← All insights
HERMS · INSIGHTS

Human-in-the-Loop AI: Designing Approval Controls for Regulated Pharmaceutical Decisions

Global life sciences AI · Published 16 November 2026 · HERMS Consultancy DMCC

See how HERMS delivers integrated life sciences solutions.

Human-in-the-loop AI in pharmaceuticals means placing competent, authorized people at the points where AI-supported work becomes a consequential decision or action. AI can propose or execute within a defined boundary, but accountability must remain assigned to the appropriate people and processes. Effective approval controls specify the evidence reviewed, the decision being authorized and the exact action permitted—not merely who clicks a button. Applicable law, GxP procedures and intended use determine where approval is required; additional gates may be prudent organizational controls. Part 11, Electronic Records; Electronic Signatures — Scope and Application EudraLex — Volume 4 — Good Manufacturing Practice guidelines ICH E6 Good clinical practice

What Human-in-the-Loop AI Means in Regulated Life Sciences

These terms describe operating arrangements, not universal regulatory categories. A workflow can use all three at different stages. The distinction is when a person can influence the outcome and whether intervention can occur before unacceptable consequences.

Operating conceptHuman involvementSuitable control boundary
Human-in-the-loopReviews or approves before a specified action proceedsConsequential decision requiring prior verification or authorized approval
Human-on-the-loopSupervises bounded execution and can interveneMonitored, appropriately reversible actions where intervention is timely
Human-in-commandSets policy, permissions and conditions for suspensionOrganizational authority over deployment, delegation and continued use

Supervision is not a substitute for prior approval where a required signature or qualified decision is involved. Likewise, approving a system does not pre-approve every future action. EU AI Act Article 14 addresses effective oversight of applicable high-risk systems, including interpretation, awareness of automation bias, override and safe interruption; it does not require a person to approve every output of every pharmaceutical AI tool. Article 14: Human oversight

Design Approval Gates Around Decision Risk—not Around Every AI Output

Classify actions rather than documents. One dossier workflow may auto-classify a file, require verification of a scientific summary and reserve final filing authorization for the approved sponsor process. Ask about consequence, reversibility, patient/product impact, record integrity, uncertainty and novelty.

A practical policy can allow auto-processing for bounded low-risk tasks, verification for recommendations entering controlled work, and explicit authorization before consequential execution. Treat novel products, unfamiliar jurisdictions, missing sources and conflicting evidence as escalation triggers. A model confidence score is not a validated probability of correctness unless demonstrated for the specific task. This is a proposed implementation pattern, not a regulatory classification. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products AI Risk Management Framework

The Approval Gate Test

  • What exact decision or action will this approval authorize?
  • Who has the competence and delegated or legally assigned authority?
  • Can the reviewer verify the source, version, uncertainty and consequences?
  • Will any material change invalidate approval before execution?
  • Can the workflow stop safely and escalate if evidence or authority is missing?

Which Pharmaceutical Decisions Need a Human Approval Gate?

The Decision-to-Approval Matrix below is an illustrative governance proposal. It combines recommended gates with existing role-specific responsibilities where applicable. It does not assert that every jurisdiction mandates manual approval for every listed activity. Actual roles and signatures follow applicable law and approved procedures. Part 11, Electronic Records; Electronic Signatures — Scope and Application EudraLex — Volume 4 — Good Manufacturing Practice guidelines ICH E6 Good clinical practice

WorkflowAI-permitted actionsHuman approval pointAccountable role (illustrative)Evidence captured
Regulatory submission sign-offDraft, reconcile and technically checkAuthorize exact final package before transmissionAuthorized sponsor/regulatory approverPackage version; destination; claims; approval
Safety signal assessmentCompile evidence and propose prioritiesValidate/assess clinical significanceQualified safety/medical reviewerCase basis; limitations; rationale
Case processingExtract candidate fields; suggest duplicatesResolve consequential or uncertain assessmentsSafety process owner/reviewerOriginal report; edits; classifications
Clinical protocol changesDraft alternatives and impact summaryApprove changes through applicable sponsor/ethics/regulatory pathwaySponsor and relevant clinical/statistical rolesVersion comparison; scientific justification
Data queriesDetect inconsistencies; draft queriesApprove consequential interpretation; routine routing may be boundedClinical data owner/investigator as relevantSource linkage; query history
Medical reviewSummarize evidence; flag issuesApprove medical conclusionsQualified medical reviewerEvidence; uncertainty; assessment
Manufacturing deviationsAssemble records; suggest hypothesesApprove investigation conclusionsQuality/process ownerRecord links; impact; root-cause basis
CAPAPropose actions and track tasksApprove actions and effectiveness closureQuality approverRationale; implementation; effectiveness
Batch dispositionCompile release evidence and exceptionsRetain required authorized quality/QP decisionAuthorized quality role/QP where applicableControlled records; certification
Supplier qualificationAggregate diligence and risk flagsApprove qualification or restrictionsSupplier-quality ownerSource reliability; unresolved findings
CGT evidence interpretationMap literature and identify gapsApprove scientific/translational conclusionsRelevant scientific/clinical specialistsProvenance; applicability; limitations
Administrative classificationApply approved tags or routingException review; no gate needed for all low-risk tagsProcess ownerRule/configuration; transaction history

