AI glossary · Building and connecting

Human in the loop

Also called: Human approval · HITL

What is human in the loop?

An AI or an AI agent runs a workflow but stops at certain points. There a person checks, edits if needed and approves.

Picture a kitchen: the cook plates the dish, and the head chef checks it at the pass before it goes out. The AI is the cook; you stand at the pass.

How does it work?

When building a workflow, you decide where a person steps in. An example from a painting business (made up):

  • AI prepares: It reads an enquiry, finds matching items and writes a draft quote.
  • Stop for approval: The draft reaches you, for example as an email with “Approve” and “Edit”.
  • Person decides: You check price and scope, correct and approve.
  • Workflow continues: Only now does the quote go to the customer and into the customer database.

Why does it matter for your business?

Because AI can be wrong (hallucination), and mistakes involving money, liability or customer relationships are costly. So approve everything that goes outside or moves money: quotes, invoices, rejections, emails to customers. Let internal, easily fixed steps run, such as sorting enquiries, transferring data or preparing drafts.

In September 2026 Ralf Wintergerst, president of the German digital association Bitkom, said that an agent acting on its own needs reliable data, connected systems and rules on what it may and may not do. Approval points are part of those rules.

How does OsAI use it?

In OsAI workflows a person approves at the important points before anything goes out, such as an email to a customer. You decide with OsAI which steps need approval before the build, for AI automation as much as for an agent.

The free AI agent readiness check shows how ready your business is for agents.

Frequently asked questions

Is human in the loop required by law?

For typical small-business tasks, usually not. Art. 14 of the EU AI Act requires that people can effectively oversee high-risk AI, from 2 December 2027 after the amendment by Regulation (EU) 2026/1744; sorting enquiries or writing drafts is generally not high-risk. Art. 22 of the EU GDPR gives people the right not to be subject to a decision based solely on automated processing that has legal effects or similarly significantly affects them. This is an overview, not legal advice.

Doesn’t approval slow automation down again?

It takes time, but only at the point where you would decide anyway. The AI does the reading, sorting and preparing. The guide on testing an AI tool (in German) shows how a trial run tells you which steps work reliably.

Does an AI agent always need human approval?

Not for every step, but for every step that goes outside or is hard to undo. Start with few permissions and many approvals; loosen them once a step has proven itself.

Sources

Last updated: