← All insights
HERMS · INSIGHTS

Agentic AI in Life Sciences: Which Decisions Can Be Automated—and Which Require Human Approval?

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

Discover the HERMS life sciences platform and our regulatory affairs expertise.

Agentic AI in life sciences can automate bounded tasks such as retrieving approved information, checking document completeness, drafting content and routing work through controlled systems. Decisions affecting patient safety, product quality, regulated commitments or legally assigned responsibilities need proportionate verification and, where required, authorized human approval. The boundary depends on intended use, jurisdiction, evidence and reversibility—not the label “agent.” Organizations should grant execution rights only within documented limits, with traceable evidence, monitoring and effective intervention. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle EudraLex Volume 4 — Good Manufacturing Practice guidelines

1. What makes an AI system agentic?

Generative AI produces content: text, code, images or structured outputs. An agentic system combines a model with objectives, planning, retrieval, tool access and multi-step execution. It may search an approved repository, compare evidence, draft a response and create a review task. Conventional workflow automation follows predefined rules; decision support informs a person without necessarily taking action. These categories overlap: an agent can orchestrate deterministic checks, and a generative model can remain a drafting aid. This article uses “agentic” as an operational description, not a regulatory classification. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

Delegated action changes the risk. A hallucinated requirement can become a wrong filing task; a tool error can update the wrong record. Prompt injection can turn retrieved content into malicious instructions. Data leakage exposes confidential or personal information. Automation bias encourages reviewers to accept plausible outputs; model drift can undermine previously demonstrated performance. Governance must cover the entire system and its permissions, not only the language model. AI Risk Management Framework Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

2. The boundary-setting test

Before allowing any autonomous action, document eight questions. These are proposed governance controls, not eight universally prescribed legal tests:

  • Impact: Could the action affect a patient, product, trial endpoint, safety report or binding commitment?
  • Reversibility: Can it be undone before external reliance or harm occurs?
  • Sign-off: Does an applicable rule, authorization or procedure assign approval to a qualified role?
  • Evidence: Are sources authoritative, current, complete and applicable to this product and jurisdiction?
  • Sensitivity: What personal, genomic, commercial or controlled information will be accessed or transmitted?
  • Traceability: Can an independent reviewer reconstruct the evidence, versions, tool calls and approvals?
  • Monitoring: Can failures be detected quickly enough to prevent unacceptable consequences?
  • Escalation: Who receives exceptions, how quickly, and what happens if that person is unavailable?

If a consequential action is irreversible, evidence is incomplete or required sign-off is absent, stop execution and route it for review. Low uncertainty alone is insufficient: an agent can be confidently wrong. Pre-agreed limits should specify permitted systems, fields, actions and exception conditions. This approach operationalizes risk management rather than creating a new legal exemption. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products AI Risk Management Framework

3. Decision-authority matrix

The matrix below is a proposed operating policy. Human-control points combine recommended safeguards with existing role-specific obligations where applicable; they are not a claim that every jurisdiction requires manual approval of every listed task. Apply the local rule and approved quality system first. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle ICH E6 Good clinical practice EudraLex Volume 4 — Good Manufacturing Practice guidelines Artificial Intelligence-Enabled Medical Devices

