An AI agent is the wrong first step in five situations: when the task seldom recurs, when it involves decisions about individuals, when nobody knows how the work goes, when nobody owns the question and when sensitive data is involved without an agreement. In those situations it is better to start with something else.
The task seldom recurs
An agent pays off on work that repeats. A task done once a year takes more effort to describe than to do. Start with something that recurs every week and that today requires someone to move information between systems by hand.
Unusual tasks can come later, once you have seen how the agent works on the ordinary ones.
The decision concerns an individual
Automated decisions about individuals and systems that may count as high risk require an assessment by a lawyer or your data protection officer. Recruitment and credit scoring are examples that the EU AI Act points out.
Compiling a status, drafting a text or sorting email normally belongs to the lower risk classes, but the assessment depends on how the system is used. Raise the question in discovery before anything is built.
Nobody knows how the work goes
If it is unclear who does what and in which order, an agent will make the lack of clarity faster. Map first how the work flows. Often the mapping is enough to find a couple of improvements that need no agent at all.
When the flow is clear, it is possible to build on top of the systems you already have.
Nobody owns the question
Without a named owner on your side there is nobody who decides on the next step. That is a common reason AI projects stall. Appoint the owner first, and choose one defined team and three measures you agree on.
The owner needs the mandate to decide, and needs no technical background.
Sensitive data without an agreement
If personal data is to be handled, a data processing agreement should be written before the work starts, and the AI services that are used should be known and approved by you. Start gladly with information that is not sensitive.
If the agreement is missing, wait with that data.
What you can do in the meantime
There is plenty to do while you wait. Write down three situations where time disappears, list the systems where information is moved by hand and appoint an owner. Then the first conversation is concrete, and when it is time the pilot goes faster.
How you know it is time
It is time when four things hold. The task recurs every week. There is a named owner on your side. You know how the work flows, so that someone can describe it step by step. And a data processing agreement is in place if personal data is involved.
When the four hold, it is fine to start small, with a single workflow and a stop point after the first step. If you answer no to any of them you have a clear list of what remains to be done first.
In short
Wait with an agent when the task seldom recurs, when the decision concerns an individual, when nobody can describe the work, when nobody owns the question and when sensitive data has no agreement. Use the time to write down where time disappears, list the systems and appoint an owner. Then the first conversation is concrete, and getting started goes faster when it is time.