Skip to content

Articles

Approve, review or watch: human-in-the-loop at work

Updated 5 min read By Operators team

Human-in-the-loop means a person stays part of an automated process at the points where judgement matters. In day-to-day operations it takes three forms: a person approves before an action happens, reviews the work after it happens, or watches the results and can step in. Good operators choose the right form for each kind of work, and keep the human part real rather than a habit of clicking yes.

The key points

  • There are three practical forms: approve before, review after, and watch.
  • Put approval before anything hard to reverse. Use review after for reversible work.
  • Requests to the person should be few, clear and backed by evidence.
  • A loop where the person always says yes is a loop in name only.
  • Every automated change should be recorded and reversible where possible.

The three forms

Form When the person acts Good for Example
Approve before Before the action, which waits Hard-to-reverse actions An agent drafts a supplier payment, and you approve it before it is sent
Review after After the action, which can be undone Reversible, routine work An agent sorts the week’s tasks, and you check the result on Friday
Watch Any time, by exception High-volume, low-risk work An agent keeps a list up to date, and you look when something seems off

People sometimes call the first form “in the loop” and the third “on the loop”. The labels matter less than the choice: for each task, where must a person stand?

How to choose the form

Ask two questions about each task.

How hard is it to undo? Sending an email to a customer, paying an invoice, signing a contract, changing access: these are hard or impossible to reverse. They need approval before. Marking a task done, re-sorting a list or drafting a document: these are easy to reverse. Review after is enough.

How much does a mistake cost? A wrong label on an internal task costs a minute. A wrong figure in an investor update costs trust. Raise the level of human involvement as the cost rises, even for reversible work.

Cheap mistake Costly mistake
Easy to undo Watch Review after
Hard to undo Review after, or approve Approve before

What a good approval request looks like

When automation stops for a person, the request should make the decision quick. A good request has:

  • What it wants to do, in one line.
  • Why, in a sentence.
  • The evidence, as a link to the draft or the record.
  • What happens if you say no, or a way to send it back with a note.

A request without a reason forces you to redo the work to check it, which defeats the purpose. On Operators, an agent that needs you flags an item with its reason and a link, and you choose Approve or Send back with a note. The agent reads the outcome next time it runs.

The rubber-stamp problem

The biggest risk is the person who stops reading. If an agent asks fifty times a day and is usually right, approving without reading becomes the easy path. The loop is then decoration.

Three ways to keep it real:

  1. Fewer requests. Move reversible work to review after. Keep approvals for what truly needs them.
  2. Better requests. Insist on the reason and the evidence. A request you cannot judge in a minute is a request that needs a better format.
  3. Track your send-backs. If you never send anything back, either the automation is excellent or you are not reading. Check a sample by hand now and then.

Review after, done properly

Review after works only if the review happens and the work can be undone. That needs two things:

  • A record of every change, with what it was before, who made it and when, so you can see the work and put things back.
  • A fixed time to look, such as a weekly review, so it does not depend on remembering.

Without the record you are guessing; without a fixed time the review does not happen.

Watching, without staring

Watching is the lightest form, and the easiest to neglect. It works when the automation tells you when something unusual happens: a run that failed, a job that went quiet, a sudden burst of activity. Look for tools that show you the rhythm of each agent, so silence and noise are both visible.

Where people belong in operations

In most small companies, these are the points where a person should stay in the loop, whatever the automation:

  • Money leaving the company.
  • Anything sent to customers, investors or the public.
  • Changes to people’s access, contracts or pay.
  • Legal filings and signatures.
  • Deleting data.

Everything else can usually move to review after or watching, once the automation has earned it. The article a founder’s guide to approvals applies the same thinking to the people on your team, not only to agents.

A simple test

For any automated process you run, you should be able to answer three questions quickly. Where does a person act? What do they see when they do? How would you undo a mistake? If any answer is unclear, the loop needs work.

What to do next

Questions people ask

What does human-in-the-loop mean?
It means a person stays involved in an automated process at chosen points, usually to approve, correct or oversee what the automation does, rather than letting it run with no human judgement at all.
What is the difference between human-in-the-loop and human-on-the-loop?
In the loop, a person must act before the process continues, such as approving a payment. On the loop, the process runs on its own while a person watches and can step in.
Does human-in-the-loop slow everything down?
It can, if every step needs approval. Put the person only at the points that are hard to reverse, and let reversible work run with review afterwards.
How do I stop approvals becoming a rubber stamp?
Keep requests few, make each one carry a clear reason and the evidence, and track how often you send work back. If you never do, you may not be reading closely.

See every company you run on one screen.

Try the demo first: no sign-up, and every screen works on made-up data. You can't connect your own tools to it.

Join the list