Global life sciences AI · Published 16 November 2026 · HERMS Consultancy DMCC
Discover the HERMS life sciences platform and our regulatory affairs expertise.
AI governance in life sciences is the operating model that determines which systems may be used, what evidence they require, who approves consequential decisions, and how failures are contained. A practical framework connects accountable ownership, risk classification, data controls, validation, human oversight and lifecycle monitoring—not simply an ethics policy.
Governance starts with an accountable decision
A pharmaceutical company needs more than an approved list of AI tools. It needs a documented way to decide whether a particular configuration is suitable for a particular activity—and whether the organization can continue relying on it after data, models or workflows change.
The relevant question is not “Is this model compliant?” It is: “Can this system, operating within these boundaries, support this decision with acceptable and demonstrable risk?” FDA’s draft framework similarly emphasizes model credibility for a defined context of use, while EMA’s adopted reflection paper addresses AI across the medicinal-product lifecycle. Neither provides blanket acceptance of an AI architecture or vendor. FDA drug/biologic AI draft guidance EMA AI medicinal-product lifecycle reflection paper
An implementable pharmaceutical AI governance framework therefore connects three records:
- A use-case record defining purpose, owner and limitations.
- An assurance record documenting risks, controls and supporting evidence.
- An operating record showing approvals, changes, monitoring and incidents.
The framework below is a proposed organizational model. Its tiers, committees and approval gates are recommendations—not universally prescribed regulatory structures.
Define the scope before assigning controls
Governance should cover predictive models, generative systems and agents that coordinate actions through connected tools. Include internally developed applications, purchased software, embedded vendor features and employee-configured assistants.
Relevant uses span discovery, clinical operations, regulatory affairs, pharmacovigilance, manufacturing, quality, supplier assessment and market access. EMA’s reflection paper explicitly considers AI from discovery through the post-authorisation setting. EMA AI medicinal-product lifecycle reflection paper
Distinguish informational use from consequential reliance. Summarizing public literature for exploration differs from preparing evidence relied upon in a safety assessment, changing a controlled record or triggering an external submission.
The same model can therefore require different controls in different workflows. Record the actual activity, users, data and downstream decisions—not merely the software name.
Seven governance pillars
1. Use-case inventory and intended use
Owner: the business process owner, supported by the AI technical owner.
Create one inventory entry for each materially distinct use. Record the objective, system configuration, users, jurisdictions, data categories, connected tools, outputs, downstream reliance and prohibited actions.
Write intended use as a testable statement: “Extract candidate fields from approved source documents for qualified review” is more useful than “automate compliance.”
Evidence should include the intended-use statement, system boundary diagram, accountable owner and review date. Document foreseeable misuse, including employees treating a draft recommendation as an approved conclusion. This operational approach supports context-specific risk assessment rather than model-level generalization. FDA drug/biologic AI draft guidance NIST AI Risk Management Framework
2. Risk classification and decision rights
Owner: the process owner, with relevant Quality, Regulatory, clinical/safety and control-function participation.
Assess inherent risk before controls and residual risk after controls. Consider patient and product impact, sensitivity, external effect, reversibility, degree of reliance and whether errors can realistically be detected.
For agents, define permissions separately for reading, recommending, writing and transmitting. Approval to analyze a repository is not approval to modify it.
Evidence should include the risk assessment, permission register and approval route. Existing legal responsibilities and authorized procedures take precedence over the organization’s internal delegation model. NIST AI Risk Management Framework EudraLex Volume 4 — GMP guidelines
3. Data quality, provenance, privacy and security
Owner: the data owner, supported by Privacy and Information Security.
Specify permitted sources, access rights, retention, geographic restrictions and supplier processing arrangements. Preserve source identity, version and relevant context so reviewers can verify claims.
Treat retrieved documents as evidence, not instructions. Untrusted text must not be allowed to expand an agent’s authority. Apply least privilege to service accounts and separate development, testing and production access.
Evidence includes source qualification, data-flow mapping, access review and security testing. Generative-AI risk management should address information integrity, privacy, security and risks introduced through third parties. NIST Generative AI Profile
4. Validation, testing and assurance
Owner: the technical owner; Quality/GxP determines applicable regulated assurance expectations with the process owner.
Define acceptance criteria before testing. Challenge the complete workflow with representative inputs, missing evidence, conflicting sources, malicious instructions, integration failures and inappropriate permissions.
