AI Governance · ISO/IEC 23894 · Pharma R&D

AI Governance Without the Theater:
Operationalizing ISO/IEC 23894 in Pharmaceutical R&D

Why the most useful AI-governance standard for pharma is the one that plugs directly into quality risk management, FAIR infrastructure, and ontology curation operations.

May 2026 ~28 min read Ali Shahmohammadi, Ph.D. 31 references
Read Article ISO/IEC 23894
26
Pages in ISO/IEC 23894 guidance
2005
Year ICH Q9 process language entered pharma operations
2023
Release year for both ISO/IEC 23894 and ISO/IEC 42001
2028
EMA network workplan horizon for data and AI in medicines regulation
Table of Contents
  1. 01The Theater Problem
  2. 02What ISO/IEC 23894 Actually Is
  3. 03The Bridge Pharma Already Has: ICH Q9(R1)
  4. 04The ISO Ecosystem That Makes 23894 Operational
  5. 05Mapping to NIST AI RMF and the EU AI Act
  6. 06A Pharma R&D Annex III Reality Check
  7. 07The Regulator-Facing Half: FDA Jan 2025 and EMA NDSG
  8. 08The GxP Edge: PIC/S Annex 22
  9. 09Wiring 23894 into FAIR Data
  10. 10Six Operating Decisions to Get Right
  11. 11What This Looks Like in 24 Months
01 — The Problem

The Theater Problem

In pharma AI governance, visible artifacts often outpace operational controls.

Many organizations can show governance theater: committees, principles posters, and readiness decks. Fewer can show the operational chain from risk identification to risk treatment to monitoring evidence.

The central argument of this article is that ISO/IEC 23894 provides the best practical backbone for pharma AI risk because it extends a process language the quality function already operates. The work is not framework selection theater. The work is integration into real systems.

Operational credibility in regulated AI comes from repeatable risk process, not from principle statements alone.
02 — Standard Scope

What ISO/IEC 23894 Actually Is

A guidance standard that extends ISO 31000 risk management for AI-specific characteristics.

Principles

Clause 4 alignment

Integration, structure, adaptation, inclusion, and continual improvement remain intact, but are interpreted for probabilistic systems.

Framework

Clause 5 operationalization

Leadership, integration, and evaluation become AI-specific through controls around data lineage, drift handling, and oversight roles.

Process

Clause 6 execution

Communication, context, assessment, treatment, monitoring, and reporting are extended for model lifecycle, explainability, and bias considerations.

Key point: 23894 does not replace ISO 31000 logic. It adds AI-specific depth where organizations typically fail in implementation.

03 — Pharma Bridge

The Bridge Pharma Already Has: ICH Q9(R1)

Q9(R1) and ISO 31000 are process-isomorphic enough that 23894 can be adopted as an extension, not a replacement.

Pharma quality teams already run risk structures with hazard identification, analysis, evaluation, control, communication, and review. ISO/IEC 23894 can be anchored into that same machinery rather than creating parallel governance tracks.

In practice, this means AI risk owners, escalation paths, and review cadences should sit in the same enterprise risk workflow already used for GxP-adjacent quality risk.

04 — ISO Stack

The ISO Ecosystem That Makes 23894 Operational

23894 becomes practical when paired with terminology, lifecycle, and auditable management-system standards.

Vocabulary + Architecture

ISO/IEC 22989 and 23053

Shared AI terminology and system framing prevent ambiguity in risk discussions and control scope.

Lifecycle

ISO/IEC 5338

Defines where lifecycle-stage risk controls should attach from planning through operation and retirement.

Risk Guidance

ISO/IEC 23894

Provides the risk-management backbone for AI-specific concerns such as drift, uncertainty, and human oversight.

Auditable Wrapper

ISO/IEC 42001

Turns risk and governance intent into management-system requirements organizations can audit and certify.

05 — Crosswalk Logic

Mapping to NIST AI RMF and the EU AI Act, Without Picking a Winner

These frameworks are complementary when used as communication and compliance layers over a single risk process.

NIST AI RMF provides organizing functions. ISO/IEC 23894 provides process depth. ISO/IEC 42001 provides auditable management-system structure. EU AI Act requirements inform treatment and oversight design where legally in scope.

The scalable pattern is one backbone, many projections: 23894 for process execution, 42001 for governance wrapper, NIST RMF for US-facing communication, and EU AI Act obligations wired into controls for applicable systems.

06 — Scope Check

A Pharma R&D Annex III Reality Check

Most pharma R&D AI systems are not automatically Annex III high-risk systems under the EU AI Act.

Assuming all pharma R&D AI is Annex III high-risk often leads to over-broad controls in low-impact contexts and under-preparation where domain-specific controls actually matter.

Even where Annex III does not automatically apply, using Article 14-style oversight patterns in design is still a robust operational choice for regulated environments.

Practical posture: treat high-risk oversight as a design default for critical R&D contexts, while keeping legal classification precise.

07 — Submission Evidence

The Regulator-Facing Half: FDA Jan 2025 and EMA NDSG

The direction of travel is consistent: regulators want structured, context-specific AI credibility evidence.

