Build an internal Q&A agent for your team

Why this matters

Every team has the same questions answered five times. “Where’s the expense policy?” “How do we handle PTO requests?” “What’s our approval threshold for vendor contracts?” “Where’s the brand kit?” The questions are easy. The answers exist somewhere. But the path from question to answer routes through one person (usually you, sometimes ops, or sometimes whoever’s been around the longest), and that person spends an hour a week answering the same things.

The fix isn’t a wiki nobody reads. It’s an agent that reads your wiki for them and answers in the channel they’re already in. Today, you build that. By the end, you’ve taken the most-asked-five questions off person-to-person Slack DMs and put them on a system that answers them at 11pm on a Sunday without complaining.

Quick start

1. Sign in — open Auxot in your browser and log in.

2. Open chat with the Admin Agent — click Chat in the left menu, make sure the agent picker reads “Admin Agent.”

3. Build the Internal Q&A agent — paste this:

   I want an "Internal Q&A" agent. Its job is answering questions from my team about how this company works — policies, processes, where things are, and who handles what. It should answer from our actual policies and handbook (I'll attach as context files), cite which document it's drawing from, and explicitly say "I don't know — ask [a person]" when it can't answer from the source material. Build it.
   

4. Answer the Admin Agent’s questions — what handbook, standard operating procedure, or policy docs to attach, what tone (formal, casual, or in-between), and who to escalate to when it doesn’t know.

5. Test on real recent questions — paste in 3–5 questions your team actually asked in the last month. Verify the answers are right and the citations are real.

Done? The agent is in Settings → Agents and selectable from the agent picker. Optional next step: wire it to a Slack channel so the team can ask it without leaving Slack (Connect Slack to your agents).

The agent can do that?

Internal Q&A goes wrong when the agent answers confidently from training data instead of your actual policies. These three prompts close that gap.

1. Have the agent cite which document it’s drawing from, every time

Open chat with Internal Q&A and ask:

For every answer you give, cite which document and section you're pulling from. Format: end the answer with "Source: [document name], section [name or page]." If you're combining info from two sources, cite both. Don't answer from general knowledge if the answer isn't in the attached docs — say "I don't see this in our docs" instead.

Why it’s non-obvious: Most internal Q&A agents fail trust because the answers look right but you can’t verify them quickly. Forcing a citation does two things: (1) the asker can verify in one click, and (2) it constrains the agent to actually pull from your docs instead of confidently making something up. The citation is what makes the answer trustworthy.

2. Have the agent flag when it doesn’t know, instead of guessing

Add this to the same chat:

Strict rule: if the answer isn't clearly in the attached context files, you say "I don't see this in our docs — ask [person or team] who handles this." Never guess. Never hedge with general advice that sounds right. The wrong answer about our policy is worse than no answer.

Why it’s non-obvious: AI models default to being helpful, which means they default to answering. For internal Q&A, refusal is the right answer when the docs don’t cover something. The “ask [person]” routing turns a dead-end into a productive handoff: the asker gets to the right human instead of being told something that isn’t true.

3. Have the Admin Agent show which questions keep getting asked

Inverted-usage move. Once the Q&A agent has been running for a few weeks, head back to chat with the Admin Agent:

Look at the questions my "Internal Q&A" agent has been getting. What questions come up most often? For each top-asked question: is the answer clearly in our docs (so the agent's handling it well), or is the answer thin/missing (so we should update our docs)? Recommend specific docs to update.

Why it’s non-obvious: The most-asked questions are diagnostic of where your documentation has gaps. Most teams don’t notice: the questions go around the wiki instead of through it. The agent’s logs make the gaps visible. Update the docs once, the agent answers correctly forever, the question stops being asked.

Go deeper

The “fake employees” framing: naming matters here more than for other agents

Of all the agents you’ll build, the internal Q&A agent is the one your team will most directly experience as a coworker. Naming matters. Don’t call it “Q&A Bot” or “Help Agent”: call it something that fits how your team talks. “Ask Aria” (after the operations lead it’s replacing). “Handbook Helper.” “Bramble” (we don’t know either, but your team will have its own joke). The naming is what makes the team comfortable asking it questions instead of pinging a human.

Choose your context files like you’re hiring an HR generalist

What would the new HR person need to know on their first day to answer team questions correctly? That’s your context file list:

• The employee handbook.
• Expense and travel policies.
• PTO / time-off policy.
• Benefits summary.
• IT and security policies (where to file a help ticket, password reset, etc.).
• Org chart or “who handles what” doc.
• Any SOPs for routine processes (onboarding, offboarding, hiring approval, or vendor contracts).

Don’t attach things that aren’t policy: meeting notes, drafts, or personal docs. The agent will quote them as if they’re authoritative. Add your first context file covers context files.

The “ask [person]” fallback list

power move 2 routes unknown questions to a person. That list of people needs to be specific:

• HR questions → HR lead’s name and contact.
• IT / access questions → IT lead.
• Finance / expense questions → finance lead.
• Anything legal → legal contact (in-house or external).
• Anything strategic / about the business → you.

Add this list to the agent’s description directly, not as a context file: it’s structural, not knowledge. The agent should reach for it the moment it doesn’t know.

Where you put the agent matters

Three reasonable places, ranked by how much your team will actually use it:

1. Slack (Connect Slack to your agents): the agent is invokable in a #ask-the-team channel, or as a DM. By far the highest-usage option for most teams.
2. A dedicated chat thread in Auxot. Lower usage: your team has to remember to open Auxot.
3. Your intranet (via API). Possible but rarely worth the engineering work.

Pick Slack unless you have a strong reason not to. The whole point is reducing friction; if asking the agent takes more clicks than asking a human in Slack, the agent loses.

