ExSitu

From an Agent That Answers to an Agent That Acts — What Has Really Changed in Enterprise Tools

The current wave of enterprise AI tools has moved from "answering questions" to "completing tasks." It is a change that requires entirely different controls.

4 min read

From an Agent That Answers to an Agent That Acts — What Has Really Changed in Enterprise Tools
Unsplash · Annie Spratt

The first half of 2026 marked a clear shift in enterprise AI tools: agent frameworks became a standard part of the major platforms. The difference is not in the model's power but in the scope of action.

The Difference in Practice

Until now, a typical request sounded like this: "Draft an email to the customer." Now it sounds like this: "Research the customer, draft the email, schedule it to be sent and log the activity in the CRM." The same task, four stages, zero approvals in between.

This is a real change in capability, and also a real change in risk. These tools work well on narrow, well-defined tasks. They work less well when the task is vague — and then they do not stop to ask, they keep going.

What It Requires of the Organization

An agent that performs actions needs its own identity, its own permissions and its own audit trail. In security terms it is a user in every sense — and usually a user that was given broad permissions "so it can work," without anyone thinking about what that means.

That is exactly why OWASP dedicated a separate risk list to agent-based applications, and why the IBM report found that most breaches targeting AI went through interfaces and permissions, not through the model.

How to Approach It

The simple rule that works: do not give an agent broader permissions than the employee it assists. And if the action is irreversible — a payment, a deletion, sending something externally — there should be a person in the middle.

It is also worth remembering that a decision not to adopt an agent is a legitimate decision. A task that takes ten minutes once a month does not justify agent infrastructure and the risk that comes with it.

What This Means for Your Organization

  • Before adopting an agent — define the task precisely. An agent on a vague task is a recipe for a vague result.
  • Give each agent a separate identity, not a shared account. Without that, there is no way to know who did what.
  • Define a list of actions that require human approval, and write it before deployment, not after.
  • Measure. If after a month you cannot point to savings in time or errors — it is a tool that has not justified itself, and that is a valid answer.

Sources