What an AI agent does in a company that is not tech
An AI agent is software that does a job, not a chat window that answers questions. In an ordinary company it works inside the tools you already have. It picks up repeatable work, follows the rules you set, and hands back anything it cannot judge.
What is the difference between an AI agent and a chatbot?
A chatbot waits for a question and produces text. An agent is given a job and produces a result.
The difference shows up in what happens after the answer. A chatbot tells you which invoices are overdue. An agent opens the accounting system, finds them, drafts the reminders and puts them in front of you to send.
That is why we call it a digital employee rather than an agent. The word describes the thing better. You do not brief a chatbot on Monday and check its work on Friday.
Which work does it take over first?
The work that repeats, follows rules, and lives in systems it can reach. Reading and sorting an inbox. Chasing a payment. Copying an order into a second system. Preparing a weekly report.
There is a simple test. Can you write down how the work is done? If the description never lands on "and then you feel it out", it is a candidate. If the instructions end in judgement, it is not.
Most companies find more of this work than they expected. It rarely sits in one job. It sits in the twenty minutes at the start of everyone's day.
What stays human?
Anything where being wrong is expensive and being right needs context that is not written down anywhere.
Pricing exceptions. An unhappy customer who has been unhappy before. A supplier relationship. The decision to make an exception at all. A digital employee can prepare all of these and should not close any of them.
The useful framing is not "which jobs disappear". It is "which part of this job is typing, and which part is deciding". The typing goes first.
Where does it go wrong?
Three ways, in our experience, and only one of them is about the model.
It is pointed at work that was never really a process, so every case turns out to be an exception. It is given access to a system nobody has cleaned up, so it acts on wrong data confidently. Or nobody agreed who checks its output, so after two weeks nobody does.
The model itself producing nonsense is the failure people expect and the one we see least. It is also the easiest to catch, because you can read what it wrote before it sends.
What do you need before you start?
One workflow you can describe out loud. Access to the systems it touches. And a named person who reads the output in the first weeks.
You do not need a data strategy, a new platform, or your systems cleaned up first. Those are reasons to postpone, and postponing is the expensive option.
What you do need is honesty about the first item. If you cannot describe the workflow, that is not an AI problem yet.
Written from AImplement implementations at Dutch companies of 10 to 200 people.