What this automation does
Employees ask in the chat tool they already use and get an answer that quotes the policy, or a fast handoff to a person.
- Ask. An employee messages the assistant in Slack or Teams, for example “How much bereavement leave do we get for a grandparent?”
- Identify. The assistant maps the chat user to the HRIS record and loads only the attributes it needs: work state, entity, employment type, exempt status.
- Screen for sensitive topics. A classifier checks the question first. Sensitive topics skip retrieval entirely and go to the HR handoff path.
- Retrieve. The assistant searches only the policy chunks that apply to this employee’s attributes and pulls the top matches.
- Answer with a citation. The model answers in plain language, quotes the relevant line and links to the section. If the retrieved text does not answer the question, it says so.
- Personal lookups. “How much PTO do I have?” is answered from the HRIS for that employee only, or with a link to self-service if you prefer not to expose balances in chat.
- Learn from gaps. Unanswered questions are collected weekly so HR can see where the handbook is silent or unclear.
When it is worth building
The signal is an HR inbox full of the same questions: holidays, PTO accrual, expense rules, benefits enrollment windows, remote work equipment. If one HR generalist spends a day or more a week answering questions already covered by written policy, an assistant gives that time back.
It is also worth building when you operate in several states, because that is where a static FAQ page falls apart. It is not worth building if your policies are not written down or are badly out of date. Fix the handbook first; an assistant will repeat whatever the documents say.
Architecture
| Layer | What we use | Why |
|---|---|---|
| Interface | Slack app or Microsoft Teams bot | No new tool to learn, and identity comes from the workspace login |
| Source documents | SharePoint, Google Drive or Notion, one approved folder | HR controls exactly what is indexed by controlling one location |
| Applicability tags | Metadata per chunk for state, entity and employment type | Filtering before search is more reliable than asking the model to sort it out |
| Retrieval | Postgres with pgvector, or Pinecone for larger libraries | Most HR libraries are small enough that pgvector is simpler to run |
| Employee context | BambooHR or your HRIS API, read-only and field-limited | The assistant knows where someone works without seeing their pay or reviews |
| Model | Claude or OpenAI’s GPT models through the API | Strong at answering only from supplied text and citing it |
| Handoff | Private case in your HR ticketing tool or a direct message to HR | Sensitive questions reach a person quickly and stay confidential |
| Logging | Question, retrieved chunks, answer and feedback, with restricted access | Lets HR audit answers; treat logs as HR records |
For the general pattern behind this, see our guide to building a RAG knowledge assistant. An HR assistant is that pattern with stricter access control and a mandatory escalation path.
How to build it
- Inventory the documents. Collect the current handbook, state addenda, benefits summaries and policy memos. Remove drafts, superseded versions and anything with individual employee data.
- Chunk by policy section. Split on headings so each chunk is one rule with its title. Tag each chunk with applicability and an effective date.
- Define the sensitive-topic list with HR. Write examples of each category, including indirect phrasing such as “my manager keeps commenting on my appearance”. Test the classifier against 100 or more examples before launch.
- Write the answer rules. Answer only from retrieved text, quote it, link it, never speculate about legal rights, and say “I don’t know, here is who to ask” when the text is silent.
- Connect identity and the HRIS. Map workspace users to employee records and limit the fields the assistant can read.
- Pilot with one department. Have HR review every answer for two weeks and mark each one correct, incomplete or wrong.
- Roll out with a clear notice. Tell employees what the assistant is, that conversations are logged and reviewed by HR, and that it does not replace talking to a person.
What breaks in production
- Stale policy versions. Someone saves last year’s handbook PDF into the indexed folder. Index only from one controlled location and show the effective date in answers.
- Confidential disclosures in the wrong place. An employee describes a harassment incident in a public channel where the bot is mentioned. The bot should move the conversation to a private message and alert HR without repeating details.
- Answers that sound like promises. “You are entitled to” reads as a commitment. Phrase answers as what the policy says, and keep your handbook’s disclaimer language intact.
- Questions about other people. “Is Sam out next week?” should be refused unless the policy and the data allow it.
- Year-specific facts. Holiday calendars, open enrollment dates and contribution limits change every year. Tag them with a year and set a reminder to replace them.
- Over-escalation. If the classifier is too cautious, ordinary questions about PTO go to HR and the assistant saves no time. Tune with real examples and review weekly.
Timeline and cost
A single-entity assistant in Slack or Teams usually takes three to six weeks including the HR review period. Lumeor builds this in the range of $12,000 to $30,000, depending on the number of states and entities, whether personal HRIS lookups are included, and how much reporting HR wants. Model and hosting costs for a company of a few hundred people are typically small; see what a RAG knowledge assistant costs for how running costs scale.
How to measure success
- Share of questions answered without an HR follow-up
- HR-graded accuracy on a weekly sample of answers
- Sensitive-topic recall: of the questions HR judges sensitive, how many the assistant routed correctly
- Median time to answer, compared with the old HR inbox
- Top unanswered topics each month, and how quickly the handbook is updated to cover them
- Employee thumbs-up rate, read alongside the accuracy sample rather than on its own