Why every workflow I ship pauses for a human

Here is the rule: nothing automated sends, pays, deletes or publishes without a person approving it first.

That sounds obvious until you've built something that works. The first time a workflow drafts a perfect customer reply, the temptation is to let it send. It'll be right almost every time. The few it gets wrong are the ones you'll remember.

Where this came from

I run a small eBay hardware business. Early on I wired up an agent to answer buyer questions overnight, because most of them are the same question — "does this include the power supply?" — and I wanted to sleep. It answered correctly for weeks. Then a buyer asked a two-part question, the agent answered the first part with confidence and ignored the second, and I woke up to a return request.

The agent didn't fail. The design did. There was no step where a person looked at the draft before it left.

What the pause actually looks like

In n8n it's one node. The workflow drafts, then waits. You get a message with the draft and two buttons: approve or edit. Nothing moves until you press one.

Trigger → Claude drafts reply → Wait for approval → Send
                                   │
                                   └─ Edit → back to Claude with your note

Three things make it work in practice:

1. The approval message shows the exact text that will be sent. Not a summary. If you can't see it, you can't approve it.
2. "Edit" is a real path, not a dead end. Your note goes back to the model and you get a new draft, so you never have to leave the chat to fix a word.
3. There's a timeout. If nobody approves in 24 hours, the workflow stops and tells you — it never sends by default.

The cost of the pause

About ten seconds per message, from your phone. For the eBay business that's maybe two minutes a day. Against that: I haven't had a wrong message leave since, and I read every reply the agent writes, which means I notice when it starts drifting.

When you can skip it

When the action is reversible and cheap. Logging to a sheet, tagging a record, sending something to yourself — automate those fully. The test is simple: if the agent gets this wrong at 3 a.m., what does it cost me at 9 a.m.? If the answer is "a minute", skip the pause. If the answer involves a customer, money or a delete key, keep it.

Try it

The steps above are enough to add the pause to anything you already run. If you'd rather watch n8n workflows being built, my free lessons are on the lessons page. If you want the finished workflow files, the approval message templates and the rules I use for my own agent, that's what Build a Safe AI Operator is.