Most companies have that one person you ask when you cannot find something. They know which folder holds the work instruction, which version of the contract template is current, and why that exception for customer X was agreed back then. As long as that person is around, it works. During their vacation things stall, and when they leave, knowledge is lost that was never written down – or was written down but cannot be found. An internal AI assistant solves the second problem above all, and it is bigger than many think.

The real problem: knowledge exists but cannot be found

Very few companies have too little documentation. They have documentation in seven places: a SharePoint, a wiki nobody has maintained for two years, network drives with folder structures from three reorganizations, email threads containing the actual decision, and PDFs named “final_v3_NEW”. The information is there. But whoever needs it has to know where it lives and what it is called – and that is precisely what only the one person knows.

This costs measurable time. As far back as 2012, the McKinsey Global Institute calculated that office workers spend nearly 20 percent of their workweek looking for internal information or tracking down colleagues who can help. That is close to a full working day per week. The number is old, and so is the problem – what is new is that it has become solvable.

Classic full-text search helps only so much here, because it finds terms, not answers. Search for “vacation carryover” and you will not find the document that talks about “remaining leave into the following year”.

What an internal assistant does differently

An internal AI assistant connects a language model to your existing repositories via RAG: SharePoint, wiki, document folders, and optionally the ticket system and manuals. Employees ask in their own words – “How do I set up a new supplier?” – and get the answer from the right document, with a link to the passage, in the current version.

The difference from search: the assistant understands the intent behind the question, finds content phrased differently, and summarizes instead of listing twenty hits. The difference from the colleague: it is reachable at 6 a.m. and in August, and it does not get tired of answering the same question for the fortieth time.

Where the benefit shows up first

From the projects we build, a few typical spots stand out where an internal assistant makes a difference quickly:

Onboarding: new employees resolve basic questions themselves instead of interrupting experienced colleagues for every small thing – ramp-up time shrinks noticeably
Rare procedures: processes that occur only a few times a year – year-end closing steps, exception workflows, rarely used machines – no longer have to be painstakingly reconstructed
Policies and rules: what currently applies to expenses, remote work, approval limits? The answer comes from the valid document instead of from memory
Knowledge of departing employees: whatever is documented stays usable – even when nobody remembers that the document exists

Permissions are not a detail – they are half the battle

The most delicate part of internal assistants is not the AI but access. Not everyone in the company may see salary bands, HR files, or the margins of individual quotes – and the assistant must not bypass the permission model by dutifully quoting from documents the person asking could never open. A properly built assistant therefore checks the user’s rights on every request and searches only what that person could read directly.

That sounds obvious, but it is a common weak spot in quickly assembled solutions. We settle the permission model before the first line of code – along with where the assistant runs and what happens to employees’ queries. Those are personal data too.

The honest part: outdated documents stay outdated

An assistant makes your content findable, not correct. If the work instruction dates from 2021 and the process changed in 2024, the assistant delivers the outdated instruction – politely, with a source reference, and completely wrong. Connect a chaotic repository and you get an assistant that delivers chaos faster.

In practice this is less dire than it sounds, and it even has a side effect customers appreciate: because every answer shows its source, outdated and contradictory documents suddenly become visible – they are being read again. The assistant becomes the trigger to clean up the repository exactly where it is needed, instead of in some theoretical grand project.

What getting started looks like

We advise against connecting the whole company on day one. Better: one delimited knowledge area with clear demand – say, one department’s process documentation or the manual knowledge around one product. That lets you test with the team’s real questions whether the answers hold up, before further sources are added. A pilot like this stays within the scope of a small software project and delivers solid results within a few weeks.

And to answer the question that is guaranteed to come from the works council: an internal assistant is a tool, not surveillance. Who asked what does not belong in any report – but which questions come up empty does, because that is exactly where documentation is missing.