How to build an internal Ask HR assistant from policy documents safely
Use this 30-day pilot, permission test, and question-routing matrix to answer repeated HR policy questions from approved sources safely.
HR teams lose time answering questions that already have an approved answer: How many days of leave do I have? Where is the expense policy? Which form starts parental leave? What applies to remote work? The expensive part is rarely writing one reply. It is reconstructing the same answer, checking whether the policy changed, and explaining where the source lives.
An internal Ask HR assistant can shorten that path, but only if it is treated as a governed knowledge service. It should answer reusable policy questions from approved documents, show the source, respect access rules, and stop when a question depends on an employee record, a legal interpretation, or a sensitive case.
This guide gives an HR operations leader a 30-day pilot that can run on real policy documents without pretending that a knowledge assistant is an HRIS, payroll system, or case-management platform.
Start with the business job
The job is not “launch an HR chatbot.” The job is to reduce repeated policy interruptions while helping employees reach the current approved answer faster.
Microsoft describes employee self-service agents that answer policy questions and escalate complex or sensitive requests to HR specialists. SAP likewise describes an integrated knowledge base that presents relevant HR policy information and reduces the need to create HR tickets. Both point to the same operating boundary: reusable information can be self-service; personal transactions and sensitive judgement still need the owning system and people.
Define the pilot in one sentence:
Employees in the pilot group can ask about approved policies and procedures, receive a sourced answer, and know when to contact HR.
That promise is narrow enough to test and useful enough to save time.
Route every question into one of three lanes
Before connecting documents, agree on the answer boundary.
| Lane | Example | Expected behavior |
|---|---|---|
| Approved knowledge | “What does our travel-expense policy cover?” | Answer from the current approved policy and cite it. |
| Personal record or transaction | “How many leave days do I have left?” | Explain where to check or whom to contact; do not invent a personal balance. |
| Sensitive case or judgement | “Is my manager treating me unfairly?” | Do not decide the case; direct the employee to the confidential HR route. |
This matrix prevents the most common scope error: connecting policy PDFs and then expecting the assistant to perform HR operations it cannot safely complete.
Use the same rule for payroll, benefits, performance, accommodations, grievances, investigations, medical information, and legal questions. A policy may explain the general process. The individual outcome still belongs to the authorized HR workflow.
Week 1: collect real demand and set boundaries
Start with questions, not folders. Ask the HR team and a small employee group for the repeated questions from the last month. Include email, chat, meetings, onboarding sessions, and existing tickets if available.
Create a simple list with:
- the employee’s original wording;
- category, such as leave, expenses, benefits, remote work, conduct, or onboarding;
- frequency or a rough high/medium/low signal;
- the current source used by HR;
- whether the answer changes by country, contract, role, location, or seniority;
- whether the question belongs in approved knowledge, a personal system, or a confidential case.
Choose 20 to 30 questions for the first pilot. Prefer frequent, low-risk questions with one authoritative answer. Do not begin with every possible exception.
Write the exclusions next to the goal. For example: no payroll calculations, no access to medical or performance records, no employment-law advice, no grievance decisions, and no changes to employee data.
Week 2: build a controlled policy source pack
Do not connect the whole HR drive. Build a small source register for the questions selected in week 1.
Copy this structure into a spreadsheet:
| Field | What to record |
|---|---|
| Policy question | The employee wording this source should answer |
| Authoritative document | One approved handbook, policy, procedure, or FAQ |
| Owner | Person accountable for the rule |
| Audience | Everyone, managers, one country, one entity, or another defined group |
| Effective date | When the rule became current |
| Review date | When the owner will check it again |
| Supersedes | The old version that must no longer appear |
| Escalation route | Where exceptions or personal cases go |
Remove drafts, duplicate copies, expired calendars, and policies without an owner. If two documents conflict, the assistant is not ready for that question. HR must decide which source governs before the question enters the pilot.
Permissions belong in this step. A general holiday calendar may be company-wide. Manager guidance, compensation procedures, investigation material, or medical documents should not become visible merely because they sit near a public policy in the same folder.
Polp’s current Google Drive workflow lets an administrator select folders for knowledge-base ingestion instead of exposing every file by default. Retrieval is filtered by organization and document-type permissions, so the pilot should test with the same role differences employees will have in real use.
Week 3: test answers, sources, permissions, and refusals
Turn the initial question list into a small evaluation set. Each test should contain the question, expected source, minimum acceptable answer, expected audience, and required escalation behavior.
Include these test classes:
- A direct question using the policy’s own wording.
- A natural employee phrasing that does not repeat the document title.
- A question with two relevant policies where one is newer.
- A question asked by an authorized employee.
- The same question asked by someone without access.
- A personal-record question the assistant must not answer from generic policy.
- A sensitive case that must go to HR.
- A question for which no approved source exists.
- A misleading question that assumes a benefit or rule the company does not offer.
- A Spanish and an English formulation if the workforce uses both languages.
Review more than fluency. For every result, check:
- Did the answer use the correct current source?
- Does the cited passage support the material claim?
- Did an older or restricted document appear?
- Did the assistant admit when evidence was missing?
- Was the next action clear when HR involvement was required?
A polished answer with the wrong policy is a failed test. A clear refusal with the correct HR route can be a successful one.
Week 4: pilot with one group and measure outcomes
Launch to one department, location, or employee cohort. Explain what the assistant can answer, what it cannot do, where sources appear, and how to report a bad answer.
Track a small scorecard for four weeks:
- policy questions asked;
- answers with the expected source;
- answers accepted by the employee or HR reviewer;
- questions that correctly moved to HR;
- questions with missing or conflicting knowledge;
- repeat contacts after an answer;
- time HR spends reviewing and correcting the system.
Do not call every avoided email a successful resolution. A person may abandon the search or ask someone privately. Use follow-up signals, repeat questions, feedback, and a sample reviewed by HR before claiming saved time.
Microsoft’s HR self-service scenario also treats feedback as an input for improving unclear policy sections. That is the useful loop: questions reveal where policy wording, ownership, permissions, or coverage needs work.
Polp can surface low-confidence answers, missing-source answers, and negative feedback as knowledge gaps for administrators. The HR owner can then fix the authoritative document, reconnect the right source, correct access, or record that the question must remain an HR escalation. The weekly knowledge-gap review gives that queue a lightweight operating rhythm.
What Polp does in this workflow
Polp is the knowledge layer in this pilot. It can ingest selected documents and Google Drive folders, retrieve within organization and document-type permissions, answer with visible sources, and expose gaps when approved evidence is missing. Employees can use the source-checking guidance to verify important answers.
Polp does not replace the HRIS, payroll platform, benefits administrator, or confidential case process. Personal balances, data changes, approvals, complaints, investigations, and legal decisions stay in their owning systems and teams.
That separation is a strength. HR gets a practical way to reduce repetitive policy work without placing sensitive employee operations inside an unbounded chatbot.
The launch decision
At the end of 30 days, expand only if the pilot shows:
- one current owner for each supported policy question;
- correct answers and citations across the test set;
- no restricted-source leakage in role-pair tests;
- reliable refusal and escalation for personal or sensitive cases;
- an active process for fixing missing and outdated knowledge;
- evidence that employees resolve more questions without creating hidden rework.
If those conditions are not met, narrow the source pack and fix the owner problem. Adding more documents will not repair unclear policy, duplicate versions, or missing permissions.
For a pilot using selected company policies and real employee questions, request a Polp demo. If your approved policies already live in Drive, review the Google Drive integration before choosing the first folders.