Separate analytical performance from operational control: an accurate answer does not demonstrate that routing, signatures or tool execution are correct.
Evidence should include requirements, test datasets, results, unresolved deviations and an acceptance decision. Assurance must match intended use; supplier benchmarks alone do not demonstrate suitability for a company’s configured process. FDA drug/biologic AI draft guidance EudraLex Volume 4 — GMP guidelines
5. Human oversight and exception management
Owner: the process owner and the relevant qualified decision-maker.
Specify what reviewers must check, what evidence they receive and when work must stop. Reviewers need competence, sufficient time and authority to reject recommendations.
Define deputies, urgent escalation and handling when required reviewers are unavailable. A human approval step is ineffective if reviewers cannot inspect sources or if the workflow executes before approval.
Evidence includes reviewer training, approval records, exception procedures and testing of review effectiveness. Applicable clinical, quality and regulated-signature responsibilities remain in force when AI supports the work. ICH E6 Good clinical practice EudraLex Volume 4 — GMP guidelines FDA Part 11 — Scope and Application
6. Monitoring, change control and incident response
Owner: the technical service owner, with the process owner and relevant Quality or security functions.
Monitor actual performance, exceptions and permission violations. Control changes to models, prompts, retrieval indexes, rules, integrations and approved sources.
Predefine suspension triggers and a fallback process. Investigate whether an error affected previous outputs, not only the current transaction.
Evidence includes monitoring reports, change assessments, incident records and corrective actions. Lifecycle monitoring should address changing system behavior and operating conditions, consistent with risk-management principles. NIST AI Risk Management Framework NIST Generative AI Profile
7. Documentation, auditability and improvement
Owner: the process owner, supported by records management and independent assurance.
Maintain a reviewable evidence chain linking the source, configured system, output, human decision and executed action.
Retain decision-relevant records under an approved schedule. Do not log sensitive data indiscriminately or assume hidden model reasoning is available.
Evidence includes version histories, access records, material tool events, approvals, audit findings and improvement actions. Electronic-record controls depend on the applicable activity and rules; FDA’s Part 11 guidance expressly distinguishes underlying record obligations from its limited enforcement-discretion policy. FDA Part 11 — Scope and Application
Apply risk tiers without confusing them with legal classifications
The following four tiers are internal governance categories. They are not the EU AI Act’s statutory risk categories.
| Internal tier | Typical characteristics | Proposed approval gate | Escalation trigger |
|---|---|---|---|
| Low | Informational; limited reliance; easily corrected; approved low-sensitivity data | Process-owner approval and basic security checks | New sensitive data, operational reliance or external action |
| Moderate | Drafts or recommendations entering an established review process | Process and technical acceptance; relevant control-function review | Unverifiable evidence, repeated material errors or expanded permissions |
| High | Material influence on regulated evidence, safety, quality or controlled records | Applicable Quality/GxP assurance and domain acceptance before production | Critical miss, unauthorized write, control failure or material configuration change |
| Critical | Potential severe harm, irreversible effects or interference with required authorized decisions | Executive risk decision plus required domain approvals; prohibit unsupported autonomous action | Lost oversight, inability to reconstruct decisions or unacceptable residual risk |
Do not average away a severe consequence because other factors score well. Poor human detectability can make an apparently reversible action dangerous: the organization cannot reverse an error it does not recognize.
A prohibited or unlawful use remains prohibited; an internal committee cannot approve it into acceptability. Legal classification must be assessed separately by intended purpose, jurisdiction and actor responsibilities. European Commission AI Act implementation overview
Make decision authority explicit
The matrix below is an illustrative HERMS governance proposal, not a universal legal rule. Actual approval roles must follow applicable law, authorizations and approved procedures.