Troubleshooting

• The agent gives confident-sounding wrong answers. It’s pulling from general knowledge instead of your docs. power move 1 (citations required) plus power move 2 (refuse if not in docs) fix this. If they still happen, your context files don’t actually contain the answers and the agent is filling the gap from training. The fix is in the docs.
• The agent says “I don’t see this in our docs” too often, even for things that ARE in the docs. Either the relevant doc isn’t attached, or the doc’s structure makes retrieval hard (one giant unsearchable PDF, headers without keywords). Break docs into shorter, well-structured files. Add your first context file has guidance.
• The team isn’t using it. Either the wrong delivery channel (see “Where you put the agent”) or the wrong name (something that feels like “official channel” instead of “easy ask”). Try renaming and re-introducing.
• Sensitive questions are being asked of the agent (compensation, HR complaints). That’s a feature and a risk. The agent should refuse those questions explicitly: “For compensation questions, please reach out to [HR lead] directly — that’s not something I’m equipped to discuss.” Add this to its description.

Variations & edge cases

• Free tier: the agent works at all tiers. Multi-team setups (Set up multi-team isolation) might want one Q&A agent per team if policies differ.
• Multi-language teams: specify in the agent’s description that it should answer in the language the question was asked. Modern models handle this well.
• Highly regulated industries: for compliance-relevant answers, attach a disclaimer to the agent’s description (“for medical, legal, or compliance questions, always verify with [specific person] — my answers are informational, not authoritative”) and consider whether human-in-the-loop (Set up a content review pipeline) belongs on certain question types.
• When NOT to use this: if your “policies” exist mostly as things people remember and tell each other instead of in docs, the agent has nothing to read. Build the docs first; the agent can wait.

Walkthrough

Step 1: Audit your existing docs

Before building the agent, take 20 minutes and ask: what docs do we have, and which ones answer actual recurring questions? Make a list. If a question your team asks weekly isn’t answered in any doc, write a one-paragraph answer and save it in the docs folder. The agent will only be as good as the docs you give it.

Step 2: Open chat with the Admin Agent

Click Chat in the left menu. Make sure the agent picker at the top reads “Admin Agent.”

Step 3: Build the agent

Paste this:

Build me an "Internal Q&A" agent for my team. Its job:

1. Answer questions from my team about how this company works — policies, processes, and who handles what.
2. Pull only from the attached context files; never invent answers from general knowledge.
3. Cite the source document and section in every answer.
4. If the answer isn't in our docs, say "I don't see this in our docs — ask [name]" and route to the right person from this list: [paste the routing list].
5. Refuse compensation, HR complaints, and other sensitive questions explicitly, with a redirect to HR.
6. Tone: [formal / casual / in-between, your choice].

Attach the context files I'll provide. Make it ready to use.

The Admin Agent will ask which context files (the docs from Step 1), the exact routing list, and your preferred tone.

Step 4: Test on real questions your team asked recently

Pick 5 actual questions from the last month: Slack DMs, water-cooler asks, and the recurring ones. Paste them one at a time into chat with the agent. Verify:

• Is the answer right?
• Is the citation real (the document and section it cites actually contain the answer)?
• For questions not in the docs, does it correctly say “I don’t see this — ask [name]”?
• For sensitive questions, does it refuse and redirect?

For every miss, update either the agent’s description (if the rule is wrong) or the context files (if the answer should be in the docs but isn’t).

Step 5: Wire it to where your team actually asks

For most teams, this means Slack (Connect Slack to your agents). Set up the agent in a channel like #ask-the-team or as a DM target. Announce it to your team:

“Heads up — we’ve got a new agent called [name] that answers most policy questions from our handbook. Ask it before pinging me. If it doesn’t know, it’ll tell you who to ask. It cites its sources so you can verify. Be skeptical for the first week — flag any wrong answers and we’ll fix them.”

The “be skeptical” framing matters. You want them double-checking until trust is earned, not rubber-stamping.

Step 6: Run the gap analysis monthly

Use power move 3 once a month. The Admin Agent reads the questions that hit “I don’t know” and tells you which docs to update. Update the docs. The agent answers correctly forever. The question stops being asked.

What’s next

• → Turn internal office hours into FAQ updates. Converts recurring questions into handbook rows that this agent cites, keeping docs current between major rewrites.
• → Enforce a living style guide and terminology. Shared glossary: check outbound or internal drafts against the same vocabulary this agent teaches.
• → Add your first context file. Your handbook, SOPs, and policies live here.
• → Connect Slack to your agents. Slack is where your team asks; put the agent there.
• → Audit and clean up your agents. Once you have several agents, audit which questions go to which.
• → Onboard a new teammate to your agents. Internal Q&A is the agent every new hire should meet first.
• → Turn incident lessons into updated runbooks. After incidents, paste lessons into the handbook context this agent already cites.
• → Turn meeting notes into action items and decisions. Meeting decisions and owners become handbook rows this agent can cite next week.

Reference

• Citations are the trust layer: every answer ends with a real source pointer.
• “I don’t know” is the right answer when the docs don’t cover something. Never guess.
• Confidently-wrong answers about policy destroy trust faster than any other failure mode. power moves 1 and 2 are the safeguards.
• Where the agent lives = Slack, almost always. Friction kills internal Q&A faster than anything else.
• Most-asked questions are documentation gaps. power move 3 surfaces them; the fix is updating the docs, not the agent.
• Refuse sensitive categories explicitly: compensation, HR complaints, anything that needs a human.
• See also: Review documents against your standard terms, Turn meeting notes into action items and decisions, Add your first context file, Connect Slack to your agents, Onboard a new teammate to your agents.