triox informatika

August 6, 2026 · Triox Informatika · 5 min read

Until now, AI made suggestions. Now it acts.

An AI agent doesn't write an answer—it does the task: it signs in to the system, fills things in, sends, makes changes. In plain terms: what this changes in a company's day-to-day work, and where it's worth stopping it.

In brief

  • AI until now answered. An agent acts: it signs in to systems, fills out a form, sends an email, changes data.
  • That hasn’t made it smarter. What changed everything is that its mistakes now have consequences.
  • The question isn’t whether we trust it, but what it has permission to do and what it can’t do without approval.
  • The rules are the same as for a new employee: narrow permissions, approval before irreversible steps, and a log of who did what.

Over the past few years, office artificial intelligence did one thing: it returned text. You asked for a summary, you got one. You asked for a draft email, you got one. What you did with the answer was up to you—you read it, corrected it, sent it.

Now a new word is making the rounds in the industry: agent. The name may sound mysterious, but what it means is simple. An agent doesn’t make a suggestion; it does the task. It doesn’t write the email—it sends it. It doesn’t tell you what to enter into the system—it enters it.

What’s the difference between “answers” and “does it”?

At first it sounds like a matter of convenience: one less click. In practice, though, an important safety element disappears from the process—the person who used to stand between the answer and the consequence.

When AI only made suggestions, every one of its mistakes stopped at a human glance. Was the amount wrong? You caught it before sending. Was the deadline off? You fixed it. This filter wasn’t perfect, but it was there.

With an agent, this step drops out—unless we deliberately put it back.

Before, the process had three steps: a question, the AI's suggestion, then a person carried out the action—stopping came naturally. With an agent, the task is followed by machine steps and the action happens on its own, so the approval point has to be deliberately put back before the irreversible step. THEN Question that you asked AI suggests returns text A person does it a natural stopping point NOW The task that you gave Agent acts login, fill, send The action is done nothing stops it on its own APPROVAL GOES HERE before the irreversible step

Why now?

It didn’t happen overnight. Two things came together.

One is that language models can now break a task down into steps reliably enough: what to look at first, what comes next, when it’s done. The other is that, one after another, software products have gained connection points through which a program can do the same things a person does by clicking in the interface.

That is what suddenly made agents usable—and that is what made the question of permissions urgent.

The turning point isn’t that AI got smarter. It’s that it no longer suggests—it presses the button.

The cost of a mistake has changed

A language model is sometimes confidently wrong. That was never good, but it was manageable: a person read the text and corrected it. With an agent, however, the mistake doesn’t stay text—it becomes a sent email, changed data or an order.

It’s important to see that this isn’t about AI “running amok.” The everyday risks are far more mundane: it misunderstands the task, picks the wrong record, does the same thing twice, or goes ahead with something you only asked it to prepare.

Not a question of trust, but of permissions

The most common mistake is that companies frame this as a question of trust: “Can we trust it?” That’s a dead end, because there’s no good answer. The useful question goes like this: what can it do, and what does it need permission for?

This is exactly the same logic that works with a new employee. They don’t get narrower permissions because we don’t trust them, but because that way their mistakes do less damage, too.

What you can safely hand over What needs human approval
Collecting, extracting and comparing data Sending an email to a client or partner
Preparing drafts, summaries, meeting minutes Financial transactions, orders, contractual commitments
Organizing internal tasks, reminders Changing or deleting data in a production system
Preparing recurring administrative work Granting permissions, creating accounts
Suggesting what the next step should be Anything visible outside the company

The point of the right-hand column isn’t that the agent shouldn’t touch these things. It’s that it can prepare them, but a person sends them.

Five rules before you switch it on

  1. Give it its own identity. Don't let it work under an employee's account, because then everything in the log appears under that person's name. With a separate identity, you can trace what the machine did and what the person did.
  2. Start with read-only access. For the first few weeks, it's enough for it to look, gather and suggest. What truly saves time often shows up at this level already.
  3. Draw the irreversible line. A sent email, a transfer, a deletion, a permission grant: these should always be preceded by human approval. This doesn't slow down the benefit; it only limits the damage.
  4. Keep a log, and have someone actually look at it. In the early period, it's worth skimming through what it did every week. Most misunderstandings come to light here—well before they turn into trouble.
  5. Say who it belongs to. Name an employee who is responsible for it: they shut it down if it behaves strangely, and they're the one people ask if something seems off. Systems without an owner cause the most trouble.

And what’s worth doing on the systems side

Technical details
  • A separate service account for the agent, with the narrowest permissions needed—not a personal account, not administrator rights
  • Permission expiration and regular review: what was needed for an experiment shouldn’t still be active three months later
  • An approval step before outbound and irreversible actions, technically enforced, not just requested in a policy
  • Full logging: what it accessed, what it changed, when—in a searchable form
  • A separate test environment for the first phase of the rollout, with no production data
  • Clarifying data handling: what data can go to the service, where it is stored, for how long—this also goes into the internal policy
  • A way to stop it: there should be a switch that stops it immediately, and someone should know where it is

What to expect, realistically

Today, agents deliver tangible results in repetitive, well-defined administration: they gather data, prepare and organize. What doesn’t work: taking over responsibility. Numbers, deadlines and commitments still have to be approved by a person.

So the success of the rollout doesn’t depend on which product you choose. It depends on whether your permissions are in order, and whether you have decided in advance where the line is.

An agent is exactly as dangerous as the permissions you gave it. That’s not about the agent—it’s about you.

Getting permissions in order is the first step—even if agents are still some way off. 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 read about our approach to security.

← Back to the blog