| Application | Proposed AI autonomy | Required human role | Key evidence/control |
|---|---|---|---|
| Regulatory submission drafting | Retrieve evidence and prepare candidate text | Regulatory and scientific reviewers approve controlled content | Verified source-to-claim links; version control |
| Final submission approval | Assemble and check an approved package; transmit only within authorized workflow | Authorized applicant/sponsor process retains approval and signatures | Approval bound to package, destination and version |
| Literature screening | Search, prioritize and identify candidate records | Domain reviewer resolves consequential or uncertain classifications | Coverage testing; originals; documented exclusions |
| PV signal evaluation | Compile patterns and propose priorities | Qualified safety/medical personnel evaluate significance | Traceable evidence; uncertainty; escalation |
| Clinical protocol design | Compare options and draft alternatives | Sponsor and relevant clinical/statistical reviewers; ethics/regulatory review as applicable | Scientific justification and participant protection |
| Gene therapy evidence assessment | Map literature and identify evidence gaps | Scientific, clinical and biosafety specialists assess conclusions | Provenance; translational limitations |
| GMP deviations | Assemble evidence and propose investigation paths | Quality approves conclusions, actions and closure | Investigation records; causal justification |
| Batch disposition | Check records and flag exceptions | Applicable authorized quality/QP responsibilities remain | Controlled records; required certification |
| Supplier intelligence | Consolidate evidence and flag concerns | Supplier-quality/procurement owner decides qualification | Source reliability; conflicts; documented decision |
The important control is not the job title alone. It is an explicit decision right linked to evidence, competence and an enforceable workflow. Relevant clinical and GMP responsibilities remain distinct from an agent’s technical ability to generate or execute an action. ICH E6 Good clinical practice EudraLex Volume 4 — GMP guidelines
Assign roles without creating a governance bottleneck
Use a RACI-like model, but retain one identified accountable owner for each decision.
| Role | Principal responsibility |
|---|---|
| Accountable executive | Approves policy, resources and risk appetite; resolves escalations within lawful authority |
| Process owner | Owns intended use, operational outcomes, training and continued suitability |
| QA/GxP | Determines applicable quality-system controls; oversees assurance, deviations and regulated change processes |
| Regulatory | Assesses jurisdictional implications, authority-facing claims and submission controls |
| PV/Medical/Clinical | Owns relevant professional interpretation and safety/clinical review |
| Legal/Privacy/InfoSec | Assesses legal basis, contracts, confidentiality, data protection and cybersecurity |
| Data/AI technical owner | Maintains architecture, configuration, integrations, testing and monitoring |
| Independent assurance | Challenges control design and operating effectiveness without owning routine execution |
A central governance forum should set standards and resolve high-risk exceptions, not approve every low-risk prompt. Otherwise, governance becomes a queue rather than a control.
Document deputies and conflicts of interest. Independence matters particularly when the same team develops the system, evaluates its performance and argues for deployment.
Govern vendors and model supply chains
Vendor qualification should examine the service actually purchased: model access, hosted application, retrieval service, integration provider or combined platform.
Require evidence proportionate to risk concerning:
- Data use, retention, training use and subprocessors.
- Access controls, incident notification and service continuity.
- Model/configuration changes and available version controls.
- Evaluation evidence, limitations and customer responsibilities.
- Exportability of records and exit arrangements.
A supplier certificate is supporting evidence, not a substitute for use-specific assessment. ISO/IEC 42001 concerns an AI management system; it does not authorize a medicine, validate a particular workflow or prove compliance with every pharmaceutical obligation. ISO/IEC 42001:2023 official overview
Contracts should support the operating controls. A policy requiring advance change assessment is ineffective if the service can change materially without usable notification or an acceptable fallback.
Understand what each regulatory resource contributes
| Source | Status/territory | Practical governance implication |
|---|---|---|
| FDA drug/biologic AI guidance | January 2025 draft; US; nonbinding and not for implementation | Assess credibility against a defined context of use; do not call it final |
| EMA AI reflection paper | Adopted; first published September 2024; EU scientific reflection paper | Consider AI risks across the medicinal-product lifecycle |
| EU AI Act | Binding EU legislation with differentiated scope and phase-in | Assess classification, actor responsibilities and applicable obligations separately |
| NIST AI RMF and GenAI Profile | Voluntary, cross-sector resources | Structure lifecycle risk management and generative-AI controls |
| ISO/IEC 42001:2023 | International AI management-system standard | Establish and improve organizational management arrangements |
| ICH/GxP and electronic-record controls | Jurisdiction- and activity-dependent implementation | Preserve applicable trial, quality, validation, record and signature responsibilities |
These resources are complementary, not equivalent. The FDA page remains explicitly draft; EMA identifies its reflection paper as adopted. NIST describes its framework as voluntary, and ISO identifies 42001 as a management-system standard. FDA drug/biologic AI draft guidance EMA AI medicinal-product lifecycle reflection paper NIST AI Risk Management Framework ISO/IEC 42001:2023 official overview
As of 16 November 2026, the Commission reports general EU AI Act application from 2 August 2026, subject to exceptions. Its current implementation page reports revised high-risk dates of 2 December 2027 for Annex III and 2 August 2028 for Annex I product-related systems following the AI Omnibus. Not every pharmaceutical AI tool is therefore automatically high-risk or subject to an identical timetable. European Commission AI Act implementation overview
For EU clinical trials, ICH E6(R3) Principles and Annex 1 became effective on 23 July 2025; Annex 2 becomes effective on 15 January 2027. These are EU implementation dates, not worldwide effective dates. ICH E6 Good clinical practice
The current EudraLex index lists Annex 11, January 2011. Revised Annex 11 and proposed AI Annex 22 were consulted on in 2025; final adoption was not established in the reviewed sources. They should not be presented as effective final requirements. EudraLex Volume 4 — GMP guidelines GMP Chapter 4, Annex 11 and proposed Annex 22 consultation
Validate proportionately and preserve trustworthy records
Not every AI tool is automatically a GxP computerized system. Determine whether its intended use affects a regulated activity, required record, patient safety or product quality.
