“Agent” is the word of the year, which means it is now attached to things that are not agents. That would be a harmless marketing problem except for one thing: building the wrong one wastes weeks. So here is the distinction in plain language, and a test you can apply to your own problem in about a minute.

An automation follows a path you drew

An automation is a fixed sequence. Something happens, then steps run in order. A form gets submitted, so a row is added to a spreadsheet, an AI step drafts a summary, and the summary lands in your inbox. Every single time, the same path.

That rigidity is a feature, not a limitation. Automations are cheap, fast, and boringly reliable. When one breaks, you can point at the step that broke. If you can sketch your process as a flowchart on the back of an envelope, you want an automation — and you can probably build it in an afternoon.

An agent decides its own path

An agent is given a goal, a set of tools, and permission to figure out the steps. You do not tell it “first search, then summarize, then email.” You tell it what a good outcome looks like and let it choose what to do, in what order, adapting as it goes.

That flexibility is genuinely powerful for work where the steps cannot be known in advance. “Research these fifteen suppliers and flag anything concerning” is agent work — what counts as concerning varies per supplier, and the path through the research depends on what turns up.

It is also slower, more expensive per run, and less predictable. An agent can take a reasonable-looking wrong turn and burn twenty steps on it. This is the part the demos leave out.

The one-minute test

Ask yourself: can I draw the flowchart?

  • Yes, easily. Build an automation. You are done deciding.
  • Yes, but it has a few branches. Still an automation. Branches are just conditions, and conditions are cheap.
  • No, because it depends on what we find. Now you are in agent territory.
  • No, because I have never actually mapped this process. This is the most common answer, and it means the real task is mapping the process — not building anything yet.

That last one deserves emphasis. A surprising amount of “we need AI for this” turns out to be “we have never written down how we do this.” Writing it down often reveals that three of the seven steps exist only because someone left the company in 2019.

Whichever you build, keep a human on the consequences

The single most important design decision is not automation versus agent. It is which actions require a person to say yes.

My rule: anything that sends, pays, publishes, or deletes gets an approval step. The system can do all the work — draft the email, prepare the payment, write the post, identify the files — and then stop and wait for a human. You keep nearly all the time savings and you eliminate nearly all the catastrophic outcomes.

Automating a task you have never verified is just making the same mistake faster and more often.

Start with something small, repetitive, and low-stakes. Get it running. Watch it for a week. The confidence you build from one boring automation that actually works is worth more than an ambitious agent you do not trust.