The HR head keeps answering the same question in different places
An HR head may begin the morning with a leave question on WhatsApp, find a travel claim doubt in email and then hear that a plant supervisor has shared an old holiday list with a new employee. None of these questions is hard on its own. The problem is that the answer lives in a policy PDF, an induction deck, somebody's inbox or the memory of one experienced team member.
This becomes more difficult when a company has a head office in Mumbai, a plant near Pune or Nashik and sales people travelling across Maharashtra. A policy may be the same for everyone, but context matters. A probation employee may have a different rule. A plant shift may follow a different calendar. A manager may need an approval that the general policy does not cover. HR spends time finding the right version, explaining it again and correcting confident but outdated replies.
The answer is not to let a public chatbot guess. It is to build a controlled company knowledge base AI system. For an HR head, that means putting approved information in a well-organised Notion workspace, connecting an AI assistant to that material and making the assistant show where each answer came from. Routine answers can move faster. Sensitive or unusual cases still go to a person.
What the system looks like
The system has three simple parts. Notion holds the current source material. An AI assistant finds the relevant pages and prepares an answer in plain language. An approval rule decides whether the answer can be shown immediately or must wait for HR review.
Notion is not useful merely because pages can be created quickly. Its value comes from order. Each policy page needs an owner, an effective date, a last-reviewed date and a clear audience. Superseded pages must be archived. Permissions must prevent an employee-facing assistant from reading confidential case notes or leadership-only material.
The AI layer should not be treated as a second policy writer. Its job is to retrieve approved facts, combine only what is needed and say when the source does not contain an answer. A useful response can include a short explanation, a link to the source page, the relevant form and the name of the team that handles exceptions.
How one question moves through the workflow
- An employee asks a question. The question can arrive through an internal portal, a company chat or a form chosen for the pilot. One channel keeps the first version easier to control.
- The assistant identifies the subject. It separates a general leave-policy question from a personal leave-balance case. This matters because the second question needs employee data and stronger controls.
- It searches approved Notion pages. The assistant reads only the collection that HR has marked for this purpose. It does not search random messages or the open internet for company rules.
- It prepares an answer with its source. The reply uses everyday language and points to the policy page and revision date. If sources conflict or no answer is found, it does not fill the gap.
- A person approves when required. The HR team reviews exceptions, personal cases and any answer below the agreed confidence level. The final decision stays with the authorised person.
What HR should put into Notion first
A common mistake is to upload every document and expect useful answers. A better pilot begins with a small group of frequent, low-risk questions. HR can start with office timings, the published holiday list, the general leave application process, standard reimbursement steps, induction checklists and links to approved forms.
Each page should answer one subject clearly. A page called "Leave" that mixes policy, payroll rules, manager guidance and special cases is difficult for employees and AI alike. Smaller pages with clear headings make the answer easier to check. They also show who must update a page when a rule changes.
- Give every page a content owner and review date.
- Mark the audience, such as all employees, managers or HR only.
- Keep one current version and archive older versions away from the assistant.
- Add examples only when HR has approved them as valid guidance.
- State the escalation path for questions that the page does not cover.
This clean-up is useful even before AI is connected. It shows where two documents disagree, where a form is missing and where an unwritten custom has been mistaken for company policy.
A practical example for an HR head
Suppose a new employee asks, "Can I take leave during probation, and which form should I use?" The assistant checks the approved probation and leave pages. It prepares a short answer, links the correct form and names the approver. If the policy gives a general rule but the employee is asking about a medical emergency, the workflow sends the case to HR instead of applying the general answer blindly.
Now suppose a manager asks whether an employee's unused leave can be paid out after resignation. That question may depend on company rules, employment terms, payroll records and current legal advice. The assistant can collect the relevant approved references for the HR head, but it must not make the final interpretation or promise an amount.
The difference is important. Good automation reduces searching and repeated typing. It does not remove judgement from a decision that affects a person.