For relevant systems, connect requirements to tests and tests to acceptance decisions. Include source provenance, access, integration behavior, reviewer controls, failure handling and retention.
Preserve originals and distinguish machine-generated candidate information from approved records. Where electronic signatures apply, maintain the required identity and signature controls; an agent must not impersonate an authorized person.
FDA’s Part 11 scope guidance states that Part 11 remains in effect and that underlying predicate-rule obligations continue despite specified enforcement discretion. A generic “Part 11 compliant” vendor claim therefore does not settle applicability or adequacy for a particular deployment. FDA Part 11 — Scope and Application
Address predictable failure modes
| Failure mode | Specific proposed safeguard |
|---|---|
| Hallucination or incorrect attribution | Verify material claims against authoritative source passages |
| Bias or unrepresentative performance | Test relevant populations, languages and operating conditions |
| Prompt injection | Treat external content as untrusted; enforce authorization outside model instructions |
| Data leakage | Restrict data flows, permissions, retention and provider processing |
| Model or retrieval drift | Monitor quality; control versions; reassess material changes |
| Inadequate traceability | Capture source versions, configuration, material events and approvals |
| Automation bias | Test reviewers; require evidence-based review; sample accepted outputs independently |
| Third-party dependency failure | Qualify suppliers; define notification, continuity and exit controls |
These safeguards operationalize risk management; they are not claims that any single technique eliminates the associated risk. Generative-AI assurance needs layered controls and continuing evaluation. NIST Generative AI Profile
Two illustrative governance scenarios
Incorrect source attribution in a dossier answer
An assistant drafts a response to an authority question and cites a genuine study, but the cited passage does not support its conclusion.
The reviewer rejects the claim before approval. The team preserves the output, assesses whether other sections are affected and investigates retrieval and attribution performance. Corrective action addresses the failure pattern—not merely the sentence.
The governance lesson is that citation presence and citation correctness are different acceptance criteria.
Safety signal prioritization requiring medical review
A system assigns low priority to a potential signal because reports are sparse, while the available narratives suggest a clinically important pattern.
Qualified safety personnel review the evidence and override prioritization. The process owner checks whether similar cases were deprioritized and adjusts monitoring or escalation criteria after appropriate assessment.
The governance lesson is that prioritization supports workload management; it must not silently replace accountable medical evaluation. Both scenarios are hypothetical.
Implement in phases, with evidence at each gate
| Period | Measurable deliverables | Prerequisites |
|---|---|---|
| Days 0–30 | Approved charter; named owners; initial inventory; temporary restrictions for unassessed consequential uses | Executive sponsorship and access to business/system owners |
| Days 31–60 | Risk-assessed priority uses; intended-use records; vendor/data reviews; defined acceptance criteria | Agreed tiering method and relevant control-function participation |
| Days 61–90 | Completed pilot testing; trained reviewers; monitoring, incident and fallback procedures; documented acceptance decisions | Suitable data, controlled environments and resolved critical findings |
| Day 90 onward | Periodic reviews; change-triggered reassessment; supplier oversight; independent sampling; retirement planning | Stable ownership, retained evidence and operating monitoring |
These are planning windows, not promised implementation durations. A high-risk deployment should not progress simply because a calendar milestone has passed.
