AI Chatbot for HR: A Handbook Q&A Service HR Consultants Can Sell

Build it as a handbook-only bot. Load the client's current employee handbook and nothing else, switch off general-knowledge answers, give the bot a short list of topics it must hand to a named adviser, show a disclaimer on every chat, and test 30 questions before launch. Charge a one-off setup fee plus a monthly retainer.

The catch is the hand-over. Limiting a bot to one document is easy on most chatbot platforms, but passing a live chat to a person usually sits on a dearer plan, and your fee has to carry that. So sell a maintained service rather than software: the bot answers the repetitive questions (leave, sickness reporting, expenses, dress code), and you stay responsible for everything that needs judgement.

Follow me on Instagram@sagnikteaches

What you are actually selling

Think of a handbook bot as a search tool that speaks plain English and quotes the client's own document, rather than an HR adviser in a box. That framing keeps the sales conversation honest and keeps your liability small. Put three promises in your proposal and nothing else.

Connect on LinkedInSagnik Bhattacharya

• Answers come from the handbook only. If a policy is not written down, the bot says so.
• Named topics go to a person. The list of those topics is agreed in writing, per client.
• You maintain it. When the handbook changes, you update the bot and re-test it within an agreed number of working days.

What clients usually value is the line manager who stops being asked, for the fortieth time, how many days' notice a holiday request needs. A bot that handles those questions well, at eleven at night, is a retainer a small business owner understands.

Subscribe on YouTube@sagnikteaches

It also pairs naturally with the drafting work you may already do with AI, which the tutorial on AI for HR policies, letters and cases covers. The bot is the client-facing side; that tutorial is the back-office side. If a client asks why they should not build it themselves, how to answer staff HR questions with an AI helpdesk shows the in-house version, and your proposal should say plainly what the retainer adds to it: the handbook clean-up, the test sheet, the escalation rota and the monthly upkeep.

Choose a platform that can hand over to a human

You need three things from the software: answers restricted to uploaded sources, a way to pass a conversation to a person, and a conversation log you can review. Not every cheap plan has all three, and the hand-over feature is the one that is usually gated.

Botpress puts routing and assignment on its Team plan ($750 a month billed annually), which is too much for one small client. That is why, for most single-client retainers, I would start by testing on Botpress Free, then decide between Chatbase Basic and a build where the bot never hands over live but always names a person and gives a contact route. Prices change; check each vendor's pricing page the week you quote.

The cost belongs in your fee, not on a separate invoice the client can question. A client with 60 staff may generate well under 250 chats a month on a handbook bot, but you should check the first month's real figures before you promise a flat price. Building the FAQ your chatbot needs covers how to write the answers the bot will lean on most.

Prepare the handbook before the bot ever sees it

When a handbook bot misleads staff, the cause is usually the document rather than the AI. Handbooks contradict themselves, mention policies that moved to an intranet page, and use "reasonable notice" without defining it. Fix that first, and bill for it as a separate line.

The pre-upload check

1. Export the handbook to a clean text-based PDF or Word file. Scanned pages with no text layer are invisible to the bot.
2. Strip anything that is not policy: the welcome letter, organisation chart, photos with captions.
3. Search for the words "see attached", "as discussed" and "from time to time". Each one is a gap the bot will either skip or invent around.
4. Add an effective date and version number on page 1 and put them in the file name, such as handbook-v4-2026-09.pdf.
5. List every policy that lives outside the handbook (pension, bonus rules, disciplinary procedure) and decide whether each one is loaded, linked or excluded.

Here is how this looked on an illustrative IT support firm with 42 staff. The handbook said sick pay depended on "length of service" without a table, the remote-working section pointed to an intranet page that no longer existed, and two sections defined the notice period for annual leave differently (one week versus ten working days). Before the bot went live the consultant listed those three conflicts, the owner chose the answer for each, and the handbook was reissued as version 5. Nothing about the AI was difficult. The clean-up took roughly six hours of the consultant's time and was billed as a discovery fee.

The same thinking applies to any client document you load; how to build a company knowledge base AI can answer from goes deeper on the format.

Write the instruction that keeps the bot inside the handbook

Every platform has a place for standing instructions. Use it to set the boundary in plain words. This is a starting version for the illustrative IT support firm; adapt the bracketed parts per client.

You answer questions for employees of [Client name] using ONLY the
employee handbook files provided to you (current version: [v5, dated 1 Sep 2026]).

Rules:
1. If the handbook does not contain the answer, reply: "That isn't covered in
   the handbook. Please ask [adviser name or HR contact] at [route]." Do not
   guess and do not use general knowledge.
2. Quote the section heading and give a one-line answer. Never give legal advice.
3. For these topics, do not answer. Say you cannot help with that here and give
   the contact route: grievances, harassment or bullying, discipline or dismissal,
   redundancy, pregnancy, disability or health conditions, pay disputes,
   whistleblowing, anything about a named person.
4. If the employee says they are upset, unsafe, or unwell, reply with the contact
   route first and nothing else.
5. Keep answers under 90 words. Ask no questions except to clarify which policy
   the employee means.
6. End every answer with: "This is general information from the handbook, not
   advice about your personal situation."

