HERMS white paper
Explore the HERMS approach and our solutions portfolio.
Download the editable Word white paper
HERMS WHITE PAPER | WP5 | NOVEMBER 2026
Governed AI in Regulatory Affairs
A Framework for Human-Accountable Intelligence
Regulators in the United States, Europe, Saudi Arabia and Abu Dhabi now agree on one point: AI can support regulatory work, but people remain accountable for every decision. This paper turns that consensus into a practical framework for regulatory affairs teams in the GCC.
HERMS Consultancy DMCC
Dubai, United Arab Emirates | info@hermsglobal.com | www.hermsglobal.com
Contents
Executive Summary
1. Why Regulatory Affairs Needs Governed AI
2. The Regulatory Landscape
3. Core Concepts
4. The HERMS Human-Accountable Intelligence Framework
5. Risk Tiers for Regulatory Use Cases
6. Human Decision Rights
7. Evidence, Traceability and Audit
8. Generative AI and Large Language Models
9. Implementation Roadmap
10. Pitfalls and Readiness Checklist
11. How HERMS Helps
References
Executive Summary
AI is already part of regulatory work. FDA's drug centre reports experience with more than 500 submissions with AI components from 2016 to 2023.5 In January 2026, FDA and EMA published ten joint principles for good AI practice across the medicines lifecycle, the first of which is that AI should be human-centric by design.1
| 10 joint FDA–EMA principles for good AI practice, January 20261 | 7 steps in FDA's draft risk-based credibility framework3 | 500+ submissions with AI components reviewed by FDA's drug centre, 2016–20235 | Oct 2025 Abu Dhabi Responsible AI Standard takes effect15 |
The direction is consistent across jurisdictions. Saudi Arabia's AI ethics principles require human oversight across the AI lifecycle and identifiable, accountable people.12 Abu Dhabi requires human-in-the-loop or human-on-the-loop oversight, override mechanisms and audit trails.15 EMA states that the applicant or marketing authorisation holder is responsible for ensuring that all algorithms, models and datasets are fit for purpose.7
Six commitments of governed AI
- Define the context of use for every AI application before it is used.1,3
- Set oversight in proportion to risk, based on how much the output influences a decision and what happens if it is wrong.3,4
- Ground every output in sources that a reviewer can check.1,15
- Assign a named human owner to every decision the AI supports.12,15
- Log inputs, outputs, versions and human interventions for audit.15
- Monitor and re-evaluate performance throughout the lifecycle.1,15
| HERMS view In regulatory affairs, the question is not whether AI is accurate enough to decide. It is whether the organisation can show who decided, on what evidence, and why. |
1. Why Regulatory Affairs Needs Governed AI
Regulatory affairs is document-heavy, source-dependent and deadline-driven, which makes it well suited to AI support. It is also a function where an error can affect patient safety, product quality or a regulator's trust in the company.
Where AI adds value
| Activity | Typical AI support | Human role |
| Regulatory intelligence | Monitoring authority websites, summarising new guidance, flagging changes | Confirm relevance and impact; decide action |
| Dossier preparation | First drafts, cross-checking modules, gap analysis against requirements | Review, correct and approve content |
| Labelling | Comparing SmPC, PIL and reference labels; tracking differences | Decide wording; approve local labels |
| Queries and deficiency letters | Locating prior responses and source data | Write and sign the response |
| Pharmacovigilance support | Case intake triage, literature screening | Medical assessment and causality |
| Tenders and pricing | Extracting requirements, checking document completeness | Commercial and compliance decisions |
Where AI creates risk
- Inaccurate output: EMA notes that large language models can produce variable results and irrelevant or inaccurate responses.8
- Data leakage: the same EMA principles highlight potential data security risks and the need for safe input of data.8
- Unclear accountability: if no person owns an AI-supported conclusion, nobody can defend it to a regulator.12
- Silent drift: model performance can change over time, which is why the joint principles call for scheduled monitoring and re-evaluation.1
2. The Regulatory Landscape
There is no single AI rule for regulatory affairs. The framework in this paper draws on the following sources.