FDA

Context-of-use credibility model

The January 2025 draft guidance describes risk-based AI credibility assessments by context of use across drug development activities.

EMA/HMA

Data and AI workplan through 2028

The NDSG workplan emphasizes AI literacy and structured capability development across the medicines regulatory network.

08 — GxP Signal

The GxP Edge: PIC/S Annex 22 and What It Forecloses

Draft language points toward strict controls in GMP-critical contexts, reinforcing the need for pre-wired risk evidence.

The consultation direction indicates strong constraints around model behavior and validation expectations in critical applications affecting patient safety, product quality, or data integrity.

Organizations that already run AI systems through a 23894-style risk process are structurally better prepared for any final GMP-facing AI requirements.

09 — FAIR Integration

Wiring 23894 into FAIR Data and Ontology Curation

The strongest implementation pattern is to land risk controls in the same FAIR and ontology systems already supporting R&D data products.

Scope/Context

FAIR plans and maturity indicators

Use FAIR DMP and maturity instrumentation to ground context and acceptance criteria for AI risk decisions.

Risk Identification

FAIRsharing standards registry

Detect divergence from binding standards early by anchoring asset-level risk checks to standards metadata.

Risk Treatment

OBO Foundry and ontology QA

Use established ontology governance principles and pitfall scanning to treat semantic drift and control quality risk.

Outcome: when FAIR, ontology, and provenance are first-class, ISO/IEC 23894 reporting becomes queryable evidence, not manual evidence assembly.

10 — Operating Model

Six Decisions Worth Getting Right

Successful rollouts treat 23894 as operating design, not policy decoration.

1. Anchor to existing Q9 register

Avoid parallel AI-only risk tracks. Extend the existing enterprise quality risk mechanism.

2. Start with one context of use

Run one complete evidence pack end-to-end before scaling portfolio-wide.

3. Use 42001 as wrapper

Adopt management-system structure early, even before certification milestones.

4. Use ATLAS for threat cataloging

Ground AI threat identification in a maintained adversarial taxonomy.

5. Treat soft-law as design input

Use OECD and WHO guidance to harden risk-acceptance criteria in health-adjacent contexts.

6. Fund capability development

Competence and role design determine whether governance remains operational after year one.

11 — Next 24 Months

What This Looks Like in 24 Months

The standards, regulator direction, and FAIR substrate are converging into one implementable governance architecture.

The practical endpoint is a coherent stack where AI risk evidence is generated continuously as systems run: context-of-use aligned, lifecycle-aware, ontology-stable, FAIR-backed, and governance-auditable.

Teams that build this integration now will treat inspections and submissions as projection and query problems. Teams that do not will keep rebuilding evidence manually per project.

Closing thought: stop performing AI governance. Operationalize it where the data and models actually live.

Back to Portfolio Related: Agentic Governance Related: Semantic Layers
References

31 References

Sources include ISO/IEC standards, ICH/FDA/EMA guidance, FAIR and ontology governance literature, and security/risk frameworks.

  1. 1ISO/IEC 23894:2023. iso.org
  2. 2ISO 31000:2018. iso.org
  3. 3ICH Q9(R1) resources via EMA and ICH. ema.europa.eu
  4. 4FDA Q9(R1) overview presentation. fda.gov
  5. 5ISO/IEC 22989:2022. iso.org
  6. 6ISO/IEC 23053:2022. iso.org
  7. 7ISO/IEC 5338:2023. iso.org
  8. 8ISO/IEC 42001:2023. iso.org
  9. 9ISO/IEC TR 24028:2020. iso.org
  10. 10ISO/IEC TR 24027:2021. iso.org
  11. 11ISO/IEC TS 25058:2024. iso.org
  12. 12ISO/IEC 24029-2:2023. iso.org
  13. 13NIST AI RMF resource center. nist.gov
  14. 14NIST AI RMF crosswalk documents. airc.nist.gov
  15. 15EU AI Act (Regulation EU 2024/1689). eur-lex.europa.eu
  16. 16AI Act Annex III explainer. ai-act-service-desk.ec.europa.eu
  17. 17FDA draft guidance on AI in regulatory decision-making. fda.gov
  18. 18HMA/EMA data and AI workplan to 2028. hma.eu
  19. 19PIC/S Annex 22 consultation draft. picscheme.org
  20. 20FAIR Cookbook. faircookbook.elixir-europe.org
  21. 21FAIR maturity evaluation framework (Scientific Data, 2019). nature.com
  22. 22FAIRsharing registry and publication. fairsharing.org
  23. 23OBO Foundry principles and operationalization. obofoundry.org
  24. 24OOPS! ontology pitfall scanner. oops.linkeddata.es
  25. 25MITRE ATLAS. atlas.mitre.org
  26. 26NIST AI 100-2 E2023 adversarial ML taxonomy. csrc.nist.gov
  27. 27OECD AI Principles. oecd.ai
  28. 28WHO AI ethics and governance for health (2021). who.int
  29. 29WHO guidance on large multi-modal models in health (2024). who.int
  30. 30Common Regulatory Capacity for AI, Alan Turing Institute. turing.ac.uk
  31. 31NeOn methodology for ontology engineering. springer.com