Two points here are worth explaining to a client. First, rule 1 is the one that protects you. A bot that happily answers "do I get paid for the public holiday if I'm on call" from general knowledge is a bot that can contradict the contract. Second, rule 4 matters more than most people expect; employees often type a personal difficulty into a chat box because it feels private. The bot should not try to counsel them.

What a good answer and a bad answer look like

Illustrative test question: "How many days' notice do I need to give for a week off in August?" A good reply from the IT support firm's bot reads:

Section 6.2 (Annual leave): requests for a week or more need at least four weeks' notice and approval from your line manager. Shorter requests need ten working days. This is general information from the handbook, not advice about your personal situation.

A bad reply adds "though managers are usually flexible" or "you may be able to negotiate this". That extra clause is invented, and a tester who sees it should fail the build. Make the failure obvious in your test sheet by marking any sentence not traceable to a section as a fail.

Escalation rules, written as a table the client signs

Escalation is the part buyers are nervous about, so show it on one page. Agree the table in a short meeting with the owner, then store it in the contract. Here is a filled-in version for a client with no dedicated HR person.

The last row is the safety net. Whatever platform you pick, make sure a frustrated employee can reach a person within two turns. On a plan without a helpdesk inbox, that means the bot posts a name and a contact route; on one with an inbox, it passes the chat to the person on the rota. The response targets are what the client will measure you on, so do not promise faster than you can deliver.

Treat anything that touches a protected characteristic or a possible legal claim as a topic where you advise the client to involve a qualified employment-law adviser. The bot's job ends at the contact route.

Disclaimers that employees see

Put the disclaimer in three places: the welcome message, the end of each answer, and the page the chat sits on. Keep it short, because long disclaimers are not read.

Welcome message:
"Hi, I'm [Client]'s handbook assistant. I can answer questions using our
employee handbook (version 5, 1 September 2026). I can't give advice about your
personal situation, and I don't handle complaints or sensitive issues. For those,
contact [name] at [route]."

Footer on the chat page:
"AI assistant. Answers come from the employee handbook only and may be out of
date if policies have changed since the version date above. Chats are logged
and reviewed by [consultancy] to improve the service."

The logging line matters. Employees should know conversations are stored and who reads them, and the client should have decided this with their own data-protection adviser. Ask the client whether staff chat logs count as personal data that their staff privacy notice must cover, and how long the logs should be kept. Raise it; do not decide it for them.

Per-client setup, step by step

This is the sequence I would use for each new client, with realistic time per step. A first client takes longer; the fifth should take about half a day plus testing.

1. Discovery call (45 minutes). Confirm which policies exist, who answers staff questions today, and which topics the client wants refused.
2. Handbook clean-up (2 to 6 hours). Use the pre-upload check above and agree fixes in writing.
3. Build (1 to 2 hours). Create the bot, upload the file, paste the instruction, set the welcome message, turn off any web-browsing or general-knowledge option, and set the contact route.
4. Test (2 hours). Run the 30-question sheet below; fix and re-run until it passes.
5. Soft launch (one week). Give the link to three or four managers and review every chat.
6. Full launch. Announce it with a one-paragraph message from the owner that includes the disclaimer.
7. Month-one review (30 minutes). Read the logs, list unanswered questions, and propose handbook amendments.

Keep each client in a separate bot, separate files and a separate login. Never build one bot that holds several clients' handbooks, even if the platform allows tagging, because one wrong tag leaks one client's pay rules to another client's staff.

The 30-question test sheet

A bot is ready when it passes a fixed set of questions you wrote before looking at its answers. Build the sheet from three buckets and score each reply pass or fail.

Set the pass mark at 100% for the last two buckets and 11 of 12 for the first. A single wrong answer to a sensitive question is a launch blocker. Also test awkward phrasing: typos, slang, a question that mixes a covered and a refused topic ("how much notice do I give, because I'm being bullied and want to leave"). The bot should refuse the whole turn and give the route.

Checking that the numbers are right

For the 12 covered questions, a person who knows the handbook should confirm every figure against the page. In the illustrative IT support firm test, the bot answered a question about overtime with "time and a half". The handbook said "time and a half after 40 hours", and the bot had dropped the condition. It was fixed by adding a rule to quote any threshold in full, and the question went on the permanent sheet.

Pricing the retainer

Price by the work and the risk, not by the software. A simple structure is a setup fee, a monthly retainer, and an hourly rate for anything outside the agreed scope. The numbers below are illustrative; use your own rates.

On those numbers the platform is about 30% of the retainer and your time is the rest, which is the right way round. A client paying $425 a month is paying for the maintenance promise. If you come out below your normal hourly rate once the second and third clients are live, that is a sign the maintenance estimate is too low, not that the software is expensive.

