-4 plain sentences, then name the document and section,
  e.g. "(Staff handbook, section 6.2)".
- If the documents don't answer the question, say "I can't find that in our
  policies" and suggest asking the operations manager. Never answer from general
  knowledge or from employment law in general.
- If a question could mean staff or volunteers, ask which.

Hand over, do not answer, if the question involves: the person's own pay, balance
or records; health, disability or sickness details; pregnancy, adoption or family
leave for their own situation; a complaint, grievance, bullying or harassment;
disciplinary matters; redundancy or job security; anything distressing or urgent.
For these, reply kindly, don't ask for details, and give: "Please contact
[operations manager] directly on [email/phone], or [chair of trustees] if it
concerns the operations manager." If someone may be at risk, give the safeguarding
lead's contact first.

Illustrative answers from the charity's testing, one from each lane:

Q: What's the mileage rate?
A: Mileage is paid at $0.35 per kilometre for business journeys in your own car. Claim
   monthly on the expenses form with dates, start and end points and the purpose of
   each trip. (Expenses policy, section 3; Expenses claim form)

Q: Can I carry over holiday?
A: You can carry over up to 5 days into the next holiday year, to be used by the end
   of March. Anything above that is lost unless your manager agrees otherwise in
   writing. For your own balance, check with the operations manager.
   (Staff handbook, section 6.4)

Q: I've been signed off by my doctor for a month, what happens with my pay?
A: I'm sorry you're unwell. The operations manager can go through your pay and
   support with you personally. Please contact [operations manager] on [contact].
   You don't need to share details here.

The third answer is exactly right: warm, brief, no attempt to explain sick pay rules, and no request for medical details in a chat.

Test it with 40 questions before anyone relies on it

Build a test set from the question log: 25 real policy questions, 10 hand-over questions phrased in different ways, and 5 awkward ones (vague, staff-or-volunteer, or about something no policy covers). Write the correct answer or route for each, run them all, and mark each response right, wrong or half right. The charity's first run scored 31 of 40. The failures were instructive:

• Answering from general knowledge. Asked "what's the minimum holiday I'm entitled to?", it gave a general legal figure it knew, not the policy. The instruction "never answer from employment law in general" was added in bold, and the question now gets the policy figure plus a suggestion to ask the operations manager about legal minimums.
• Missing a hand-over. "My partner's due in March, what can I take?" was answered with a summary of the family leave policy. It is a personal family leave question, so the trigger list now includes phrasing like "my partner", "expecting" and "due".
• Staff and volunteer confusion. "Can I claim for parking?" got the volunteer answer for a staff member. The instruction to ask which fixed it.
• Old wording surviving. One answer quoted the 2023 TOIL rule, because a copy was pasted inside another document. Removing it fixed the problem and showed why the clean-up matters.

The second run scored 38 of 40, and both remaining misses were "half right" answers the manager was happy to live with. Re-run the whole test set after every policy change.

Pilot with four people for two weeks

Before announcing it to everyone, give the assistant to a small group: at the charity, two staff from each site, including one sceptic and one person who asks a lot of HR questions. Ask them to use it for every HR question for two weeks and to forward any answer that looked wrong, vague or oddly worded. The pilot group asked 19 questions. Seventeen answers were fine. One was right but quoted the handbook's section number from an older layout, which confused the reader, so the handbook's headings were renumbered and the file replaced. The other was a question nobody had anticipated, "can I bring my dog to the office?", which no policy covered; the manager wrote a one-line answer into the question-and-answer sheet. Small fixes, but finding them with four friendly testers is far better than finding them with eighteen staff and a reputation to win.

Tell staff what it is, and what it isn't

How you launch it decides whether people trust it. Keep the message short and honest:

"We've set up an AI helpdesk for quick questions about our HR policies: leave, expenses, time off in lieu, sickness reporting and where to find forms. It answers only from our current policies and tells you which one it used. It can't see your personal records, and it isn't the place for anything personal or sensitive: for those, please contact [operations manager] directly, as before. If an answer looks wrong, tell us so we can fix it. Our policy documents are always what count."

Being clear that it is an AI matters for trust, and if you employ people in the EU, the EU AI Act's transparency duties, which have applied since 2 August 2026, include telling people when they are interacting with an AI system. The name and first message of the assistant should make that obvious.

What staff type in, and who can see it

However clear the launch message, some people will type personal details into the assistant: a diagnosis, a family situation, a complaint about a manager. Plan for it. Use a business or nonprofit plan whose terms keep business content out of model training by default, find out what chat history admins can see and how long it is kept, and tell staff plainly. Add a line to your AI or data policy that the HR assistant is for policy questions only. The wider rules for staff AI use are covered in the HR tasks you can hand to AI safely.

