Pharma AI glossary
Human-in-the-loop (HITL)
An oversight design where a human reviews or approves AI output before it takes effect.
Definition
An oversight design where a human reviews or approves AI output before it takes effect. Regulators increasingly distinguish real HITL from rubber-stamping: Annex 22 requires the human be qualified, with defined responsibilities, and, remarkably, that their "training and consistent performance should be monitored"; FDA treats human review as part of the COU that can *reduce* model risk only if the review is genuine. The jagged-frontier research explains why: unexamined approval decays into "falling asleep at the wheel," and an unmonitored loop is a fictional control.
Oversight, not rubber-stamping
HITL means a qualified human reviews, approves, or overrides agent output before it affects the record or decision. The workflow gate before QMS or batch-record commit is the control across investigations, CAPA, release support, and change control; Annex 22 and FDA credibility both assume oversight scales with risk.
Design patterns
Centaur (divide and judge) and cyborg (interleaved editing) patterns from BCG preserve human agency; passenger mode (accept fluent output wholesale) is the anti-pattern. Routing low-confidence predictions to humans is necessary but not sufficient. Reviewers need literacy and time.
Frequently asked questions
Does HITL satisfy EU AI Act requirements?
It supports deployer obligations for oversight proportionate to risk, but must be documented, trained, and evidenced, not an informal glance.
When can AI run without HITL?
When context of use is low consequence, monitoring is strong, and rulebooks allow. Examples often include predictive maintenance alerts or environmental trend flags routed for human follow-up, not automatic batch disposition. Not by default for critical GMP decisions under draft Annex 22’s static/deterministic bias.
Which GMP decisions always need human review?
Batch release and disposition, critical manufacturing parameter changes, official QMS commits, and any generative text entering Part 11 records. Documentation drafts and search results still need qualified review proportional to how the output will be used.
Related terms
Benchmark your AI literacy
Free 20-question assessment, eight modules, and instant scoring, built for pharma quality and manufacturing teams.
Open the Literacy CenterBrowse all terms
See the full 71-term glossary index or search interactively in the Literacy Center.
Back to glossary hub →Educational content only, not legal or regulatory advice. Regulatory guidance cited here includes drafts (FDA AI credibility guidance; EU GMP Annex 22) as of August 2026; verify against final texts before relying on them in submissions. Company-reported figures (Merck CSR timings, Sanofi results) are labeled where used. MIT's ~95% pilot figure carries its own caveat (~150 interviews, contested definitions, not peer-reviewed). Re-check sources on module finalization.