Human in the loop: where the person goes

A human in the loop belongs at the one step in an automation where a wrong move costs real money, damages a relationship, or can't be undone. Everything upstream and downstream of that step can run on its own. The mistake most teams make is putting the person everywhere, or nowhere, instead of at that one point.

What human in the loop means

It's a single checkpoint inside an automated workflow where a person looks at what the system produced before it goes out into the world: an email before it sends, a refund before it processes, a public post before it publishes. The automation still does the drafting, the formatting, the sorting. The person's job shrinks to one decision: approve, edit, or stop.

Why full automation backfires on these steps

Let the machine send the email, issue the refund, or post the update without a check, and it will eventually do all three with confidence and no judgment behind it. A drafting mistake is cheap: reword it and move on. A sent mistake is not. The customer already read it, the money already left the account, the post is already public. Full automation on a low-stakes step saves you nothing you'd notice. Full automation on a high-stakes step is where a small error turns into an expensive one, fast.

Where to put the human in the loop

Ask one question at each step: what does it cost if this is wrong and nobody catches it before it goes out? If the answer is money leaves an account, a customer receives something you didn't check, or the action can't be reversed, that step gets a person in front of it. If the answer is a document gets sorted wrong or a draft sits in a queue for ten minutes, let the machine run it alone. Most workflows only have one or two steps that actually meet the first bar.

Tell me which step in your workflow you'd never want to run unattended, and I'll show you where the checkpoint goes and what can run without one.