Blueprint for an AI Approval Workflow

Use an explicit state machine: intake → source/data check → recommendation → policy/risk check → authorized review → execution → reconciled history → periodic review. The controls below are proposed engineering requirements that make the organizational policy enforceable.

  • Intake: assign a unique item identifier, product/context and urgency; preserve the original input.
  • Source/data check: verify provenance, permissions, version and required fields; stop or escalate missing critical information.
  • Recommendation: label AI-generated content, supporting evidence, assumptions and unresolved uncertainty.
  • Policy/risk check: apply approved rules outside the model; determine the role, approval scope and permitted tools.
  • Review: offer explicit approve, reject or request-clarification decisions; silence and elapsed time are not approval.
  • Execution: bind approval to the exact content, target, version and action; reject stale approvals and re-check authority.
  • History and reconciliation: record the executed result and acknowledgement, not only the intent; retain tamper-evident history.
  • Periodic review: sample approved and rejected work, review exceptions and reassess whether the gate remains effective.

Implementation detail matters. Use duplicate-safe execution identifiers so retries do not create a second filing or case. If content changes after approval, return it for reassessment according to defined materiality rules. An audit trail can show that a flawed process ran exactly as designed; it is evidence, not proof of compliance.

Decision Rights, RACI and Segregation of Duties

Define who proposes, reviews, approves, administers and audits. A service account may prepare work but should not grant itself approval rights. Delegation needs scope, effective dates, competence and a deputy—not shared credentials. Use dual approval where justified or required, not as a universal ritual.

ActivityResponsibleAccountableConsulted / independent check
Define approval policyProcess ownerDesignated business/quality ownerRegulatory; Medical/PV; Legal; security
Prepare recommendationAI operator/system ownerProcess ownerDomain specialist
Approve regulated conclusionAuthorized domain reviewerRole assigned by applicable procedure/lawRelevant cross-functional experts
Administer accessSystem administratorSystem/service ownerSecurity; Quality; access reviewer
Release configuration changeTechnical/change ownerAuthorized change approverQuality and affected domain owners
Audit control effectivenessIndependent assuranceAssurance leadProcess and technical owners

This RACI is illustrative. One named accountable role should own each defined approval decision, while professional or statutory responsibilities remain separately identifiable. Administrators should not be able to rewrite approval history without detection. Emergency access must be time-limited, justified and independently reviewed. Part 11, Electronic Records; Electronic Signatures — Scope and Application EudraLex — Volume 4 — Good Manufacturing Practice guidelines

Electronic Records, Signatures and Inspection-Ready Evidence

A workflow confirmation is not automatically a legally meaningful electronic signature. Determine whether the record and signature fall within applicable requirements. FDA’s final Part 11 scope guidance is nonbinding; Part 11 itself remains in effect where applicable, together with underlying predicate rules. Its limited enforcement discretion is not a blanket exemption from record integrity, authorized access or signature controls. Part 11, Electronic Records; Electronic Signatures — Scope and Application

EU GMP Annex 11 addresses computerized systems, while Annex 16 addresses Qualified Person certification and batch release. Automated checks cannot replace required authorized certification. Apply controls to the regulated activity rather than branding software “GxP-compliant” without an intended-use assessment. EudraLex — Volume 4 — Good Manufacturing Practice guidelines

An approval evidence packet should connect the original source and version; relevant system/model/configuration identifiers; material inputs and output; policy checks; reviewer identity and role; timestamp with defined time-zone reference; decision and rationale; approved payload; execution result; and subsequent corrections. Record what is necessary to reconstruct the decision without indiscriminately retaining sensitive data.

