What an AI Agent Actually Does, in Plain Terms

Three words that get used interchangeably and should not be

A chatbot answers questions in a conversation. An automation runs a fixed sequence of steps. An agent is given a goal and some tools, and works out the sequence itself, step by step, reacting to what it finds.

The distinction matters commercially because they have wildly different costs and risk profiles. Automations are predictable and cheap to run. Agents are flexible and cost more per task, both in tokens and in the effort required to keep them behaving.

The test for whether you need one

Ask whether the steps are the same every time. If yes, do not buy an agent. A fixed pipeline that does the same six things in the same order is cheaper, faster, easier to test and impossible to talk into doing something strange.

If you can draw the process as a flowchart and it fits on one page without loops, you want an automation. If drawing it requires the phrase “it depends what we find”, an agent starts to earn its cost.

Where agents genuinely earn their keep today

1. Triage and routing. Read an enquiry, work out what it is, gather what is needed, route it with a summary. The variety is the point.
2. Research and consolidation. Pull from several sources, reconcile, produce a briefing with citations. Slow for a person, well suited to a tool-using agent.
3. Multi-step resolution with checks. A support case that needs a lookup, a policy check and an action. Each step is simple; choosing which comes next is not.

What they still get wrong

Agents fail in characteristic ways. They loop, repeating a step that did not work. They stop early and declare success. They take a plausible but wrong path when the data is ambiguous. None of these are exotic bugs — they are the normal behaviour of a system doing probabilistic planning.

Which is why every production agent needs boundaries: a step limit, a cost ceiling, a list of tools it may call, and a rule about which actions require human confirmation. Anything that spends money, sends external communication or deletes something belongs behind that confirmation.

The cost model people get wrong

Agents charge per step, not per task, and a task can take fifteen steps. Budget by expected steps rather than by request, and instrument the distribution — the average is fine, the tail is what surprises you on the invoice.

How to start without betting the quarter

Choose one workflow with genuine variability, that a person currently does in ten to thirty minutes, and where a wrong answer is embarrassing rather than catastrophic. Run the agent in draft mode for a fortnight — it produces the output, a person approves it, and every disagreement is logged.

That log is the whole project. It tells you the real accuracy, it shows which failures are systematic, and it is the evidence you need to decide whether to let the agent act without approval.

Frequently asked questions