A human in the loop is not yet a control.
In their article for ISPE iSpeak of 4 September 2026, Frank Henrichmann and Oliver Herrmann (QFINITY) show why Human-in-the-Loop in AI-supported GxP processes often offers less protection than is assumed in practice: anyone who only reviews results instead of producing them moves into a purely supervisory role, in which errors are noticed less often. From this, the article derives a multi-layered control concept.
Human-in-the-Loop comes with preconditions.
Human-in-the-Loop (HITL) is a risk-based control concept in which qualified people review, approve or override the results of automated or AI-supported GxP processes. Drawing on the research on human factors and automation bias, Henrichmann and Herrmann test whether HITL is inherently a reliable control. The conclusion is sober: as soon as the human moves from active analysis into a purely supervisory role, the likelihood of catching errors declines. Whether the control works is decided by its design.
What hollows out Human-in-the-Loop.
Where the psychology becomes operational.
Two examples from computerized system validation show how the effects would play out.
AI-supported test review
An AI-enabled tool rates an OQ result as passed because the acceptance criteria are formally met. The result looks correct, and the reviewer approves it. Nobody has checked the unspoken assumption that the dataset represents the production configuration. An outdated configuration state is technically a simple error and still hard for the reviewer to see. The risk grows when one language model generates test cases and a second one reviews them, because both can follow similar patterns. Tests from language models also often degrade into smoke tests that confirm the flow without verifying the requirement behind it.
Traceability agents
An agent that cross-checks requirements, specifications and test artifacts reports the traceability of an audit trail requirement as complete: a test case references the requirement, and the test mentions the function. The green status says nothing about whether completeness, immutability and attribution of the audit trail were verified at all. The clean artifact still counts as evidence.
The obvious reaction, adding further review stages, does not solve the problem. In controlled experiments, correction rates fell as soon as corrections took effort. Human review then risks becoming a ritual, in the words of the article formally compliant but substantively hollow. As the system matures, the visible error rate may also fall while the review volume grows. That need not mean the risk is decreasing; it may mean that errors have become harder to detect. Which reading is true only becomes clear by looking at model quality, the review process and the control design.
Detectability is designed in, not assumed.
The regulatory direction is unambiguous. The AI Act, Regulation (EU) 2024/1689, requires effective human oversight for high-risk AI systems, while EU GMP Annex 11 requires clearly defined responsibilities and risk-based controls. The draft Annex 22 identifies human oversight as a key control mechanism. An April 2026 FDA warning letter to Purolea Cosmetics Lab cites the use of AI agents that generated GMP records without sufficient verification. The authors answer with a layered model. HITL remains one layer in a control system of people, processes and technology, and the technical controls of that control system keep working even when vigilance declines. How vendors and regulated users divide these requirements in practice is the subject of our recap of the SAP expert workshop.
| Risk driver | Typical failure of the control | What compensates for it (examples after Henrichmann/Herrmann, ISPE iSpeak 2026) |
|---|---|---|
| Automation bias | Proposals from the system are adopted without anyone challenging them | Mandatory own assessment and justification, review by a second independent party, guardrails in the system |
| Plausibility trap (cognitive ease) | The reviewer confirms what looks plausible instead of analyzing it | Acceptance and rejection must be justified; the interface shows uncertainty instead of traffic lights |
| Review fatigue | Approvals become routine at high throughput | Limited review volumes per person, rotation of reviewers, risk-based triggers instead of full review |
| Human-AI loop | Documentation and decisions imperceptibly align with the logic of the system | Monitoring of drift and data quality, periodic re-verification of the training data |
| Attitude effects | A person's attitude toward automation shapes review behavior | Building AI literacy, peer review, changing reviewers |
Behind this sits the risk logic of ICH Q9(R1): severity, probability, detectability. Human-in-the-Loop only ever claims the third factor, and that is exactly the one that has to be demonstrated. Guardrails belong in the operational frame of the system, meaning system prompts, orchestration, tool rules and execution constraints. Their execution and every exception are logged. That keeps them observable and open to review, and their effectiveness is demonstrated across the lifecycle. Only then do guardrails ease the burden on human review, following the same logic as transferring a human task to automation: the control moves, the evidence stays. That evidence does not end at release. In operation it continues through ongoing process monitoring with drift and data quality control. For results that can be objectively derived from source data and reproduced, human review can be lighter; where the system delivers conclusions, it becomes more intensive. Classifying a result as objectively derivable does not replace the review: results flagged that way can trigger a particularly strong automation bias.
From the practice of the GAMP committees.
Frank Henrichmann, Senior Executive Consultant at QFINITY, is Chair of the GAMP Global Steering Committee and a member of the global GAMP AI SIG. Oliver Herrmann, founder and CEO of QFINITY, is Chair of the ISPE GAMP Europe Steering Committee and a member of the ISPE AI Community of Practice. The article picks up the argument both began in Pharmaceutical Engineering in January/February 2026: AI as the next stage of validation, with the expert who remains indispensable in the loop.
From oversight to demonstrable control.
Is your Human-in-the-Loop a control or a gesture?
Bring one AI-supported process, such as test review or traceability. In the initial consultation we map it to the mechanisms and name the technical control that is missing once attention fades. Free of charge, about 30 minutes, directly with one of the authors.
Book an initial consultation