Keep evidence in, or reliably linked to, the designated system of record. A separate chat transcript is insufficient if the approved output cannot be identified in the controlled record. Protect retention, accessibility and change history; demonstrate that exported evidence preserves content and meaning. Part 11, Electronic Records; Electronic Signatures — Scope and Application

Exceptions, Overrides and Human-AI Disagreement

Route missing sources, unexpected inputs, contradictory evidence and permission failures to named exception queues. Assign urgency from the business consequence, not from how easy the item is to resolve. In safety workflows, review queues must not defeat applicable reporting timelines; define an approved urgent fallback.

Require a proportionate reason for material overrides and preserve the original recommendation. Neither automatic acceptance nor automatic rejection of disagreement is appropriate. Distinguish a reviewer correction from an authorized exception to policy; the latter needs the designated escalation route. Track near misses as well as executed errors. AI Risk Management Framework

Hypothetical example: an agent proposes approving a supplier because a certificate appears current, but the reviewer finds that it covers another manufacturing site. Reject the recommendation, preserve the mismatch and investigate similar outputs. Correct the source-matching control through change management rather than allowing the model to alter its own approval rules.

Rollback is useful only when the effect is reversible. An external communication may need correction and impact assessment, not silent deletion. A failed execution should remain visible until reconciled; a retry must not reuse approval for materially changed content or a different destination.

Validation, Monitoring and Change Control

Define intended use and acceptance criteria before testing. Validate both the AI-supported task and the approval mechanism: can an unauthorized role approve, can a stale approval execute, can a rejected item proceed, and can the evidence be reconstructed? Challenge missing sources, conflicting versions, tool failures and concurrent edits. Part 11, Electronic Records; Electronic Signatures — Scope and Application EudraLex — Volume 4 — Good Manufacturing Practice guidelines

Test reviewers as part of the workflow. Seed consequential errors and assess whether users detect them with the supplied evidence and realistic workload. Training should cover model limitations, escalation and automation bias—not merely interface operation. Oversight is weak if a competent person lacks time, evidence or authority. Article 14: Human oversight

Control changes to models, prompts, retrieval indexes, permissions, routing rules and integrations. Assess impact and rerun relevant regression tests. Monitor performance drift and review load, with thresholds for reduced autonomy or suspension. Periodic review complements—not replaces—change-triggered reassessment. AI Risk Management Framework

Regulatory Landscape: FDA, EMA, EU AI Act and GxP

SourceStatus / scopeApproval-control implication
FDA drug/biologic AI guidanceJanuary 2025 draft; nonbinding; not for implementationContext-of-use credibility; no general permission for autonomous approval
EMA AI lifecycle reflection paperAdopted September 2024; EU scientific reflection paperLifecycle risk assessment and appropriate accountability
EU AI Act Article 14Binding provision for applicable high-risk systems; phased applicabilityEffective risk-proportionate oversight, interpretation, override and safe stop
FDA Part 11 / scope guidanceRegulation where applicable; September 2003 guidance is nonbindingAssess records/signatures, identity, authority and predicate rules
EU GMP / ICH GCPActivity- and jurisdiction-dependent expectationsRetain applicable quality and sponsor/investigator responsibilities
NIST AI RMFVoluntary framework; AI RMF 1.0 revision underwayOrganize oversight, evaluation and monitoring without claiming certification

These sources are not interchangeable legal instruments. FDA’s January 2025 document remains explicitly draft, EMA’s reflection paper is adopted, and NIST’s framework is voluntary. ICH E6(R3) Principles and Annex 1 are effective in the EU from 23 July 2025; Annex 2 from 15 January 2027. Regional implementation differs. Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products AI Risk Management Framework ICH E6 Good clinical practice

As of 16 November 2026, the Commission reports general AI Act application from 2 August 2026, with revised high-risk dates of 2 December 2027 for Annex III and 2 August 2028 for Annex I product-related systems. Classification depends on the actual system, purpose and relevant provisions; not all pharmaceutical AI is high-risk. Medicines and GxP obligations remain separately applicable. The timetable here is attributed to the Commission summary rather than a complete consolidated-text legal assessment. AI Act

Implementation Roadmap and Oversight KPIs

Days 1–30: inventory consequential actions, identify authorized roles and map current records, signatures and exceptions. Days 31–60: configure gates, evidence packets, permissions and fallback paths; test against prespecified acceptance criteria. Days 61–90: complete human acceptance, resolve critical findings and enable monitored production within approved limits. These are planning windows, not guaranteed deployment times.