Use caseWhat AI may automateHuman-control pointRationale / governing risk
Regulatory intelligenceRetrieve, classify and compare published changesVerify applicability before changing obligationsSource freshness; jurisdiction and product scope
Dossier draftingPrepare evidence-linked text and consistency checksAccountable experts verify claims and approve controlled contentUnsupported claims; incorrect source interpretation
Final filing authorizationAssemble package and run technical checksAuthorized sponsor process approves exact filing before transmissionExternal commitment; signatures and portal authority
Protocol designCompare designs; draft alternativesSponsor, clinical/statistical and ethics/regulatory review as applicableParticipant protection; scientific validity
Eligibility / recruitmentPrescreen approved data against defined criteriaQualified site staff confirm eligibility and consent processMisclassification; privacy; investigator responsibility
Safety intake / triageExtract fields; identify potential duplicates; route urgentlyReview exceptions and consequential classificationsMissed cases; reporting clocks; incomplete narratives
Signal validationScreen patterns and compile supporting evidenceQualified safety personnel validate significanceStatistical association is not established causation
Medical assessmentDraft narratives and assemble relevant evidenceAppropriately qualified medical reviewClinical interpretation; benefit–risk consequences
Deviation / CAPALink records; propose causes and actionsQuality unit approves investigation conclusions and closureFalse causal attribution; ineffective corrective action
Batch releaseCompile records; flag exceptions; support release testingRetain required authorized quality / QP decisionProduct quality; legally assigned certification
Gene-editing strategyCompare literature and propose research hypothesesScientific, biosafety and ethics approval before experimentsOff-target effects; incomplete translational evidence
Market-access analysisRun approved scenarios and evidence comparisonsHEOR / access team verifies assumptions and claimsBiased inputs; unsupported payer value claims
Device change controlDetect changes; execute approved testsResponsible manufacturer assesses regulatory impactIntended use; safety/effectiveness; change-plan boundaries

4. Regulatory affairs and submission authorization

AI in regulatory affairs is well suited to evidence assembly: matching requirements to dossier sections, detecting inconsistent terminology and flagging missing references. Retrieval must preserve document version, publication date, product scope and source location. A proposed requirement should remain a candidate interpretation until reviewed; scraping an authority website does not establish that a rule applies. FDA’s January 2025 draft focuses on credibility for a specified context of use, not blanket permission to automate submissions. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products

Separate drafting, publishing and authorization. An agent may prepare an electronic submission and, under an appropriately controlled process, perform a technical transmission after approval. That does not make the agent the accountable applicant or authorized signatory. Bind approval to the exact package version and destination; any material subsequent change should invalidate that approval. These are recommended controls, while the applicable pathway determines actual filing and signature requirements. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products Part 11, Electronic Records; Electronic Signatures — Scope and Application

5. Clinical research and drug development

AI clinical trials tools can compare protocol options, identify inconsistent criteria and support feasibility analysis. A candidate design is not a protocol approval. Review must test endpoint relevance, statistical assumptions, operational feasibility and participant burden. Likewise, eligibility prescreening should identify candidates for qualified assessment, not silently exclude people on the basis of unverified summaries or missing data. Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle ICH E6 Good clinical practice

ICH E6(R3) anchors trial conduct in sponsor and investigator responsibilities and risk-proportionate quality. In the EU, its Principles and Annex 1 have applied since 23 July 2025; Annex 2 is scheduled to become effective on 15 January 2027. Those dates should not be generalized to every country. Validate the actual trial workflow, including data provenance and escalation, rather than treating a model benchmark as proof of GCP suitability. ICH E6 Good clinical practice

6. Pharmacovigilance: intake is not medical assessment

AI in pharmacovigilance should distinguish clerical preparation from consequential safety judgment. Extraction, language assistance and duplicate suggestions can support case processing. Seriousness, reportability, causality and signal significance can involve medical interpretation and jurisdiction-specific obligations. The proposed control is to route uncertain or consequential decisions to qualified safety staff, while preserving original wording and the basis for every changed field. Do not infer a universal legal requirement for manual review of every case from general AI guidance. Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

Hypothetical example: an agent receives an adverse-event email, records receipt time, preserves the original, extracts candidate fields and flags possible hospitalization. It opens an urgent review task rather than deciding that the report is non-serious because discharge information is missing. A reviewer checks the original and approves the assessment; the controlled workflow prepares the report and tracks acknowledgement. If reporting is time-critical, the escalation pathway must prevent the review queue from becoming a reason for delay. This is an illustrative workflow, not a validated HERMS implementation.

