Human approval: when automation should stop
Automating does not mean no longer deciding. Why approval points, visible steps and the ability to interrupt matter in AI workflows.
Related project: ETUNIM AI Studio
The more a tool does on its own, the more it matters to know what it is doing. It is one of the principles guiding ETUNIM: people decide.
Three questions before automating
- Is a mistake easy to undo? Rearranging a draft is different from sending a message to a customer.
- Can someone see what happened? Without visibility of the steps, an error can go unnoticed.
- Who answers for the result? That responsibility does not pass to the tool.
The harder an action is to undo, the more sense it makes to require approval before running it.
What the references say
Anthropic’s technical guide argues that agents should show the steps they plan, stop to ask for guidance at set points or when blocked, and have stopping conditions such as a maximum number of iterations.
In the European Union, Article 14 of the AI Act deals with human oversight of high-risk systems. Among other points, it mentions the risk of over-relying on a system’s outputs and the ability to intervene or to stop it safely. Not every system is high-risk, and this is not legal advice, but the idea is useful as general good practice.
Gartner, for its part, predicted in 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. It is a prediction, not a fact, but it is a reminder that control is part of a project from the start.
How we think about control
In ETUNIM AI Studio, still at the concept stage, we want it to be possible to see the data between steps, identify errors and set human approvals before important actions. Stay in control is more than a phrase: it is a requirement of the product we want to build.