Glossary
AI Agent
An AI agent is a program that takes on a goal, works out the necessary intermediate steps on its own, and calls on tools such as search, databases, or other programs along the way. The difference from a chatbot isn't language ability: it's that an agent acts instead of just answering.
What an AI Agent Actually Does
A language model on its own generates text. An agent gets two additional things: tools it can call, and a loop in which it checks after every step whether the goal has been reached. That turns “write me an answer” into “find out what applies here, and take care of it.”
An everyday example: an invoice arrives by email. A fixed process would file it according to a rigid rule. An agent reads it, looks up the matching order in the system, notices a discrepancy in the amount, flags the case for review, and notes what the discrepancy consists of. None of these intermediate steps were written down anywhere.
Where It Pays Off and Where It Doesn’t
Agents pay off when a process has many exceptions. That’s exactly where classic automation fails: every exception needs its own rule, and at some point nobody maintains the rulebook anymore.
They do not pay off for processes that are always the same. Anyone moving the same data from A to B every day gets a better result with fixed automation: faster, cheaper, and traceable when something goes wrong. An agent doing a job that a simple rule could have done is expensive luck.
The second limit is traceability. Where a decision has to be justified, for instance to an audit, a self-chosen path to the solution is a drawback. In that case you need either fixed processes or a log of every step.
What Deployment Requires
Three things before the first agent goes into production:
- Reachable systems. An agent is only as useful as the interfaces it’s allowed to use. Where everything runs through user interfaces only, it gets slow going.
- A clear boundary. What is it allowed to decide on its own, and where does it need to check first? That boundary needs to be set before the first deployment, not after.
- A measurable case. Without a baseline, there’s no way to tell afterward whether it made a difference.
Not to be confused with
- Chatbot
- Answers a question and is done. An agent pursues a goal across multiple steps and decides along the way what needs to happen next.
- Process Automation
- Runs a predefined process the same way every time. An agent works out the process itself within the scope of its task. For stable, uniform processes, classic automation is faster, cheaper, and easier to trace.
- RPA
- Operates existing interfaces the way a person would: clicking through screens. Agents work through APIs instead of the user interface. RPA use has dropped noticeably over the past twelve months.