Before you build an agent
Does this step need an AI agent?
Check before you plan one. Name one step in a client's workflow and answer up to five questions. You will see which of four labels fits the step (Automate, AI-assisted, Agent, or Keep human) and why.
Your step:
The result
This checks one step, from your answers alone. A full workflow needs the whole picture: who does each step, the systems, what goes wrong, and which step is worth building first. That is what AgentStepper does.
Your answers stay in your browser. Nothing is sent or stored. Back to the home page
The check needs JavaScript to run. These are its questions, in order:
- Should this step stay with a person? Yes, if it depends on judgment only a person can make, on a relationship, or on someone who must answer for the outcome. If yes: Keep human.
- Does finishing this step take several actions, not just one?
- Does what happens next change from case to case, in ways too varied for fixed rules?
- Does the work cross two or more systems? If questions 2, 3, and 4 are all yes: Agent.
- Does the step read or write free-form text, such as emails, notes, or replies? If yes: AI-assisted. Otherwise: Automate.
The four labels
Every step in a workflow fits one of four labels. Only one of them is Agent.
| Label | What it means |
|---|---|
| Automate | Rule-based and repeatable. No judgment needed. |
| AI-assisted | AI drafts or suggests. A person decides. |
| Agent | Pursues a goal over several actions across tools, with a person approving key steps. |
| Keep human | Judgment, relationships, or risk that should stay with a person. |
The check asks its questions in the same order AgentStepper uses for every step: first whether a person must stay in charge, then whether an agent fits, and only then whether AI can help a person or plain rules are enough. A "Not sure" answer is treated as No for now, so the result is never more ambitious than your answers support. It shows what the label could change to, and it becomes a question to ask the client.