How to roll it out in 30 days
Days 1 to 7: choose one boundary
List the questions that consume the most HR time and choose one low-risk group for the pilot. Define what is included, what is excluded and who owns the answers. Record a baseline: weekly question volume, average response time and the number of cases passed between people before an answer is given.
Days 8 to 14: clean the source
Move the approved material into a simple Notion structure. Remove duplicates, name the owner of every page and add review dates. Ask HR and one manager to test whether a person can find the right answer without AI. If people cannot understand the source, the assistant will not repair it reliably.
Days 15 to 21: build and test the workflow
Connect the assistant only to the approved pages. Create sample questions with spelling mistakes, incomplete details, mixed Hindi and English, and cases that should be refused or escalated. Check whether every response uses the source, shows uncertainty and follows the approval rule. Keep a log of the question, source, draft, approver and final answer.
Days 22 to 30: run a small supervised pilot
Release the system to one location, function or employee group. Let HR review answers during the pilot. Hold a short weekly review to find missing pages and unclear wording. At the end of the month, compare response time, repeat questions, correction rate and the percentage of cases that needed escalation. Expand only when the answers are accurate and the content owner can maintain the system.
What success should mean
The goal is not the largest number of AI answers. The useful measures are fewer repeated queries, shorter waiting time, fewer outdated documents in circulation and more consistent escalation. HR should also watch the percentage of answers changed by reviewers. A high change rate usually points to poor source material, weak instructions or a pilot scope that is too broad.
Employees should have a simple way to say that an answer was unclear or wrong. That feedback belongs in the review queue. It should lead to a source-page correction where needed, not just a one-time rewrite of the reply.
What AI must never be trusted with
An HR knowledge assistant must never decide a disciplinary action, harassment complaint, termination, compensation change, medical accommodation or legal interpretation. It must not expose one employee's information to another person. It must not invent a policy when the source is missing, silently choose between conflicting documents or present an old page as current.
Access control deserves the same attention as answer quality. Payroll files, investigation notes, medical information and personal identifiers should not enter the employee-facing knowledge collection. The company should decide what is logged, who can see the logs and how long they are retained. A named HR owner must be able to pause the assistant immediately if a source changes or a wrong answer appears.
The safest rule is clear: AI can find, summarise and draft from approved material. An authorised person owns exceptions, sensitive cases and decisions.
Build the habit, not just the bot
A useful knowledge system depends on people maintaining it. Page owners need reminders. HR needs a regular review meeting. Managers need to stop circulating downloaded copies after a policy is updated. The assistant should point people back to the controlled source instead of creating another pile of documents.
The Future Corporate builds practical AI systems around real company workflows, with approval and supervision designed into the work. The company also runs its own operations on more than a dozen AI agents with a supervising agent, a CRM built with AI, a WhatsApp enquiry agent and a Telegram assistant. This operating experience helps separate a useful workflow from an impressive demonstration.
Founder Avinash Chate has trained teams at more than 80 organisations. For an HR head, the starting point is still modest: one trusted source, one narrow question set and one accountable owner. Companies can review corporate AI training for teams, see AI ideas by department and learn about the people behind the work.
Frequently asked questions
What is a company knowledge base AI system?
It is an approved collection of company information, such as policies, process notes and standard answers, connected to an AI assistant that finds the relevant source and prepares a clear response. The source remains controlled by the company, and a person reviews sensitive answers.
Why use Notion for an HR knowledge base?
Notion can give HR one organised place for current policies, frequently asked questions, forms and process owners. It is useful when the team sets permissions, names a content owner and keeps revision dates visible.
Can an AI assistant answer every employee question automatically?
No. It can answer low-risk questions from approved material or prepare a draft. Personal cases, disciplinary matters, legal interpretation, payroll changes and exceptions should always go to an authorised person.
Can an HR team pilot this system in 30 days?
Yes. A 30-day pilot can cover one narrow set of questions, clean the source material, build the answer workflow, test difficult cases and release it to a small group with human approval and clear measures.