Agents vs. automations: when do you actually need an AI agent?

Every week a client asks us for "an AI agent". About half the time, what they really need is a well-built automation with one or two AI steps. Here is the framework we use to decide.

1. Is the path known in advance?

If you can draw the process as a flowchart, start with a workflow. Deterministic steps are cheaper, faster and easier to test. Add an LLM only where you need to read, classify or write.

2. Does the task need tools in an unknown order?

Agents shine when the next step depends on what was just learned: researching a lead, debugging a failed order, answering a question that spans three systems.

Start with the workflow. Promote it to an agent only when the flowchart becomes a hairball.

3. What does a mistake cost?

• Low: let the agent act and log everything.
• Medium: agent drafts, human approves in Slack.
• High: agent recommends, human decides.

A quick checklist

Use evaluation sets of 50–200 real examples before launch, track cost per task, and review a random 2% sample weekly.