Measure control effectiveness
Define denominators and evidence sources before reporting:
- Inventory coverage: identified uses recorded against the population established through an approved discovery method.
- Risk-assessment coverage: active uses with current approved assessments divided by active uses.
- Validation completeness: applicable acceptance requirements supported by approved evidence.
- Human-review turnaround: elapsed time to accountable review, segmented by urgency.
- Override rate: materially changed or rejected recommendations, categorized by reason.
- Incident containment time: elapsed time from detection to documented containment.
- Overdue reviews: systems past their approved review date.
- Provenance completeness: sampled outputs with sufficient verified source lineage.
- Audit-trail completeness: sampled tasks containing all required decision and action records.
Pair speed metrics with quality sampling. A falling override rate may indicate improvement—or complacency. No universal target is appropriate across all uses.
HERMS perspective
HERMS Global positions its portfolio around connected life-sciences intelligence across development, regulation, manufacturing, access and safety. Traceable evidence, bounded autonomy, explicit decision rights and human accountability are appropriate principles for evaluating such workflows. This article does not independently establish product validation, regulatory acceptance or deployment performance. HERMS Solutions
Frequently asked questions
What is an AI governance framework for a pharmaceutical company?
It is the combination of ownership, policies, decision rights, controls and retained evidence governing AI use from proposal to retirement. It should connect business processes with applicable quality, safety, privacy and regulatory requirements rather than exist as a separate ethics document.
Who is accountable when an agent makes a regulated recommendation?
Accountability remains with the organization and the applicable authorized roles. Identify the process owner, technical owner and qualified reviewer explicitly. A model-generated recommendation does not transfer statutory, professional or procedural responsibility to the software. ICH E6 Good clinical practice EudraLex Volume 4 — GMP guidelines
Which uses need formal validation or human sign-off?
Assess intended use, regulated impact, records, applicable law and approved procedures. Consequential uses generally warrant stronger assurance. Required clinical, quality or signature responsibilities remain applicable; there is no universal rule that every AI output requires manual approval. FDA Part 11 — Scope and Application ICH E6 Good clinical practice EudraLex Volume 4 — GMP guidelines
Can one risk tier apply across the entire company?
A common method can, but the assessment should remain use-specific. The same platform may support exploratory research and consequential record processing. Classify each configuration and workflow, then reassess when its data, permissions or reliance changes.
Does ISO/IEC 42001 certification prove pharmaceutical compliance?
No. It concerns an AI management system. It does not replace use-specific assurance, applicable GxP obligations, required approvals or a product authorization. Evaluate certification scope alongside deployment evidence rather than treating it as universal acceptance. ISO/IEC 42001:2023 official overview
When should a system be suspended or retired?
Suspend when critical controls fail, unauthorized actions occur or acceptable performance cannot be demonstrated. Retirement should revoke access, disable integrations, preserve required records and assign ownership of unresolved outputs. Test the fallback before it is needed.
To discuss governance readiness, traceable evidence and appropriate decision authority, explore HERMS Solutions and request a workflow-focused demonstration through the relevant solution enquiry.
Relevant HERMS solutions
Explore HERMS solutions relevant to the workflows discussed in this article: ATTESTRA — integrity and data governance; VERIQIS — evidence integrity and assurance; REGAXIS — registration lifecycle intelligence. Evaluate each solution against your intended use, evidence and approval requirements.
Sources and further reading
- FDA drug/biologic AI draft guidance
- EMA AI medicinal-product lifecycle reflection paper
- NIST AI Risk Management Framework
- EudraLex Volume 4 — GMP guidelines
- NIST Generative AI Profile
- ICH E6 Good clinical practice
- FDA Part 11 — Scope and Application
- European Commission AI Act implementation overview
- ISO/IEC 42001:2023 official overview
- GMP Chapter 4, Annex 11 and proposed Annex 22 consultation
Explore the pharmaceutical AI series
- AI decision authority and human oversight
- Agentic versus generative AI in pharmaceuticals
- Human approval controls for pharmaceutical AI
- Real-world evidence for regulatory and market access
- AI medical device validation and change control
- Global pharmaceutical market access, HTA and pricing evidence
Explore HERMS life sciences solutions and AI governance and traceability, or request a workflow-focused demonstration.