7. Cell and gene therapies

AI in cell and gene therapy can connect genomic evidence, process knowledge and literature to research questions. Gene-editing strategy recommendations remain hypotheses until supported by appropriate experimental evidence and specialist evaluation. An apparently strong sequence-level prediction does not establish clinical safety, manufacturability or acceptable long-term risk. Separate research exploration from authority to initiate experiments or make treatment decisions. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle

Use stricter boundaries where sensitive genomic information, donor identity or product–patient linkage is involved. Recommended controls include restricted access, approved data environments, independently checked identity mapping and explicit experimental approval gates. EU GMP includes dedicated advanced-therapy guidance in Part IV; a generic chatbot assessment is not a substitute for the product’s applicable controls. EudraLex Volume 4 — Good Manufacturing Practice guidelines

8. GMP manufacturing and quality

In manufacturing, agents can assemble batch records, flag atypical results and propose deviation or CAPA investigation paths. They should not turn correlation into a confirmed root cause or close investigations merely because required fields are populated. Quality review should assess evidence, impact, actions and effectiveness against approved procedures. Existing GMP computerized-system and validation expectations remain relevant when an AI component is introduced. EudraLex Volume 4 — Good Manufacturing Practice guidelines

Batch-data review is distinct from batch release. EU Annex 16 addresses Qualified Person certification and batch release; automated evidence preparation does not replace that required responsibility. Real-time release testing is a separate controlled concept, not permission for a general-purpose agent to approve a batch. The Commission’s 2025 consultation proposed revised Annex 11 and new AI Annex 22; the reviewed sources do not establish final adoption of Annex 22, so this article treats it as draft. EudraLex Volume 4 — Good Manufacturing Practice guidelines Stakeholders’ Consultation on EudraLex Volume 4 — Chapter 4, Annex 11 and New Annex 22

9. Regulated medical devices

A back-office assistant and a diagnostic or treatment-related software function may face different regimes. FDA regulates medical devices based on intended use and applicable law, not AI as an abstract technology. An agent incorporated into a regulated device therefore needs an assessment of its actual function, users, patient impact and lifecycle changes. Artificial Intelligence-Enabled Medical Devices

Device change control cannot be reduced to “the model improved.” Changes to models, prompts, retrieval or connected tools may alter performance or intended use. Where a Predetermined Change Control Plan is applicable, its authorized scope matters; it is not unrestricted permission to self-modify. FDA’s resource page distinguishes final PCCP guidance from draft AI-device lifecycle guidance. Artificial Intelligence-Enabled Medical Devices

10. HERMS AI Decision Authority Framework

The HERMS AI Decision Authority Framework below is an illustrative, proposed model—not an official regulatory standard. It does not assert that HERMS currently implements or has validated every level. Assign levels to individual actions, not entire departments or products.

LevelPermitted authorityRequired boundary
L1 — RetrievalRetrieve approved informationSource visibility; access controls
L2 — Evidence analysisCompare and structure evidenceProvenance; quality checks; uncertainty
L3 — Decision recommendationsRecommend an actionVerification before consequential reliance
L4 — Conditional executionExecute defined steps behind approval gatesApproval bound to action, content and destination
L5 — Bounded autonomyExecute predefined low-risk actions autonomouslyDocumented limits; monitoring; intervention; escalation

L5 is not a maturity prize and does not mean unrestricted autonomy. A workflow may use L5 to tag documents, L2 to compare evidence and L4 to transmit an approved package. Required clinical, regulatory, safety or quality sign-off remains outside the agent’s delegated authority. Choosing the lowest sufficient authority reduces the consequences of errors without blocking useful automation.

