What does human-in-the-loop mean?
At defined control points, a person reviews the AI result. They can approve, correct fields, answer follow-up questions or reject an action. This control is a binding part of the workflow.
Four forms of human control
Approval
A result is passed on or an action performed only after confirmation.
Correction
A person corrects individual fields before the workflow continues.
Clarification
The agent detects missing information and requests a specific decision.
Escalation
Risk, uncertainty or technical failure leads to a defined manual handling route.
When must a person review?
The decision cannot depend on a model score alone. Important factors include financial impact, legal commitments, privacy, reversibility, safety and whether the result can be checked objectively. An external commitment or order normally needs stricter controls than an internal summary.
| Risk level | Example | Suitable control |
|---|---|---|
| Low | Internal document tagging. | Sampling and reversal are possible. |
| Medium | Associating an invoice with a case. | Automatic checks; a person decides if uncertain. |
| High | Binding order or external price commitment. | Explicit approval by an accountable person. |
| Not permitted or unresolved | Processing data without a legal basis or necessary permissions. | Do not execute; clarify the process first. |
Review what matters, not everything
If people must fully review every standard case, automation offers little benefit. We therefore define which cases may proceed automatically and when someone must check: for example, missing details, illegible documents, contradictory values or unusual amounts.
What needs to be logged?
- Which case and version were processed?
- Which input data and rules were used?
- Which model or provider was used?
- Which validation rules passed and where were there deviations?
- Who approved, corrected or cancelled the case?
- Which action was actually performed afterwards?
Limitations of this approach
Human review does not replace a well-defined workflow. If people constantly correct confusing AI outputs, the process is poorly designed. Control points must be understandable, quick to use and focused on real risks. Cover arrangements and a manual fallback are also needed when the system fails.
Examples from projects
PictureFix includes quality controls and does not leave final judgement entirely to the last automated check. The process scanner on pauderer.de provides a non-binding initial assessment, not an automated business decision. The AITI learning app combines AI preparation with human validation of learning content.
Is human-in-the-loop always manual?
A person decides at the control point. Fixed rules can ensure that only relevant cases are presented – for example, those containing deviations.
Can the level of review be reduced later?
Yes, if sufficient measurements demonstrate that clearly defined standard cases are identified reliably. Changes must be tested again.
Who is responsible?
Professional and organisational responsibility remains with the process operator. An AI model is not an accountable decision-maker.