KPIDefinition / limitation
Override rateMaterial changes/rejections divided by reviewed recommendations; low rates can reflect complacency
Escalation rateItems escalated divided by eligible items; segment by cause and urgency
Review turnaroundTime from review-ready status to accountable decision; inspect urgent outliers
Unsupported recommendation rateSampled recommendations lacking sufficient supporting evidence
Audit-trail completenessSampled decisions containing every required evidence element
False-negative escalationAdjudicated items needing escalation that the workflow missed
Post-approval correctionsApproved items later corrected for material error; distinguish severity

Establish denominators, sampling, owners and intervention thresholds before launch. Pair speed with independent quality review and queue capacity. A faster gate is not better if reviewers miss consequential errors or approve items outside their competence.

How HERMS Supports Human Accountability

“Model proposes; rules engine decides; audit trail records” is a proposed HERMS operating approach, not a claim of validated deployed capability. Here, the rules engine decides routing and whether execution conditions are satisfied; it does not replace a person’s required professional or legally assigned judgment.

Under this approach, recommendations remain linked to verifiable evidence, role-based gates control consequential actions, and uncertainty triggers escalation. HERMS describes solution areas covering regulatory work, safety, manufacturing and advanced therapies. Their suitability and actual approval controls must be established for the intended deployment, not inferred from portfolio descriptions.

Explore HERMS Solutions to discuss a demonstration focused on your approval boundaries, evidence requirements and accountable roles.

Frequently Asked Questions

Does every AI output need human approval?

No. Bounded low-risk processing may be automatic within approved limits. Prior approval belongs where required by law or procedure, or justified by risk. Document which actions are auto-processed, verified or explicitly authorized; do not equate a generated draft with permission to act.

Can an agent approve its own recommendation?

Not where independent or authorized human approval is required. As a proposed control, separate recommendation generation from permission checks and approval rights. The model should not grant itself authority, change its own threshold or use another person’s credentials.

Is a confidence threshold enough?

No. Confidence may be poorly calibrated and cannot establish source validity, regulatory applicability or authority. Combine evaluated task performance with hard policy checks, evidence availability and exception triggers. Novel or consequential cases may require review even when the score is high.

What makes an electronic approval defensible?

A defensible record identifies the authorized person, decision, evidence, approved version and subsequent action. Formal electronic-signature requirements depend on applicable rules. A generic approval button or detailed audit log alone does not establish that those requirements were met. Part 11, Electronic Records; Electronic Signatures — Scope and Application

How should urgent exceptions be handled?

Use predefined urgency rules, a named deputy, timed escalation and an approved fallback. Silence must not become authorization. Record what was done and why, then review the exception. Urgency should accelerate accountable handling, not silently remove required safety or quality controls.

How do teams avoid review bottlenecks?

Reserve gates for consequential decisions, route by competence and urgency, provide concise evidence packets and monitor queue capacity. Test whether reviewers can detect errors. Improve source checks and routing rather than responding to congestion by weakening a necessary approval.

When must an approval be repeated?

Repeat it when a material change affects the approved content, destination, action or supporting basis, according to defined procedures. Also reassess when authority expires or a relevant new risk appears. Design the system to reject stale approval rather than expecting reviewers to notice every downstream change.

For related perspectives, the preceding articles address automation boundaries, agentic versus generative architecture, and enterprise AI governance. This article focuses on the execution controls that connect those policies to each consequential decision.

Discuss a governance-readiness assessment or workflow-focused demonstration through HERMS Solutions , with approval rights, exception handling and inspection evidence as explicit evaluation criteria.

Relevant HERMS solutions

Explore HERMS solutions relevant to the workflows discussed in this article: ATTESTRA — integrity and data governance; VERIQIS — evidence integrity and assurance; AXION — manufacturing quality. Evaluate each solution against your intended use, evidence and approval requirements.

Sources and further reading

Explore HERMS life sciences solutions and AI governance and traceability, or request a workflow-focused demonstration.

FOR TECHNICAL ENQUIRIES,
CALL +971505528088

ENGINEERING INTELLIGENCE. REVOLUTIONIZING LIFE SCIENCES.

HERMS is the intelligence platform. Its applications are the decision engines. HERMS expert services support implementation and regulated execution of the HERMS intelligence ecosystem, from regulatory and market access to pharmacovigilance, compliance and quality.

Drug registration timelines by country · SFDA, MOHAP, EDE, EDA, EMA and FDA timelines · GCC central registration