What Is Human-in-the-Loop and When to Use It?

One of the most important processes in AI workflows is HITL (Human-in-the-Loop). This essentially means requiring a human to verify AI output or assist the AI in achieving a specific goal. However, the difficult part is using it smartly - finding the balance between over-using and under-using it. Read on to find out how.

What is Human-in-the-Loop?

HITL refers to any point where a human needs to intervene in the middle of an AI process.

In witch cases the human need to intervene AI process? - There can be a many cases the user need to intervene in the middle of ai process:

• Approval: This is the most common form. The user must approve a high-stakes action the AI wants to take, such as approving a deployment plan or deleting a file.
• Security: A more specific type of approval. This applies to sensitive processes you don't want the AI to handle autonomously, like running a production DB migration or logging into a LinkedIn account.
• Complex MVP (The "Wizard of Oz" approach): This is a fascinating concept. Let’s say you are building an incredibly complex application, for example, a tool that creates a two hours Hollywood-quality movie with CGI from a single prompt. That is a massive undertaking. To "sell" the vision to investors while the tech is still developing, you can use HITL. You take the user's prompt, but instead of the AI doing the heavy lifting, humans handle the task behind the scenes. This allows you to validate the market while you build the AI to eventually handle it autonomously.

When should you use HITL?

This is one of the trickiest questions that exists. Why? Because if we overuse this method, we won’t be able to achieve the maximum output from the AI.

For example: If we check the quality of the code after each line, it will take too long; in this case, it would even be faster to write the code yourself.

So, when should we use it? I think the most important time is when you need to understand what the AI is going to do.

For example: after planning or ideation, we must understand what it is going to do. We need to create a workflow where the AI will ask us when it needs us.

My solution

As I said, we need a way to "talk" with the AI like a conversation, and we need to be able to do it from anywhere. This is why I created claude-slack-bridge a way we can talk in the middle of the process with Claude Code.

You can try it out. It has helped me a lot when I needed him to continue exploring something in the middle of a session, or if he found an issue and I wasn't near the computer.

The most important thing to remember from this blog post is that we will always need HITL, we just need to be smart about how we use it.