AI & Data
AI agents in operations: where they deliver and where they do not yet
AI agents already carry out tasks in finance, procurement, customer service and IT. Where they work, where they still fail and how to start safely.
In 30 seconds
- 01An AI agent does not just answer: it queries systems, decides the next step and carries out tasks.
- 02It works well on tasks with a clear goal, known rules and results that are easy to check.
- 03Start with the agent preparing the work and a person approving it. Full autonomy comes later, if at all.
An AI assistant answers questions. An agent goes further: it receives a goal, queries the systems it needs, decides the next step and carries out tasks until the work is done or it needs help. That difference is what takes AI out of the chat window and into operations.
What an agent needs to work
To act, an agent needs access to the company’s systems, with well-defined permissions. That is exactly the role of standards like MCP, which we covered in MCP in practice. Without controlled access, the agent can only suggest, and someone keeps copying and pasting.
Where they already deliver
The best first agent is boring: it does one repetitive task, with clear rules, that eats up someone’s hours today.
Where they do not yet
Agents still make mistakes on open-ended tasks, tasks with many exceptions or tasks that depend on context not written down in any system. They are also a poor fit where a mistake is costly and hard to undo, such as unchecked payments or contract changes.
The risk is not only the AI’s mistake. An agent with broad permissions executes that mistake at speed and at scale. Least privilege, logging every action and volume limits are part of the project from day one.
In practice · the first-agent test
- Does the task have a clear goal and known rules?
- Can a person check the result quickly?
- Can a mistake be undone without significant loss?
Three “yes”: a good place to start.
How to start
Start with the agent preparing the work and a person approving it before it runs. Measure time saved and accuracy for a few weeks. With data in hand, increase autonomy where the agent is consistently right and keep human approval everywhere else.