| Source | Status | Relevance |
| FDA–EMA Guiding Principles of Good AI Practice1 | Published January 2026; voluntary | Ten principles from human-centric design to lifecycle management; basis for future guidance2 |
| FDA draft guidance on AI for regulatory decision-making3 | Draft, January 2025 | Seven-step, risk-based credibility framework for a defined context of use |
| EMA reflection paper on AI in the medicinal product lifecycle6 | Adopted September 2024 | Considerations for safe use of AI from discovery to post-authorisation |
| EMA/HMA guiding principles on large language models8 | Published September 2024; updated regularly | Safe input, critical review of outputs, continuous learning, escalation |
| EU GMP Annex 22: Artificial Intelligence9 | Consultation draft, July 2025 | Draft excludes generative AI from critical GMP use; scope under review10 |
| EU AI Act11 | Phased application; high-risk rules from December 2027 and August 2028 | Transparency rules apply from August 2026 |
| SDAIA AI Ethics Principles and GenAI Guidelines for Government12,13 | In force, Saudi Arabia | Human oversight; accountability; no GenAI in critical government decisions |
| DoH Abu Dhabi AI Policy and Responsible AI Standard14,15 | Policy 2018; Standard October 2025 | Risk tiers, human oversight, audit trails and documentation |
| ISO/IEC 4200116 and NIST AI RMF17 | Published 2023 | Management system and risk framework: govern, map, measure, manage |
| GCC implication Companies operating in the Gulf should expect regulators to apply national AI ethics and health-sector standards alongside FDA and EMA expectations. A single governance model aligned with all of them is more efficient than separate controls per market. |
3. Core Concepts
Context of use
FDA defines context of use as the specific role and scope of the AI model used to address a question of interest.3 The joint principles also require a well-defined context of use for every AI technology.1 In practice, this means one written statement per use case: what question the AI addresses, what it produces, and how the output will be used.
Model risk
FDA's draft guidance assesses model risk using two factors: model influence, meaning how much the output drives the decision, and decision consequence, meaning the impact of an incorrect decision.4 The higher the risk, the more validation, oversight and documentation are needed.3
Credibility
Credibility is trust, established through evidence, in an AI model's performance for a particular context of use.3 FDA's seven steps are: define the question, define the context of use, assess model risk, plan the credibility assessment, execute the plan, document results and deviations, and decide whether the model is adequate.3
Human in the loop and on the loop
Abu Dhabi's Responsible AI Standard requires human-in-the-loop or human-on-the-loop mechanisms, clear escalation protocols by risk level, and override mechanisms.15 Human in the loop means a person approves each output before use; human on the loop means a person supervises and can intervene.
4. The HERMS Human-Accountable Intelligence Framework
The framework below is a HERMS synthesis of the sources in Section 2. It organises AI governance in regulatory affairs into six pillars.
| Pillar | What it requires | Anchored in |
| 1. Purpose | A written context of use and question of interest for every AI application | FDA draft guidance3; joint principle 41 |
| 2. Proportion | Risk tier based on model influence and decision consequence; controls scale with tier | FDA draft guidance4; DoH risk classes15; joint principle 21 |
| 3. Provenance | Every output linked to identifiable, current sources; data lineage documented | Joint principle 61; DoH documentation rules15 |
| 4. People | Named owner for each AI system and each decision; defined decision rights; trained users | SDAIA accountability12; DoH roles15; joint principle 51 |
| 5. Proof | Logs of inputs, outputs, versions, reviews and overrides; audit-ready records | DoH traceability15; ISO/IEC 4200116 |
| 6. Performance | Pre-use validation, monitoring, drift checks and periodic re-evaluation | Joint principles 8 and 91; NIST AI RMF17 |
| The accountability test For any AI-supported regulatory output, the organisation should be able to answer within minutes: what was the AI asked, what sources did it use, who reviewed it, what did they change, and who approved it. |
5. Risk Tiers for Regulatory Use Cases
The tiers below are a HERMS planning model. They apply the FDA logic of model influence and decision consequence4 and the Abu Dhabi principle that controls rise with risk.15
| Tier | Examples | Minimum controls |
| Tier 1: Low | Internal search, meeting notes, first-pass translation for internal reading | Approved tool list; no confidential data in unapproved tools; user responsible for checking |
| Tier 2: Moderate | Regulatory intelligence summaries, first drafts of non-critical sections, document completeness checks | Source citations required; qualified reviewer approves before use; sample audits |
| Tier 3: High | Content supporting safety, efficacy or quality claims; labelling text; responses to authority questions; PV case triage | Validated workflow; human approval of every output; full logging; documented credibility evidence; periodic re-evaluation |
| Tier 4: Not permitted | AI as sole decision-maker on submission, labelling approval, safety causality or batch-related decisions | Prohibited; AI may support but never replace the accountable person |
Tier 4 reflects the wider regulatory position. Saudi Arabia's government GenAI guidelines bar generative AI from critical decisions affecting individuals and require human review of results.13 The draft EU GMP Annex 22 states that generative AI and large language models should not be used in critical GMP applications.9
6. Human Decision Rights
Governance fails when it is unclear who decides. Each AI-supported process should define four roles.