11. Governance, GxP validation and audit trails

Start GxP AI validation with intended use and system boundaries. Specify inputs, outputs, users, connected systems, allowed actions and unacceptable outcomes. Evaluate representative and challenging cases, including missing information, conflicting sources, multilingual records, malicious content, unavailable tools and incorrect permissions. Establish acceptance criteria before testing, and verify the complete workflow rather than the model in isolation. This is a proposed risk-based implementation approach grounded in existing credibility and quality-system principles. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products AI Risk Management Framework EudraLex Volume 4 — Good Manufacturing Practice guidelines

Maintain an AI audit trail sufficient to reconstruct relevant source versions, model/configuration identifiers, material instructions, tool calls, outputs, reviewer changes, approvals and execution results. Retain decision-relevant evidence under an approved policy; do not log sensitive data indiscriminately or assume hidden model reasoning is available. FDA Part 11 applicability depends on electronic records, signatures and predicate rules. Its guidance includes limited enforcement discretion, not a blanket exemption from trustworthy records or underlying obligations. Part 11, Electronic Records; Electronic Signatures — Scope and Application

Control updates to models, prompts, retrieval indexes and integrations through impact assessment and proportionate regression testing. Give agents least-privilege service identities, separate read from write permissions and treat retrieved documents as data, not instructions. Human oversight is meaningful only when reviewers have competence, time, source access and authority to reject or stop the action. Monitor for automation bias as well as model error. AI Risk Management Framework Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

12. International regulatory landscape

The US FDA’s January 2025 drug/biologic AI document remains labelled draft, nonbinding and not for implementation on the verified official page. Its central contribution is a risk-based credibility assessment tied to context of use. EMA’s September 2024 reflection paper is adopted and spans the medicinal-product lifecycle, but is not itself a universal authorization or prohibition on autonomous agents. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle

The EU AI Act is a separate cross-sector legal framework. Do not classify every pharmaceutical AI tool as high-risk: assess intended purpose, scope, provider/deployer role and the relevant classification route. For systems within applicable high-risk provisions, human oversight is a system and operational requirement; it should not be paraphrased as a universal rule that a person must click approval for every AI output. AI Act — regulatory framework and application timeline

As of 16 November 2026, the Commission’s updated implementation pages report general application from 2 August 2026, with earlier AI-literacy/prohibition provisions from 2 February 2025 and GPAI rules from 2 August 2025. Following the AI Omnibus, the Commission reports Annex III high-risk rules applying from 2 December 2027 and Annex I product-related high-risk rules from 2 August 2028. Its 27 July 2026 announcement confirms entry into force of that amendment package. Older 2026/2027 summaries should not be reused uncritically. The amended operative legal text was not fully retrievable in this research session; legal deadline reliance requires a final legal-text check. AI Act — regulatory framework and application timeline AI Omnibus enters into force

NIST AI RMF and its Generative AI Profile are voluntary risk-management resources, not product approvals or GxP certifications. International GxP principles become operational through local implementation and applicable quality systems; ICH adoption alone does not establish identical worldwide effective dates. Use these resources together without treating them as interchangeable legal instruments. AI Risk Management Framework Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile ICH E6 Good clinical practice

13. Deployment roadmap and meaningful KPIs

Deploy by action and risk, not by a broad promise to “automate regulatory.” First inventory decisions and assign accountable owners. Second map permissions, records, obligations and review points. Third validate in a segregated environment using realistic failures. Fourth run in shadow mode against qualified review. Fifth enable limited production actions with monitored exceptions and a tested stop mechanism. Expand authority only after documented evidence supports the additional risk. These are recommended deployment steps, not regulator-mandated project phases. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products AI Risk Management Framework EudraLex Volume 4 — Good Manufacturing Practice guidelines

Use KPIs that expose safety and control performance: critical-field error rate against adjudicated records; unsupported-claim rate; missed urgent-case rate; exception resolution time; approval-bypass attempts; source-to-claim traceability completeness; and post-change regression failures. Track turnaround time alongside rework and quality, not instead of them. Define denominators, sampling, severity weighting, owners and intervention thresholds before launch. No universal numerical target is defensible across all use cases. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products AI Risk Management Framework

14. Frequently asked questions

Which decisions can AI automate?