Keeping the helpdesk's answers current

An HR assistant goes stale the day a policy changes and nobody updates its files. Give it one owner (here, the operations manager) and three routines:

• Policy change: the new version replaces the old in the assistant's documents the same day it is approved, and the test questions for that policy are re-run.
• Monthly review: 20 minutes looking at any "I can't find that" answers staff reported and any hand-overs, adding answers to the question-and-answer sheet where a policy gap caused them.
• Annual check: the whole test set, and a read of the instructions to make sure contacts and names are still right.

If holiday requests are still handled by email, the helpdesk will keep getting "how do I book leave" questions; automating holiday requests and absence tracking removes the question by making the process obvious. For how customer-facing assistants handle the same hand-over problem, when an AI chatbot should hand over to a human has useful rules that transfer well.

What it changed, and where else it fits

Before the helpdesk, the operations manager spent about two and a half hours a week on HR questions (roughly 25 questions at six minutes each). Three months after launch, in this illustration, about 60% of questions went to the assistant, and her HR question time fell to about an hour a week. The quality of the remaining time changed too: the personal questions reached her directly and sooner, rather than arriving mixed in with mileage queries.

Two other shapes of the same idea:

• A sports equipment shop with part-time shift staff gets most questions about the staff discount, rota swaps, uniform and breaks, usually asked from a phone before a shift. An assistant inside its HR or rota app, where staff already look, works better there than a chat tool on a laptop they rarely open.
• A community interest company with many volunteers builds a separate volunteer assistant from its volunteer handbook: expenses, what to do if a session is cancelled, and who to contact. Its hand-over list puts safeguarding first: any mention of a concern about a child or vulnerable adult gets the safeguarding lead's contact immediately and nothing else.

HR helpdesk questions from managers

Can the helpdesk tell staff their own holiday balance?

Only if it is connected to the system that holds balances, with each person seeing only their own record. A general assistant loaded with policy documents cannot and should not try. HR platforms with built-in assistants can answer personal questions like balances because they already manage who can see what; otherwise, have the assistant point staff to where they can check.

Should volunteers use the same assistant as staff?

Usually not the same one, because their rules differ: volunteers typically have expenses and role guidance rather than pay, leave and contracts. A single assistant holding both sets of documents will mix them up. Build a separate volunteer assistant from the volunteer handbook, or at minimum make the instructions ask whether the person is staff or a volunteer before answering.

What if the assistant gives a wrong answer that someone relies on?

State clearly at launch that the policy documents are what count and the assistant is a quick way to find them. When a wrong answer is reported, correct it with the person directly, fix the source document or instructions the same week, and add the question to your test set. Treat a wrong answer that someone acted on like any other HR error and put it right fairly.

Further reads

• Employment Contracts With AI: Drafting Steps and Legal Checks — Get the contracts behind the answers right.
• AI Employee Onboarding: Plan a New Starter's First Two Weeks — Introduce new starters to the helpdesk in week one.
• Standard vs Premium AI Seats: Who Needs the Bigger Plan? — Work out which staff need which AI plan.
• Microsoft 365 Copilot Pricing: Business vs Enterprise Licences — Understand Copilot licence costs before choosing it.
• How to Measure Whether Your AI Chatbot Is Actually Working — Measure whether an assistant is actually answering well.
• How to Write an AI Policy for a Small Charity — Set the wider rules for AI use in a charity.
• 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.
• How to Deflect Password Resets and Routine Requests With AI — Cut password and routine IT tickets: self-service reset licensing, which requests are safe to deflect, bot guardrails, sample chats and honest metrics.
• How HR Consultants Use AI for Policies, Letters and Cases — Three workflows for HR consultants, with a client fact sheet, letter before-and-after, a grievance chronology prompt and the lines AI must never cross.
• AI Recruiting Tools for Small Businesses: What's Worth Paying For — Nine kinds of AI recruiting tool, what each costs, which a small business should pay for, and three to skip, with examples from real hiring tasks.
• How to Run Staff Surveys and Exit Interviews With AI Analysis — A 12-question survey, an exit interview script, anonymity settings, a theme-coding prompt with checks, and an online clothing shop's first round of results.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Gemini Apps Help, Share a Gem; Microsoft Learn, Copilot pay-as-you-go overview; BambooHR, Ask BambooHR product page; OpenAI and Anthropic plan and nonprofit pricing pages; the EU AI Act, Article 50.