Approval Loop

An approval loop is a built-in checkpoint within an automated process. The program works autonomously up to a certain point and then stops. It shows a human what it would do next and waits for a yes or no. If the human says yes, the process continues. If they say no, the process is aborted or the program works out a new proposal. The term literally means a loop of approval: proposal, review, feedback, and, if necessary, a new proposal.

Why software needs a handbrake

Modern text programs built on large language models generate proposals that sound convincing. Yet they can be factually wrong without this being apparent from the tone. If such a proposal goes straight into an invoice, a termination notice, or a bank transfer, real damage can quickly result. The approval loop is the point at which someone with responsibility is inserted into the process.

This is not only about errors but also about accountability. If a bank rejects a loan application, a human must be able to answer for that decision. The European AI Act explicitly requires human oversight for high-risk applications. A documented approval step is a simple way to demonstrate this.

Conversely, every approval step comes at a cost. It takes time and ties up staff, which is exactly why it is a subject of debate. Those who allow too much to be approved lose the benefit of automation. Those who allow too little risk mishaps that end up costing more than the staff time saved.

From proposal to approval

Technically, an approval loop consists of three building blocks. First, a point in the process where the program is required to stop. Second, a display that presents the planned step to the human in an understandable way. Third, a record that logs who decided what and when. This log is often more important during audits than the decision itself.

Think of it like a lock on a canal. The water does not simply flow through but collects in front of the gate. Only once someone opens the gate does it continue on. Without the lock, the canal would be faster, but no one could control the flow.

In practice, approvals are usually tiered by risk. A text proposal for an internal memo goes through without review. A payment of 500 euros requires one signature, one of 50,000 euros requires two. This gradation is what distinguishes an approval loop from a fixed dual-control principle, where everything is always checked. A common misconception is that any human involvement already constitutes an approval loop: someone who merely reviews samples after the fact is not controlling anything, just watching.

Approval loops in chatbots and enterprise software

Currently, the term is most often encountered with AI agents. These are programs that not only respond but also carry out steps themselves, such as editing files or sending emails. Coding tools like Claude Code or Cursor therefore ask for confirmation before every action. The screen displays the planned command, with two buttons below for approval and rejection.

The principle is old even outside of AI. Anyone requesting vacation at a company triggers an approval loop: request, supervisor, approval. What is new is that now a piece of software makes the request rather than a human. Enterprise software providers such as SAP or Microsoft are therefore reworking their existing approval chains to handle AI-generated proposals.

In business news, the term often appears as an argument. Companies emphasize their approval loops to reassure customers. Critics then point out that when faced with hundreds of requests a day, people reflexively click approve. This fatigue problem is considered the greatest weakness of the approach.