"Where's the supplier contract template?", "How do I process a refund if the customer paid by card?", "Who do I ask about building access?" Questions like these land in the team group chat every day. Whoever sees them first answers, usually from memory and not always correctly. An hour later a new hire asks the same thing, because yesterday's answer has scrolled out of sight. This article shows how to run an internal help desk on top of your knowledge base: AI answers from your own policies with a link to the source, questions it cannot answer become tickets for the right owner, and every gap turns into a new knowledge base article. We cover where to start, which topics to hand to AI and which to keep, and how to avoid drowning in setup.
Why a group chat makes a poor help desk
A group chat works for announcements and urgent issues. As a reference desk it fails for three reasons.
- Answers don't accumulate. Something answered yesterday is as hard to find today as it was last week. Chat search returns a dozen mentions, not the right answer.
- Nobody owns the question. If no one replies, the question simply scrolls away. The employee asks again in a private message or guesses.
- The load falls on the same people. The accountant, the office manager and the senior salesperson answer the same questions on repeat and lose focus on their own work.
A help desk doesn't fix this by banning the chat. It fixes it with a different route: every question gets an address, every answer rests on a document, and an unresolved question cannot quietly disappear.
The route of a single question
- Ask. The employee writes to a dedicated bot in Telegram, Slack or Teams, in plain language, with a screenshot if needed.
- Search. The agent looks for the answer in your documents: policies, how-to guides, templates, notes on your software. Under the hood this is retrieval over documents plus answer generation; we explain the mechanics in our article on RAG systems for business.
- Answer with a source. If it finds the answer, it replies briefly and links to the exact section it used, so the employee can check rather than take it on trust.
- Ticket if not found. If the knowledge base has no answer, or two documents contradict each other, the agent does not improvise. It opens a ticket for the topic owner and tells the employee who will reply.
- Close and capture. The owner answers, and the agent drafts a knowledge base article from that answer. A person edits and publishes it, so next time the base answers on its own.
Step five is the one that matters. Without it, the help desk is just another chat with a bot attached. With it, the knowledge base grows from real questions rather than from whatever someone once decided to document.
Which questions go to AI and which don't
| Question type | What AI does | What stays with a person |
|---|---|---|
| "Where is it?" — templates, forms, folder links | Answers immediately with a link | Keeping links current |
| "How do I?" — steps in the CRM, accounting software, refunds | Walks through the guide step by step, with the source | Updating the guide when tools or processes change |
| "Am I allowed to?" — exceptions, extra discounts | Quotes the rule and routes a ticket to the owner | The decision on the exception |
| HR and personal — leave, sick days, pay | General policy only; anything personal goes to HR | Everything about a specific person |
| Complaints and conflicts | Doesn't engage on substance, forwards to a manager | All of it |
The rule of thumb: AI answers where the answer is written down and the same for everyone. Where a decision is needed, or a specific person is involved, it routes rather than advises.
A sample exchange
An illustrative example from a retail business:
Employee: A customer wants to return an item after 20 days. They paid by card. What do I do?
Agent: Under the returns policy, returns are accepted within 14 days. After that, only with the sales manager's approval (section 3.2, "Exceptions"). I've opened a ticket for the sales manager, who handles exceptions during working hours. Meanwhile you can tell the customer the request has been passed on and they'll hear back today. Link: "Returns policy, section 3".
Employee: If it's approved, how do I refund to the card?
Agent: The guide "Card refunds at the till" covers it in six steps: create the return document, attach the customer's request, process the refund on the terminal. Link attached.
The agent didn't decide for the manager, but it didn't leave the employee stuck either. The 14 days is a placeholder; your policy will have its own number.
Getting started: the first few weeks
- Collect questions, not documents. Export the last month or two of the group chat and list every "where", "how" and "can I" question. That is your real topic list, and most of it will be repeats.
- Assign topic owners. One person per topic: returns, purchasing, documents, IT access, HR. The owner answers tickets and keeps that section current.
- Cover the most frequent questions first. You don't need a full company handbook. Short answers to the most common questions are enough, each with an owner and a last-updated date.
- Set access rights. Warehouse staff shouldn't get answers from finance documents, and interns shouldn't see the bonus policy. The agent searches only the sections open to the asker's role.
- Pilot with one team. Two or three weeks with one team will show where the base has gaps and where answers are off. Then roll out further.
- Agree on chat etiquette. "How/where" questions go to the bot; the group chat is for urgent issues and news.
Building our own ERP taught us what most teams discover when they digitise: half the procedures existed only in people's heads (see the case study). A help desk doesn't create that problem; it exposes it. Every unanswered question points to a rule nobody wrote down.
Honest limitations
A knowledge-base help desk is only as good as the knowledge base. If your policies contradict each other, the agent won't fix that; it will say it found two different answers and hand the question to a person. Useful, but not magic.
Maintenance is the second limit. The base goes stale with every process change. Without topic owners and a weekly review of unanswered questions, within a couple of months the agent will confidently quote last quarter's rules.
The third is habit. For the first few weeks people will keep posting in the group chat. It helps when managers reply to "how/where" questions with a pointer to the bot instead of the answer itself.
More on automating internal processes is in our Operations section.
FAQ
How is an internal help desk different from a customer support bot?
The mechanics are similar, the job is not. An internal help desk works on private documents with role-based access and turns unanswered questions into tasks to extend the knowledge base. A customer bot works with public information and hands complex cases to a person.
Do we need a finished knowledge base to start?
No. A short list of answers to the most frequent chat questions is enough. The base then grows from tickets: each human answer becomes a draft article.
Will the AI make answers up?
Any model can, which is why the agent is set up to answer only from a retrieved document and always cite it. With no source, it opens a ticket instead. Answers flagged as wrong by employees go to the topic owner for review.
Where should the knowledge base live?
Wherever you can keep it up to date: Google Docs, Notion, Confluence, an internal wiki. The agent needs access to the current version and to know which sections are open to whom.
What does it cost?
It depends on the volume of documents, the number of teams and which systems need integrating. The first conversation is to scope it for your company: get in touch, or read how we approach AI agents.