Do You Need an AI Agent, or Just a Workflow?
A practical test for deciding which tasks deserve an autonomous agent and which are better as fixed workflows.
Every week brings more talk about AI agents — systems that pick their own tools, chain multiple steps, and retry when something fails. Yet in teams of three to ten people, the automations that actually save time tend to be far simpler: a human fixes the sequence in advance, and the model just fills in the blanks. Here's how to decide whether your work really calls for an agent.
What actually separates a workflow from an agent?
The difference is who decides the order of steps. In a workflow, you fix the steps and the model handles each task inside them — classify, summarize, draft. In an agent, you give a goal and the model decides which tools to call and how many times. Same model, very different cost, stability, and debugging story.
- Workflow: email arrives → classify type → route to owner → draft reply → human approves and sends. Same path every time.
- Agent: "handle this refund request" → look up the order, check policy, judge the exception, maybe check inventory. The path changes each run.
- Cost: a workflow has a fixed number of calls per item; an agent might finish in three steps or twelve.
- Debugging: with a workflow you can say "step 3 was wrong." With an agent you have to reconstruct the reasoning path.
Which one fits your task?
If inputs are repetitive and mistakes are expensive, choose a workflow. If inputs vary and the work requires open-ended searching, an agent earns its keep. Most small-business tasks sit firmly in the first group. Run through these five checks.
- Is the input structured? Order numbers, support emails, receipts — repeating shapes mean a workflow is enough.
- Is the step count fixed? If you can sketch it with three or four branches, you don't need autonomy.
- Does a human need to look? Sending, refunding, signing — irreversible actions belong behind an approval step.
- Does it require hunting for new sources? "Summarize this week's competitor pricing changes" varies every time; that suits an agent.
- Is failure cheap? If a bad result is obvious within 30 seconds, an agent is worth trying.
For an online store, answering shipping questions is workflow territory — the question types converge on about ten patterns. Gathering research before writing a new product page is agent territory: which reviews and which forums to check changes each time, and a human can filter weak results instantly.
If you do add an agent, what comes first?
Set permissions and limits before you add capability. Most agent incidents happen not because the model is dumb, but because it was allowed to do too much. Follow this order.
- Start read-only: allow lookups and search; keep writes (send, charge, delete) behind human approval.
- Cap the loop: limit tool calls and wall-clock time per request to prevent runaway loops and bills.
- Log everything: record which tool was called with which input. Without it, root-cause analysis is impossible.
- Add a kill switch: make it one click, and name the person responsible for pulling it.
- Watch for two weeks: keep a human in the loop, then auto-approve only the categories with a stable approval rate.
You don't need to agentify everything to keep up. Write down five repetitive tasks and sketch their steps by hand. If the line runs straight, build a workflow; if arrows bounce around, then consider an agent. For a small team, predictable automation usually beats clever automation.