The onboarding assistant that knows your policies
The problem
Every new joiner asks the same forty questions. How do I book leave. What's the expense limit. Who approves a purchase. Where's the VPN guide. Is Friday remote.
The answers exist — in Confluence, a shared drive, an old PDF, and three people's heads. Every answer costs a senior person fifteen minutes and an interruption.
What it costs today:
- Interruption tax. Forty questions × every new joiner × the person who happens to be nearest.
- Inconsistency. Two people give two different answers about the expense limit, and both are confident.
- Documentation that rots unnoticed. Nobody discovers the policy page is two years out of date until someone acts on it.
What good looks like
Questions get answered in the tool people already have open, from the documents you actually maintain, with a link to the source — and the assistant says "I don't know" instead of inventing a policy, because an invented HR policy is a genuine liability.
What you build
Peer "Workplace Assistant"
├── Data sources Confluence space, HR drive, IT runbooks
├── Topics leave · expenses · IT access · facilities
├── Channels Slack (channel prompt: be brief, you're in a channel)
│ Teams (same peer, different identity)
│ Web Chat (for the intranet)
└── Evaluation a question set you re-run after every changePieces used: a Peer, Data sources, Topics, Channels, Evaluation.
This build has no flow in it at all. Not everything needs one — a question-answering assistant is a Peer, and reaching for a flow here adds machinery without adding value.
Implementation
1. Index the sources you actually maintain
Connect data sources for the systems that hold the answers — Confluence, SharePoint, a drive, or uploaded files.
Be selective. Indexing everything makes retrieval worse, not better: a decommissioned 2019 handbook competes with the current one and sometimes wins. Index what is maintained, and keep it synced.
2. Write a prompt that permits ignorance
The default failure of an internal assistant is not rudeness, it is fluency. Say so explicitly:
You answer questions about how this company works, using only the connected
sources.
If the sources do not contain the answer, say you don't know and name who to
ask. Never infer a policy from what is normal at other companies.
Always link the page you used. Keep answers to a few sentences unless asked
to expand.Turn on Enable RAG, and consider Include Datasources as Tools so the Peer can target one source rather than searching everything at once — useful when IT runbooks and HR policy would otherwise blur together.
3. Fence it with topics
Add Topics for the areas it should handle, and configure off-topic handling for everything else. This is what stops a workplace assistant from confidently answering a legal or medical question because someone asked it in the same channel.
4. Put it where people already are
Add channels:
- Slack or Microsoft Teams — whichever your company lives in. Teams uses a peer-per-app model, so the assistant appears as its own chat target rather than one shared bot.
- Web Chat on the intranet, for people who look things up rather than ask.
Give each channel its own system prompt. This is the feature that saves you from maintaining three Peers:
| Channel | Channel prompt |
|---|---|
| Slack | "You are answering in a public channel. Two or three sentences, then the link." |
| Web Chat | "You are answering on the intranet. Give the full steps, not just the link." |
Same Peer, same knowledge, different manners.
5. Measure it, then change it
Build an Evaluation set of thirty real questions with known-good answers — including the ones you want it to refuse.
Run it before and after every prompt or source change. Without this, prompt tuning is vibes: someone changes a sentence, the answers feel better, and nobody notices that three previously correct answers regressed.
Use AI Analysis on real conversations to find the questions people actually ask, which is never quite the list you predicted.
What happens at run time
Someone types in Slack. The Peer retrieves from the indexed sources, answers in two sentences with a link, and stays quiet when it has nothing — telling the person who to ask instead.
The same question on the intranet gets the fuller version, because the channel prompt says so.
What you get
- A drop in repeat questions, measurable in the channel itself.
- Consistent answers — everyone gets the current expense limit, not the one someone remembers.
- A live signal on documentation quality. Questions the assistant cannot answer are the pages you are missing, and they arrive as a list instead of as complaints.
Extensions
Let it do the thing, not just describe it. Attach tools so "book me Friday off" creates the request, with agent approval so the Peer proposes and the human confirms.
Escalate into a real ticket. Attach a flow as a tool that opens an IT ticket when the assistant cannot help — passing the conversation context so the person does not have to explain twice.
Remember the person. Turn on User Memory so the assistant knows someone is in the Berlin office and answers the leave question with the right policy without being asked twice.
Add a voice. Enable voice conversation for people who would rather ask than type.
Pitfalls
Indexing everything. Retrieval quality falls as the index fills with material nobody maintains. Prune it.
Fixing retrieval problems in the prompt. If answers are wrong because the right document was never retrieved, no amount of prompt editing helps. Check search weights, chunking and the threshold first.
One Peer per channel. Duplicating the Peer to change its tone means every future knowledge change has to be made three times. Use channel prompts.
No refusal test. If your evaluation set only contains questions it should answer, you will never notice the day it starts answering everything.