The right tool in the wrong hands can have disastrous effects. Something similar can happen with APIs. When they're used inside uncontrollable workflows, the results can be totally unpredictable. Companies like Replit know very well what a rogue workflow is. In 2025, one of its AI agents managed to gain access to a live database and completely delete it. I'm not saying something like this will happen to you. But it could, right? That's why it's always better to be prepared and take every precaution to avoid risks. So, how do you make sure that AI-driven workflows are always under your control? Stay with me to learn more.
This article is brought to you with the help of our supporter, n8n.
n8n is the fastest way to plug AI into your own data. Build autonomous, multi-step agents, using any model, including self-hosted. Use over 400 integrations to build powerful AI-native workflows.
Do you know what’s great about workflows? They’re predictable. You understand what makes them succeed or fail. What’s not so great is that when you add non-predictable nodes, you lose the ability to explain their behavior. Especially with AI nodes whose task is to interpret API responses and decide what to do next. That happens because AI nodes behave like black boxes. You give them an input, they do their own thing, and decide what API operation to call next. It’s this decision step that is the riskiest. Leaving the decision purely to the AI node can put you in an undesirable situation. So, how do you avoid that? Fortunately, there’s a way.
One solution is to stop giving AI nodes full permission to decide. Since there’s an inherent risk in their decision process, you need to step in and guide them. That’s what’s called a human in the loop, or HITL. Every time an AI node makes a decision, it will ask you for approval before it makes a request to an API operation. You’ve probably seen similar approval requests whenever you’re using local AI agents like GitHub Copilot. Whenever it wants to make changes to a file, for instance, it will ask you for approval. Or, whenever it has several options, it asks you for clarification. Something similar is also possible in AI workflow nodes. Let’s look at three tools that let you use HITL.
Zapier, a popular iPaaS, offers a dedicated HITL workflow step you can easily use. Well, it actually lets you set up four types of HITL steps. The first, and perhaps the most popular one, is the approval flow. It simply sends a message to a person and asks for approval to continue executing the workflow. Then there’s what they call a confidence-based routing, where certain parts of a workflow are sent to a person for review. That happens whenever some outputs are marked as uncertain, especially when they come from AI nodes. Another type of HITL is the escalation path, which lets you define situations where a person is definitely needed to complete the workflow. And, finally, there’s the feedback loop, where a person is asked to provide an opinion about the output of an AI node. The feedback will be used in future executions to help the AI node improve itself.
Another popular iPaaS that offers HITL is n8n. Workflow steps here are called nodes because n8n uses a visual diagram-style workflow editor. You have essentially two options to pull a human into the workflow. The most popular option is the wait node, which pauses the workflow until a person reacts. Whenever the wait node is executed, it sends a message using a channel such as Slack, asking a person to review the output of the previous node and approve it so that the workflow can continue. Chat channels include Slack, Google, and also n8n’s own chat interface. The second option you have is to create a custom n8n form where you can ask a person to input certain elements that will help the workflow to continue its execution.
Finally, one more example of an iPaaS that offers HITL is Retool. One type of HITL they offer pauses a workflow execution and pulls a person into a full UI. Yes, the main goal is to obtain approval from a human before moving to the next workflow step. However, in this case, you get to build rich UI experiences that help users get the context they need to help them decide if they want to approve the workflow execution. The downside, as you can imagine, is that someone will have to implement the UI. Even though Retool helps you build the application, you still need to have some basic knowledge to make things work.
As much as these options sound great, they don’t come for free. Making one or several steps of a workflow dependent on the response of a person can create bottlenecks. The whole workflow becomes as slow as its slowest step. In the case of HITL, the slowest step will always be the slowest response from a human. Another drawback is related to the way the workflow handles state. While it’s easy to manage in a short-lived workflow, holding execution data for hours or days requires a different kind of platform. Even if you’re using a third-party solution, this ability to keep long-running workflow state will come at a cost. Finally, all those requests for approval will generate fatigue. People will get to a point where they’ll simply approve things without looking. Altogether, these are the most important drawbacks you can face with HITL.
Now you know the difficulties that you might face. Are they worse than living with the risk of losing data or something even worse? That’s up to you to decide. You need to be your own human in the loop now, if you know what I mean. To me, whenever a workflow has nodes that have unpredictable outputs, HITL sounds like a viable solution.
