Ideas · Human-in-the-lead · Field note · 3 min read

“Human-in-the-loop” isn’t the same as a human in charge.

Five levels of AI autonomy, and the one question that decides whether yours is safe to ship.

JP Galido · Blank Canvas Thinking · First published June 2026 · rewritten October 2026

“Is there a human in the loop?”

“Yes!”

Great. What does that human actually do?

“They… look at a dashboard. Sometimes.” 😂

Here’s the thing. “Human-in-the-loop” gets used in meetings like it means one thing. It doesn’t.

To one person, it means somebody approves every action. To another, it means somebody glances at a report once a day.

Same words. Completely different risk.

So before we argue about how much freedom an AI system should get, let’s get the words straight. There are four common levels. And a fifth one people keep forgetting.

Autonomy ladder Four levels from AI-in-the-loop to human-out-of-the-loop. Running down the side of every level: human-in-the-lead. AI-in-the-loopHuman acts · AI assistsHuman-in-the-loopAI proposes · human approves eachHuman-on-the-loopAI acts · human supervisesHuman-out-of-the-loopAI acts alone · audit after HUMAN-IN-THE-LEAD ↓ less human effort per action · same human accountability →
Four levels of autonomy. Human-in-the-lead runs alongside all of them.
  1. 1AI-in-the-loop. You do the work. The AI helps. It drafts, finds, summarizes, suggests. Nothing happens that you didn’t do.
  2. 2Human-in-the-loop. The AI proposes. A person approves every action. The invoice gets drafted and queued. Somebody clicks approve before it goes out.
  3. 3Human-on-the-loop. The AI acts. A person supervises. Routine stuff flows through. Weird stuff gets escalated. There’s a real stop button.
  4. 4Human-out-of-the-loop. The AI runs on its own. People set the rules and check afterwards. Only OK when mistakes are small and reversible.
  5. 5Human-in-the-lead. A named person is in charge. Always. They set the intent, own the outcome, and can stop it. Whatever level you picked.

The real question isn’t where the human sits in the loop. It’s whether a human is in the lead.

You can be on the loop, or even out of it, and still be firmly in charge: you set the direction, you own the results, and you have a stop button that actually works.

And you can be in the loop and not in charge at all. Clicking “approve” on things you don’t understand isn’t control. It’s theater. It looks like safety. It isn’t.

So how do you pick a level? Don’t start with the technology. Start with the decision, and ask three questions:

  • How bad is a mistake? A typo, or a wire transfer?
  • Can you undo it? Or has the email already left the building?
  • How many are there? Ten a day, or ten thousand an hour?

Big mistake, can’t undo it? A person approves every action.

Lots of routine, easy to undo? Let it run, watch it properly, keep the stop button handy.

Huge volume, small reversible mistakes, clear rules? It can run on its own, with an audit afterwards.

And in every case, somebody’s name is on it. If nobody can stop it and nobody owns the outcome, you didn’t ship an autonomous system. You shipped an orphan. 🤔

One more honest note. No level is safe on its own. Approvals turn into rubber stamps. Dashboards go unwatched. Mistakes turn out bigger than you thought. The label is a promise. Your job is to make the system keep it.

In an AI-DLC, this is the test-and-approve stage. Business AI Architecture decides the level for every decision before anything gets built. That’s how you keep a human in the lead from canvas to customer.

Try this Monday

Pick one AI-assisted decision in your business. Ask: who can stop it?

If nobody can answer in ten seconds, you’ve found your first fix.