Kennisbank

Agents that take steps themselves instead of giving answers

A chat window gives an answer. An agent does something with that answer: it looks up a piece of data, fills in a field, sends a message, waits for a response, and continues. That difference — from answering to acting — is where most organizations stumble over AI. Not because it isn't technically possible, but because acting places different demands than answering.

What happens under the chat window

An agent that carries out work needs access to something. It reads a system, writes to another system, and the steps in between must be recorded somewhere — as a process, as a rule, as a workflow. That is a different layer than a conversation. It concerns chains in which steps follow one another, where the outcome of one step is the input for the next, without a human placing themselves in between each time.

That requires access: rights in systems, a place in the infrastructure, a status that is not human but also not nothing. An agent that is allowed to read and summarize documents works differently than an agent that is allowed to prepare a decision or record and follow up on a conversation. Each of those forms has its own relationship with control, with what goes wrong when it goes wrong, and with who is then involved.

The forms this takes

Agents that carry out work appear in a limited number of recognizable forms. They signal: they keep an eye on a process or a system and report when something deviates, which relates to monitoring and signaling and the question of when a notification should reach a human. They support a decision by ordering options and gathering arguments, without making the decision themselves — the domain of decision support. They switch between departments that need each other but do not use the same systems or the same language, which requires coordination between departments without anything falling through the cracks. And they record conversations and set out follow-up steps, where a commitment made by phone comes back as an action in a system.

None of these forms stands on its own. An agent that signals needs something to refer to. An agent that coordinates needs a system on both sides that allows it in. The form determines not only what happens, but also what already needs to be in order for it beforehand.

What this requires of the organization before it is possible

An agent that carries out work depends on what is already in place. Without clear agreements about who decides on what, an agent does not know which step it may take itself and which step must go back to a human. Without infrastructure that is stable enough, a chain grinds to a halt the moment a system changes or fails to respond. Without data management that is in order, an agent carries out steps on data that is not current or not reliable — with an outcome that is just as unreliable, but with the appearance of precision.

Those three — organization, infrastructure, data management — come first. Not because they are more important in a general sense, but because an agent that needs to build on something requires that foundation first. An agent that coordinates between departments gets stuck if those departments themselves do not already clearly record who is responsible for what. An agent that summarizes documents is useless if nobody knows which version of a document is the valid one.

Beyond that, there is a question that is not technical: what happens if the agent does something that is not right? Who notices that, how quickly, and what is the step back? An organization that has not thought that through only notices once it has already happened.

Where this runs aground

Most pilots with agents do not run aground on the technology itself. They run aground on a system that is not accessible without a manual intermediate step, on a process that exists on paper but in practice has three exceptions, or on a department that does not know that an agent is now also doing part of its work. Those are not small obstacles — they are precisely the things that become visible as soon as something has to actually work instead of merely give an answer.

That is also why a pilot that seems to be going well in a test environment often gets stuck as soon as it moves into practice. Practice has exceptions, old systems, and people who work differently than the process describes. An agent that is not prepared for that does its work halfway or does it wrong, without anyone noticing immediately.

From readiness to task

This page describes what is possible once an organization is ready to let agents carry out work, and what needs to be in place for that first. That readiness is one question. The other question is which tasks within your organization concretely qualify, and which part of that an agent can take on. That is what the werkscan van FTE TO AI calculates per task: not whether your organization in a general sense is ready, but how much of a specific piece of work can be taken over, and what remains of it for a human.