| Role | Responsibility |
| System owner | Accountable for the AI system, its context of use, risk tier, validation and monitoring |
| Qualified reviewer | Checks each output against sources; corrects or rejects it; records the review |
| Decision owner | Approves the regulatory outcome, such as a submission, label or response, and signs it |
| Governance body | Approves new use cases and tier changes; reviews incidents, overrides and performance |
- Named people: SDAIA requires designers, developers and implementers to be identifiable and to assume responsibility.12
- Defined monitoring duties: Abu Dhabi requires named individuals or teams to monitor outputs, review flagged cases and overrides, and report concerns.15
- Final authority stays human: the Abu Dhabi standard states that clinicians should retain ultimate decision-making authority in AI-assisted healthcare scenarios.15 The same logic applies to regulatory sign-off.
- Company responsibility: EMA places responsibility for the fitness of models and data on the applicant or marketing authorisation holder.7
7. Evidence, Traceability and Audit
A regulator or auditor may ask how an AI-supported conclusion was reached. The record should answer that question without relying on memory.
What to record
- Sources: documents, versions and dates used to generate the output.1
- Inputs and outputs: prompts or queries, outputs and the model or system version.15
- Human actions: reviews, edits, overrides and approvals, each linked to a responsible person.15
- Validation: test results, acceptance criteria and monitoring results.15
- Data lineage: data provenance, processing steps and analytical decisions, documented in a traceable way in line with GxP.1
Retention and access
The Abu Dhabi standard requires version-controlled documentation available for internal and external audits, and logging of all access to AI systems.15 Regulatory teams should align AI records with existing document retention and quality-system rules.
| Source-grounded by design A regulatory AI output without a citation is an opinion. A cited output is evidence that a reviewer can verify. Governed systems should make citations mandatory for Tier 2 and Tier 3 use. |
8. Generative AI and Large Language Models
Generative AI is the fastest-growing category in regulatory work and the one regulators treat most cautiously.
What regulators have said
- EMA and HMA: four principles for staff: ensure safe input of data, apply critical thinking and cross-check outputs, keep learning, and know whom to consult when concerns arise. Organisations should define governance, permitted use cases, training and risk monitoring.8
- EU GMP Annex 22 draft: generative AI and LLMs should not be used in critical GMP applications; in non-critical use, qualified people must confirm outputs are suitable.9 EMA is reviewing this boundary after industry feedback.10
- SDAIA: government entities must not use generative AI in critical decisions affecting individuals, and all results should be human-reviewed.13
- Upcoming EU guidance: the EU network workplan foresees guidance on AI in clinical development and pharmacovigilance.18
Practical controls
- Use only approved, enterprise-controlled tools for confidential or product data.
- Restrict generation to approved source libraries where possible, with citations.
- Require a qualified reviewer to confirm every statement used externally.
- Prohibit unreviewed AI text in submissions, labels and authority correspondence.
- Test outputs regularly against known answers and record the results.
9. Implementation Roadmap
The roadmap is a HERMS planning model for a regulatory affairs function adopting governed AI.
| Phase | Timing | Actions |
| Foundation | Days 0–90 | Inventory current AI use; write an AI policy for regulatory affairs; approve tools; assign system owners; train users on safe input and review |
| Control | Days 90–180 | Write context-of-use statements; assign risk tiers; set review and logging rules; define reviewer and decision-owner roles |
| Assurance | Days 180–365 | Validate Tier 3 workflows; set acceptance criteria; start monitoring and sample audits; report to the governance body |
| Mature | Year 2 onwards | Align with ISO/IEC 42001; periodic re-evaluation; update for new FDA, EMA and GCC guidance |
Measures of success
- Share of AI-supported outputs with complete source citations.
- Share of Tier 3 outputs with a recorded human review and approval.
- Reviewer correction rate and time saved per task.
- Number of audit findings related to AI use.
10. Pitfalls and Readiness Checklist
Ten pitfalls
- Using AI without a written context of use.3
- Applying the same controls to every use case regardless of risk.3
- Entering confidential dossier or patient data into unapproved tools.8
- Accepting outputs without checking sources.8
- No named owner for the AI system or the decision.12
- No log of who reviewed, edited or approved an output.15
- Using generative AI for critical decisions.9,13
- Validating once and never monitoring again.1
- Training users on the tool but not on its limitations.8
- Treating vendor assurances as a substitute for company responsibility.7
Readiness checklist
| Area | Ready when |
| Policy | AI policy for regulatory affairs approved; permitted and prohibited uses listed |
| Inventory | Every AI use case recorded with owner, context of use and risk tier |
| People | Reviewers and decision owners named and trained |
| Evidence | Citations, logs and approvals captured for Tier 2 and Tier 3 outputs |
| Assurance | Validation, monitoring and re-evaluation schedule in place |
| Alignment | Controls mapped to FDA, EMA, SDAIA and DoH Abu Dhabi expectations |
11. How HERMS Helps
HERMS Consultancy DMCC helps life sciences companies and healthcare organisations adopt AI in regulatory work with clear human accountability.
