An Internal AI Assistant for Staff Questions
By CodexierPublished 5 min read
Every company has questions that get asked again and again: how to report sick, where the travel policy is, what the routine is when a customer complains. An internal AI assistant answers them from your own handbook and routines, around the clock, and points to the source. This guide covers which questions are worth automating, how to feed the assistant, how to keep people from seeing what they should not, and how to tell if anyone uses it.
Which internal questions repeat
Before building anything, ask the people who get interrupted. Office managers, team leads and whoever handles payroll can usually list twenty recurring questions from memory. Typical clusters look like this:
- HR and leave: sick reporting, vacation requests, parental leave routines, the notice period in the collective agreement.
- Money: expense claims, which card to use, how receipts are submitted in the expense tool.
- Operations: opening and closing routines, key and alarm handling, supplier contacts.
- Customers: complaint handling, return rules, what we promise on delivery times.
- IT: resetting passwords, connecting to the printer, what to do with a suspicious email.
Sources: handbook, routines, policies
The assistant is only as good as the documents behind it. It works by retrieving the most relevant passages for each question and letting the language model phrase an answer from them, so contradictions and outdated files come straight through to the reply. A clean-up round before launch is part of the job, not an extra.
| Source | Include? | Watch out for |
|---|---|---|
| Staff handbook | Yes, the core | Old versions lying around in shared folders |
| Written routines and checklists | Yes | Routines that differ between sites or teams |
| Policies (IT, travel, GDPR) | Yes | Legal wording that needs a plain-language summary next to it |
| Collective agreement | Yes, with care | The assistant should quote and link, not interpret edge cases |
| Personnel files, salaries, health data | No | Never index documents about individuals |
| Email archives and chat history | Rarely | Full of personal data and outdated decisions |
Our guide to building a chatbot from company documents goes deeper into formats and structure. The short version: one topic per document, a clear date, and headings that match how people ask.
Permissions by role
This is where internal assistants go wrong. If every document is indexed into one shared pool, a new hire can ask about management bonuses and get an answer quoted from a board memo. The fix is to carry permissions into the retrieval step: the assistant searches only what the signed-in user is allowed to open.
- Sign in with the accounts you already have, such as Microsoft 365 or Google Workspace, so the assistant knows who is asking.
- Group sources by audience: everyone, managers, specific teams or sites.
- Map each group to the roles that may see it, and test with a low-privilege account before launch.
- Log questions without storing more personal data than you need, and say so in your staff privacy notice under GDPR.
Keeping the content owned and current
Give every source a named owner and a review date. When HR changes the leave routine, the updated document replaces the old one in the index the same day. A quarterly check of the questions the assistant could not answer shows which routines are missing altogether. Those gaps are often the most useful output of the whole project.
Set the assistant to say 'I don't know, ask HR' rather than guess. Our article on stopping a chatbot from making things up explains the settings that make that behaviour reliable.
Measuring whether it is used
Questions per week
Should grow in the first month, then settle. A flat zero means people do not know it exists or do not trust it.
Unanswered questions
The list of what it could not answer is your documentation backlog.
Thumbs up and down
A simple rating on each answer flags wrong or outdated sources quickly.
Interruptions
Ask the people who used to get the questions. Fewer repeats is the real result.
When you do not need a custom assistant: if you are already on Microsoft 365 with Copilot licences and your documents sit in SharePoint with sensible permissions, Copilot may cover most of this out of the box. A small team of five that sits together will also get more from a tidy shared folder than from a chatbot. A tailored setup earns its place when you have several sites, shift workers without desk access, or sources spread across systems. Our AI chatbot setup covers that case, and you can book a short call to find out which side you are on.
Frequently asked questions
Is it GDPR-compliant to let staff ask an AI questions?
It can be. Use a provider with an EU data processing agreement, keep personnel data out of the sources, log only what you need and describe the processing in your internal privacy notice.
Can the assistant answer in both Swedish and English?
Yes. Modern language models answer in the language of the question even when the source is in the other language, which helps mixed-language teams.
What happens when a routine changes?
The owner updates the document and the index refreshes, automatically or on a schedule. Without an owner the assistant keeps quoting the old routine, which is worse than no assistant.
Can it take actions, like registering leave?
It can be connected to systems through their APIs, but start with answering questions. Actions add permissions, testing and risk, and are best added once people trust the answers.
Wondering if your team would use it?
Bring your list of recurring questions to a 15-minute call. We tell you whether an internal assistant is worth building, or whether a better handbook would do the job.
Book a free 15-minute call