triox informatika

August 4, 2026 · Triox Informatika · 5 min read

You switch on AI in the office. The first surprise won't be what it can do.

The AI assistant doesn't break into anything—it finds exactly what the employee already has access to. The problem is that at most companies, nobody knows exactly what that covers.

In brief

  • An office AI assistant isn’t magic: it reads across the company’s own email, files and messages.
  • It only sees what the given employee has permission to access. That sounds reassuring—until we look at what they have permission to access.
  • The first week of a rollout typically isn’t about the AI. It’s about discovering the real state of permissions.
  • That’s no reason to put it off. It’s a reason to do the cleanup before switching it on.

More and more Hungarian companies are facing the same question: is it worth switching on the artificial intelligence built into their office software? The question is usually put like this: “What can it do, and is it worth the price?” Experience shows, however, that this won’t be the interesting part in the first days of a rollout. A very different realization will.

What does an office AI assistant actually do?

It’s worth stripping away the marketing. In practice, an office AI assistant does three things: it summarizes, it finds and it drafts.

It summarizes the meeting that left behind a two-hour recording. It finds, among the company’s own materials, what we promised this business partner last year, and it cites the document it took that from. It drafts a polite reply to an email from the three words you tossed at it.

What’s new here isn’t the writing. It’s that it searches the company’s own data: email, shared folders, group chats. A traditional search box has to be told the file name. This one only needs “what did we promise Kovács Kft. about the delivery deadline.”

The AI doesn’t break into anything. It just searches very well.

This sentence is the heart of the article. The system works on the user’s behalf, so it reaches exactly the materials that person could open themselves. It doesn’t make a single file accessible that was locked before.

That isn’t where the problem is. The problem is that at a typical company, “what they have access to” and “what they have ever opened” are two very different sets—and until now nobody noticed, because nobody searched for it.

An employee asks a question; the assistant searches the email, cloud storage and shared folders the employee has permission to access, then answers with citations. The key point is that the scope of access is wider than what the employee has ever opened. AN EMPLOYEE ASKS A sentence in everyday words WHAT THEY CAN ACCESS email and calendar cloud storage, documents shared folders, group chats ANSWER A summary with sources cited Access rights ≠ what they ever opened until now, the gap was invisible

What the company believes, and what’s actually true

The gaps below aren’t theoretical. They are the ones we run into most often during an assessment.

What the company assumes What an assessment typically finds
“Payroll is in a separate folder, only HR can see it.” The folder inherited permissions from a parent folder, and half the office can see it.
“I only shared that file with someone via a link.” The link was created with the “anyone with the link” setting, and it has been forwarded since.
“We’ve locked the accounts of people who left.” The account is locked, but its group memberships and shares are still active.
“Only a few folders are shared.” Over the years, hundreds of individual shares have piled up that nobody keeps track of.
“Our documents are well organized on the server.” The actual work happens in email attachments and personal cloud storage.

None of this is the AI’s fault. All of it built up over years, in complete good faith: someone wanted to share a file quickly, someone created a new folder, someone left the company. Until now it stayed hidden, because searching was hard. The assistant doesn’t break this state—it reveals it.

AI is not a security hole. It’s a spotlight. It lights up what was already there in the dark.

It’s not the AI you need to switch off

There are two conclusions you can draw from this. One is: then better to have no AI at all. That’s an understandable reaction, but it solves nothing: messy permissions remain just as much of a risk, only nobody sees them. And a disgruntled or departing employee can still find that folder, without any help.

The other, more useful conclusion: the rollout is a good opportunity to finally get permissions in order. This is the job that has been sitting at the bottom of the to-do list for years, because it was never urgent. Now it is.

Four steps before you switch it on

  1. Find out what's where. What folders, groups and shares exist, and who each one is meant for. The goal isn't perfect order; it's that the set of confidential material is known and kept narrow.
  2. Close down broad shares. Start by reviewing “everyone in the company” and “anyone with the link” type permissions—most of the surprises come from these.
  3. Start small. One team, a few weeks, real tasks. This both measures how much it's worth to the company and shows where permissions go wrong—before everyone is using it.
  4. Spell out what can go into it. A short, plain-language policy: what data can be entered, what needs management approval, and what no one double-checks. It's important that employees know: the output always has to be reviewed by a person.
Technical details
  • Permissions review on file shares and cloud storage: mapping inheritance, unique permissions and orphaned access
  • Review of sharing links: handling links that point outside the organization or never expire, tightening the defaults
  • Directory cleanup: reviewing group memberships, departed users, service accounts and external guests
  • Labeling confidential data: introducing sensitivity labels for the material that truly needs protection
  • Logging and reporting: who accessed what, and how access changes over time
  • Licensing and legal framework: what is available in which subscription, where the data is stored, and what goes into the internal policy

What to expect, realistically

Today, office AI takes repetitive, text-based work off your plate: summaries, searching, writing the first draft. What it can’t do: take responsibility. Numbers, deadlines and legal content still have to be checked by a person—the assistant sounds convincing even when it’s wrong.

So the success of the rollout doesn’t depend on which product you choose. It depends on whether the system behind it is in order, and whether employees know where the line is.

Getting permissions in order pays off on its own—even if the AI never ends up being rolled out. If you’d like to see where things stand at your company, let’s start with an assessment: at the end you get a documented picture of the current state and a list of risks. Request an assessment, or see what managed IT operations looks like with us.

← Back to the blog