- Governance design: AI policy, risk tiers, decision rights and governance bodies for regulatory affairs.
- Use-case assessment: context-of-use statements, risk assessment and control design for each application.
- Assurance: validation plans, acceptance criteria, monitoring and audit readiness.
- Source-grounded regulatory intelligence: structured, cited regulatory content for GCC markets with human review.
- Training: safe use, critical review and accountability for regulatory teams.
| Request a governed AI readiness review HERMS will map your current AI use against the framework in this paper and deliver a prioritised control plan. info@hermsglobal.com | +971 55 600 8944 | www.hermsglobal.com |
References
1. FDA and EMA. Guiding Principles of Good AI Practice in Drug Development, January 2026. https://www.fda.gov/media/189581/download
2. EMA. EMA and FDA set common principles for AI in medicine development, 14 January 2026. https://www.ema.europa.eu/en/news/ema-fda-set-common-principles-ai-medicine-development-0
3. FDA. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (draft guidance), January 2025. https://www.fda.gov/regulatory-information/search-fda-guidan…
4. DLA Piper. Key takeaways from FDA's draft guidance on use of AI in drug and biological life cycle, January 2025. https://www.dlapiper.com/en-at/insights/publications/2025/01/fda-releases-draft-guidance-on-use-of-ai
5. FDA CDER. Artificial Intelligence for Drug Development. https://www.fda.gov/about-fda/center-drug-evaluation-and-research-cder/artificial-intelligence-drug-development
6. EMA. Reflection paper on the use of Artificial Intelligence in the medicinal product lifecycle (EMA/CHMP/CVMP/83833/2023), September 2024. https://www.ema.europa.eu/en/scientific-guidelines/use-artificial-intelligence-ai-medicinal-product-lifecycle
7. EMA. Delivering on stakeholders' requirements for AI (presentation). https://www.ema.europa.eu/en/documents/presentation/presentation-delivering-stakeholders-requirements-ai-l-pinheiro-ema_en.pdf
8. EMA. Harnessing AI in medicines regulation: use of large language models, September 2024. https://www.ema.europa.eu/en/news/harnessing-ai-medicines-regulation-use-large-language-models-llms
9. European Commission. EudraLex Volume 4, Annex 22: Artificial Intelligence (consultation draft), July 2025. https://health.ec.europa.eu/document/download/5f38a92d-bb8e-…
10. EMA. GMP multistakeholder workshop on AI guidance development (Annex 22), 2026. https://www.ema.europa.eu/en/events/good-manufacturing-pract…
11. European Commission AI Act Service Desk. Timeline for the implementation of the EU AI Act. https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act
12. SDAIA. AI Ethics Principles. https://sdaia.gov.sa/en/SDAIA/about/Documents/ai-principles.pdf
13. SDAIA. Generative Artificial Intelligence Guidelines for Government. https://sdaia.gov.sa/en/SDAIA/about/Files/GenAIGuidelinesForGovernmentENCompressed.pdf
14. Department of Health Abu Dhabi. DoH becomes region's first entity to develop an AI policy in the healthcare sector, May 2018. https://www.doh.gov.ae/en/news/the-department-of-health-becomes-region
15. Department of Health Abu Dhabi. Responsible Artificial Intelligence (AI) Standard, DoH/ST/DDGO/RAI/V1/2025, October 2025. https://www.doh.gov.ae/-/media/B84FF5DD027846699C51F36C439A580F.ashx
16. ISO. ISO/IEC 42001:2023 Artificial intelligence: Management system. https://www.iso.org/standard/42001
17. NIST. Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
18. Osborne Clarke. EU pharma regulators publish annual AI observatory report, 2026. https://www.osborneclarke.com/insights/eu-pharma-regulators-publish-annual-ai-observatory-report
Disclaimer: This white paper is for information only and does not constitute legal, regulatory or technical advice. It reflects public sources available on 5 November 2026. Several sources are drafts or voluntary principles that may change. The framework, risk tiers and roadmap are HERMS planning models, not regulatory requirements. © 2026 HERMS Consultancy DMCC.