Add scope limits to the proposal: number of policy updates a month included, number of languages (one), and what happens to out-of-scope questions. Link the retainer to a quarterly review, in line with the advice in whether an AI consulting retainer is worth it, so the client sees what they got.

Contract points that protect you

• Source of truth. The handbook is the client's document and the client confirms each version is current. If they change a policy and forget to tell you, the bot's answer is out of date through no fault of yours.
• No advice. The bot gives general information only and is not a substitute for advice about an individual case.
• Data. State what is stored, where, for how long, who can read chat logs, and what is deleted on termination. A short NDA, signed before you receive the handbook and any staff data, is worth having first.
• Your AI wording. Add a clause covering the AI service; how to mention AI use in client contracts gives the sort of text to adapt.
• Take-down. Either side can switch the bot off at any time, and you can do it immediately if a legal change makes a policy unsafe to repeat.

Ask your own legal adviser to review the final wording, since it is your liability being limited.

Running it month to month

The monthly review is the part that justifies the retainer. Open the logs and sort the chats into four piles: answered correctly, answered but poorly, not covered, and sensitive topics routed. Then act on each.

• Not covered. If the same missing policy turns up five times, it is probably worth writing. Offer the client a one-page policy draft as paid extra work.
• Answered poorly. Rewrite the relevant handbook paragraph rather than patching the bot's instruction. The fix then helps humans too.
• Sensitive routed. Check that someone did respond within the target. If not, that is a conversation with the owner, not a bot fix.
• Volume. Compare chats to the plan's allowance. If you are near the limit three months running, move plan or renegotiate.

An illustrative month-one report for the IT support firm might read: 118 chats, 96 answered from the handbook, 14 not covered (nine about the expenses claim deadline, which the handbook gave only in an appendix), 8 routed as sensitive, with all eight acknowledged within a day. The handbook change was a two-line edit, and a good reason for the client to keep the retainer another quarter.

Where a handbook bot goes wrong in practice

Three failures show up again and again. A policy was updated but the old file stayed in the bot, so staff got two answers; delete old versions rather than adding the new one beside them. A manager pasted a private email thread into the chat to ask a question; add a line to the welcome message asking people not to enter personal details or names. And the bot answered confidently from a table it misread; avoid heavily formatted tables in the source and write key thresholds as sentences.

If the bot will also sit on a client's public website, treat it as a separate build with its own instruction, since the audience and the risk are different. The same checks belong in the AI acceptable use policy you keep for your own practice.

If you would rather map the retainer and the first client build with someone else, that is what my AI implementation consultation covers; see the contact page.

Questions HR consultants ask before selling a handbook bot

Should the bot ever answer a question the handbook does not cover?

No. A handbook bot earns trust by saying 'that is not in the handbook' and routing the question to a named adviser. The moment it improvises from general knowledge it can contradict a client's actual policy, and you carry the risk. Switch off any general-knowledge fallback and test with questions you know are missing.

Who owns the chatbot if the client ends the retainer?

Decide in the contract. A common arrangement is that you own the build and configuration, the client owns its handbook and conversation logs, and the bot is switched off and the knowledge files deleted within an agreed number of days of termination. Whatever you choose, write it down before the first upload.

Can employees use the bot to report a grievance?

Treat that as an escalation, not a bot job. The bot should recognise the topic, say it cannot handle it, and give the client's real reporting route. Never let it collect details of an allegation, because the record then sits in a chat log nobody is managing.

How often should the handbook in the bot be refreshed?

Whenever the client changes a policy, plus a scheduled check every quarter. Version the file name, upload the new one, delete the old one and re-run your test sheet. A stale holiday or sickness rule is the commonest way these bots give wrong answers.

Further reads

• Reusing Past Client Work With AI Without Leaking Client Data — Keep one client's material out of another's bot.
• What to Check in an AI Tool's Privacy Policy and Terms — Read a vendor's privacy terms before you upload.
• How to Set Up Human Review for AI Work Without Slowing Down — Set up the review step behind your escalation rules.
• The AI Tool Stack a One-Person Consultancy Actually Needs — Keep your own tool costs in proportion.
• How to Roll Out an AI Policy So Staff Actually Follow It — Help clients announce the bot to staff.
• How to Write an AI Usage Policy for Your Small Business — Give clients a short policy to sit beside the bot.
• Should You Sign an NDA Before Sharing Data With an AI Consultant? — See what an NDA covers when a client shares staff data.
• AI Acceptable Use Policy for a Small Professional Firm — Write the AI policy your own consultancy works to.
• AI for Mediators: Summaries and Agreements That Stay Neutral — A neutrality checklist for AI session notes, a balanced summary prompt, agreement drafting rules and a consent clause for every party.
• Employee Handbook Template for Clients: Build a Master Copy With AI — How an HR consultant builds one master handbook with UK and US modules, a clause library and client fields, plus a checklist for the employment lawyer.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Chatbase and Botpress pricing and help pages, checked October and November 2026; vendor privacy documentation.