Poradnik
AI agent or automation? How to choose a solution for a process in your company
An AI agent is worth considering when completing a task requires choosing the next action to suit the situation. With a fixed sequence of steps you can start with automation. If a person mainly needs help preparing content or analysing material, an AI assistant may be enough. The choice should follow the task, the data and the way the result is checked.
Vendors use different names for these things. So when you discuss a deployment, ask what the system actually does: whether it performs set steps or chooses the tools and the course of the work itself. Anthropic describes a similar distinction between a predesigned workflow and an agent that directs actions dynamically. Anthropic’s description of the approach.
How to recognise what the company needs
Ask to be walked through one real task. Establish what data is available at the start, what has to be ready at the end and what decisions arise along the way. A concrete example, where the current work gets stuck, usually explains the most.
If a file always has to be checked in the same fields and the result saved in the same place, the scope is predictable. If you have to work out what is missing, look for it in various sources and decide whether the information is sufficient, the need for flexibility grows.
Even then, check first whether a few explicit rules solve the problem. Extra freedom for the system makes sense when it brings a benefit that justifies the greater complexity of assessing and supporting it.
| Situation | Solution to consider | What to check in a test |
|---|---|---|
| Fixed input data and a set order of steps | Automation | Correctness of rules and handling of bad data |
| A person runs the task and judges the prepared content | AI assistant | Usefulness of the result and time spent checking |
| The next action depends on information found along the way | AI agent | Whether the actions chosen are apt and when it stops |
- AutomationThe steps are set in advanceCheck the rules and bad data.
- AI assistantA person runs the taskCheck usefulness and time spent checking.
- AI agentThe next step depends on the situationCheck the choice of actions and stopping.
How to limit the scope of a first AI agent
Choose one result and define the sources the system may use. Set which actions it may take on its own and which need approval. Separating preparing a recommendation from carrying out an action makes it easier to assess each stage on its own.
For instance, a tool can gather changes to terms of cooperation and prepare a summary. Deciding whether to accept new terms is a further step, requiring its own definition of responsibility. Likewise, preparing a quote and sending it to a client have different consequences.
You also need a stopping condition: missing required data, exhausting the agreed limit of work, or a case outside the scope. The system should be able to hand a matter to a person along with what it did and what is missing.
Why it pays to simplify the process first
Organisational habits can hide unnecessary steps. Automating every current step preserves the ones that are not needed to get the result.
Model example, fictional figures: a process has 17 steps. Six only re-enter information, and two checks can be merged into one. After the change 10 activities remain. That does not automatically mean a proportional time saving, but it changes the scope to be designed.
So start from a description of the result and the acceptance criteria. What this looks like in practice is shown in the case study The process had 17 steps. Ten were enough.
How to compare vendors’ proposals
Ask them to show the solution on the same cases and to state what the company gets after the test. Compare the quality of the result, people’s working time, the cost of launch and maintenance, how updates work and whether a named person can take over support.
Clarify what happens after a failed test. Is there a record of what was agreed and why it failed? Is it clear what would need to change? That information can be useful even when the decision is to narrow the scope or stop the project.
Which decision to start with
First choose a task whose effect can be checked and whose owner is ready to cooperate. Only then agree how autonomous the system should be and what technology is needed. One process can combine fixed rules with AI where it helps interpret variable content.
At dyrektor.ai we run this work from recognising the need, through an agreed scope, to assessing the result. For the board that means a basis for a decision, and for the team a clearly defined next step.