Good candidates are bounded, reversible actions with reliable inputs and observable outcomes: retrieval, completeness checks, document tagging and controlled task routing. Execution rights should remain narrower than analytical capability. Apply the boundary-setting test before allowing any write action, and escalate when the source, permitted scope or downstream effect is uncertain. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products AI Risk Management Framework

Which decisions require human approval?

Retain approval wherever applicable law or an authorized process assigns responsibility to a qualified person. As a recommended policy, also require review of consequential clinical, safety, quality and regulatory judgments. There is no single worldwide list covering every AI application; define the boundary by jurisdiction, intended use and risk. Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle ICH E6 Good clinical practice EudraLex Volume 4 — Good Manufacturing Practice guidelines

Can AI submit a regulatory dossier?

It can support assembly, validation and potentially controlled transmission. Distinguish that technical step from sponsor authorization, certification and signature. A recommended gate binds approval to the exact package and destination, with material changes triggering renewed approval. The applicable filing pathway and electronic-record requirements still govern the submission. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products Part 11, Electronic Records; Electronic Signatures — Scope and Application

Can AI approve a medicine batch?

Do not equate automated record review or real-time release testing with the legally assigned release responsibility. In the EU, Qualified Person certification and batch release are addressed in Annex 16. AI may support evidence assessment, but does not replace required authorized certification. Check the jurisdiction and product-specific release framework elsewhere. EudraLex Volume 4 — Good Manufacturing Practice guidelines

How does human-in-the-loop differ from human-on-the-loop?

In this article, human-in-the-loop means a person verifies or approves before the consequential action proceeds. Human-on-the-loop means a person supervises a bounded automated process and can intervene. These are operational descriptions, not universal statutory categories. Supervision is inadequate when intervention would come too late or a required prior approval is absent. AI Risk Management Framework

What does the EU AI Act actually require?

Requirements depend on the system, purpose, actor and applicable provision. Relevant high-risk requirements include oversight and other risk controls, but not all pharmaceutical tools are high-risk. Commission pages report revised high-risk phase-in dates in December 2027 and August 2028. Confirm the amended legal text before relying on deadlines for a specific deployment. AI Act — regulatory framework and application timeline AI Omnibus enters into force

How are AI agents validated in GxP settings?

Define intended use, assess risk and test the complete configured workflow against prespecified acceptance criteria. Include source quality, permissions, exception handling, integrations and human review, not only answer accuracy. Maintain evidence, controlled versions, change assessment and ongoing monitoring proportionate to the regulated activity. Validation is use-specific, not a permanent badge for a foundation model. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products EudraLex Volume 4 — Good Manufacturing Practice guidelines

Is a human reviewer enough to make an agent safe?

No. A reviewer cannot compensate for missing sources, inappropriate permissions, an invisible action or an unmanageable queue. Review needs competence, time, authority and usable evidence. Combine it with technical restrictions, meaningful exception handling and monitoring; otherwise the review step can become ceremonial while automation bias persists. AI Risk Management Framework Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

15. A practical HERMS perspective

HERMS Global presents a connected life-sciences intelligence portfolio spanning discovery, development, regulation, manufacturing, market access and patient safety. The practical evaluation question is whether a proposed deployment can demonstrate evidence provenance, role-specific authority and appropriate controls for its intended use—not whether it carries an “agentic” label. This article does not independently validate platform features or regulatory suitability. AI Life Sciences Software | 25 HERMS Solutions

Explore HERMS Solutions and use the relevant solution enquiry to request a demonstration focused on your workflow, approval boundaries and evidence requirements.

Relevant HERMS solutions

Explore HERMS solutions relevant to the workflows discussed in this article: RULEXIS — regulatory strategy; SYNCRA — submission intelligence; VIGILIX — safety signal intelligence; GENIX — gene therapy and genome editing. 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