All posts

When should an AI agent ask a human?

A simple way to decide which steps in an automated workflow need a person, and what that person should see when they do.

By Human In Circuit

AI agents and automations are good at the boring middle of a task: drafting the reply, matching the invoice, classifying the expense. The risk usually sits at the edges, where the work leaves your system and becomes real. A refund is paid. A post goes live. A customer reads an email.

Most teams handle this in one of two ways. They either let the workflow run unattended and hope, or they keep a person watching everything, which defeats the point of automating. There is a better middle: pause only the steps that need judgment, and make that pause quick to clear.

Three questions for every step

Before you add a human checkpoint, ask these about the step:

  1. Can it be undone cheaply? Saving a draft can. Sending money, deleting records or publishing usually cannot.
  2. Who notices if it is wrong? If the answer is “a customer” or “the finance team at month end”, a person should look first.
  3. Is the agent guessing? Low confidence, missing data or an unusual amount are good reasons to stop and ask.

If a step is irreversible, visible to others, or uncertain, it is a good candidate for review. Everything else can keep running on its own.

Make the review fast, or it will be skipped

A checkpoint only works if people actually clear it. That means the reviewer should see:

  • What the agent wants to do, in one line: “Send refund reply”, not a log dump.
  • The evidence, short and readable: the matched purchase order, the customer’s message, the draft itself.
  • The values that matter, editable: the refund amount, the category, the publish date.
  • A clear way to say no, with a reason the agent can use next time.

The answer should go back to the workflow as structured data, not a chat message someone has to interpret. “Approved, amount changed from $50 to $42” is something the next step can act on.

Start with one step

You do not need to redesign your automation. Pick the one step that worries you most, usually the one that moves money or talks to customers, and put a checkpoint in front of it. Watch how often the agent is right. Tighten or relax from there.

How Human In Circuit fits

Human In Circuit is built for exactly this pause. Your workflow, whether it runs in n8n, Zapier or your own code through the API, sends a decision request. It appears in a calm inbox on your phone or browser. You approve it, edit the details, or reject it with a reason. The workflow then continues with your decision, either through a signed callback or by checking for the result.

Each request your workflow sends uses one credit, and the Free plan includes 100 credits every month. If you want to try it on that one worrying step, start free or read the quickstart.