德勤解读EU-GMP附录22草案,认为AI验证需由系统级转向模型级
德勤解读EU-GMP附录22-AI验证从系统级转向模型级
德勤爱尔兰发布对欧盟GMP附录22草案的解读,认为AI赋能系统的验证需从系统级扩展到模型、数据与性能监测。该附录于2025年7月7日发布征求意见、预计2026年底定稿,解读将其要求归纳为人员、文件与质量风险管理三项原则。文中还提示草案仅覆盖静态模型,动态模型在GMP环境被视为重大合规风险,监管机构正在评估限制措施。
题图 无菌灌装隔离器内部的机器人操作区与手套口,GMP 生产现场的一类典型 AI 落点(图源:steriline.it)
制药 AI 现场观察 | 主体材料:Deloitte Ireland(2026-11-10)
来源:Deloitte Ireland(德勤爱尔兰),审计与鉴证业务视角的监管解读
URL:https://www.deloitte.com/ie/en/services/audit-assurance/perspectives/annex-22-ai-governance-gmp-assurance.html
发布日期:2026-11-10
体裁:咨询机构解读,英文原文;本页为英文原文与中文译文逐段对照
监管背景:EudraLex 第 4 卷附录 22 于 2025 年 7 月 7 日发布征求意见,预计 2026 年底定稿
As Artificial Intelligence (AI) evolves and becomes increasingly prevalent in the highly regulated life sciences industry, the regulatory landscape is also evolving to address the unique challenges AI presents. Eudralex Volume 4 Annex 22 was released for stakeholder consultation on July 7, 2025 and is expected to be finalised by the end of 2026. The draft Annex 22 which focuses on AI systems is a significant step in setting clear compliance expectations for companies utilising AI in their validation efforts for GxP systems.
随着人工智能(AI)不断发展并在高度受监管的生命科学行业中日益普及,监管格局也在演进,以应对 AI 带来的独特挑战。EudraLex 第 4 卷附录 22 于 2025 年 7 月 7 日发布供利益相关方征求意见,预计于 2026 年底定稿。聚焦 AI 系统的附录 22 草案,是为在 GxP 系统验证工作中使用 AI 的企业设定明确合规预期的重要一步。
But this isn't about adopting new technology, it is about fundamentally rethinking how you validate, govern and oversee it – a shift that separates compliant organisations from industry leaders. Those organizations that treat AI as a tool will fail inspections; those that treat it as a system requiring governance will lead the industry.
但这不是采用新技术的问题,而是从根本上重新思考如何验证、治理和监督它。这一转变把“合规组织”与“行业领导者”区分开来。把 AI 当工具的组织过不了检查;把 AI 当作需要治理的系统的组织将引领行业。
Annex 22 is the first GMP guidance dedicated to AI and Machine learning in regulated pharmaceutical environments. It represents a watershed moment, regulators are not rejecting AI but rather establishing clear boundaries and expectations for its safe, accountable and transparent use. For organisations that embrace this shift, Annex 22 represents an opportunity to lead the industry. It is a signal that the industry's future belongs to those who can demonstrate responsible AI governance.
附录 22 是首份专门针对受监管制药环境中 AI 与机器学习的 GMP 指南。这是一个分水岭时刻:监管机构不是在拒绝 AI,而是在为其安全、可问责、透明的使用划定期望与边界。对拥抱这一转变的组织而言,附录 22 是引领行业的机会。这是一个信号:行业的未来属于那些能证明自己进行了负责任 AI 治理的组织。
This article explores three critical dimensions: why over reliance on AI fails; what Annex 22 expects from organisations; and how to integrate AI governance into your validation and SDLC frameworks.
本文探讨三个关键维度:为什么过度依赖 AI 会失败;附录 22 对组织的期望是什么;以及如何把 AI 治理融入验证与 SDLC(软件开发生命周期)框架。
一、风险:为什么过度依赖 AI 会失败
The Risk – why over reliance on AI fails
Quality Risk Management for AI and machine learning should be risk-based, proportionate, and applied throughout the system lifecycle. Organizations should define intended use, validate data and model performance, maintain human oversight, and monitor for drift or changes, that safeguards patient safety, product quality, and Data Integrity (DI).
AI 与机器学习的质量风险管理应当基于风险、成比例,并贯穿系统全生命周期。组织应定义预期用途、验证数据与模型性能、保持人类监督,并监测漂移或变化,以保障患者安全、产品质量与数据完整性(DI)。
Gao and Feng et al. found that AI penetration drives substantial efficiency, with each 1% increase in AI adoption translating to a 14.2% improvement in total productivity, but overreliance should be avoided. Use of AI-enabled system does boost efficiency but over reliance needs to be monitored.
Gao 与 Feng 等人的研究发现,AI 渗透带来显著效率提升:AI 采用率每提高 1%,全要素生产率改善 14.2%;但应避免过度依赖。AI 赋能系统确实提升效率,但过度依赖需要被监控。
Overreliance on AI can lead to gaps such as inaccurate or incomplete data review, failure to identify compliance risks, lack of proper validation of AI-generated outputs, and insufficient documentation of decision-making processes. These gaps increase the risk of non-compliance with regulatory requirements and can compromise product quality and patient safety.
过度依赖 AI 会造成缺口:数据审查不准确或不完整、未能识别合规风险、对 AI 生成输出缺乏适当验证、决策过程记录不足。这些缺口抬升违反监管要求的风险,并可能危及产品质量与患者安全。
To address these risks and prevent compliance failures that over-reliance creates, regulators have established a new framework, Annex 22. Rather than rejecting AI, this framework redefines how it must be managed.
为应对这些风险、防止过度依赖造成的合规失败,监管机构建立了新框架,即附录 22。该框架并非拒绝 AI,而是重新定义 AI 必须被管理的方式。
二、期望:对 AI 赋能 GxP 系统的新监管要求
The Expectation - New regulatory expectations for AI-enabled GxP systems
Annex 22 is based on three principles, Personnel, Documentation and Quality Risk Management.
附录22 建立在三项原则之上:人员(Personnel)、文件(Documentation)与质量风险管理(Quality Risk Management)。
· Personnel - Subject matter experts, data scientists, QA and IT Consultants, they need to be qualified to understand the intended use and risks associated with the integration of AI into GMP environment.
· Documentation - process of documenting all the activities of the process. Additionally regulated users should review these documentations irrespective of whether models are developed, tested and validated internally or externally.
· Quality Risk Management - implementation based on impact or risk to Patient safety, product quality and DI throughout the model lifecycle.
· 人员:主题专家、数据科学家、QA 与 IT 顾问,需要具备资质以理解 AI 融入 GMP 环境的预期用途与相关风险。
· 文件:记录流程全部活动的过程。此外,无论模型是内部还是外部开发、测试与验证,受监管用户都应审查这些文件。
· 质量风险管理:基于对患者安全、产品质量与数据完整性的影响或风险,在模型全生命周期内实施。
By embedding these principles into their governing policies and procedures, organisations can move beyond mere compliance and build a culture of confidence and integrity in their data management practices.
把这些原则嵌入治理政策与程序后,组织可以超越单纯的合规,在数据管理实践中建立信心与诚信的文化。
The draft regulation specifies that organisations must ensure their policies comprehensively address both traditional data management and emerging AI/LLM (Large Language Model) technologies.
草案规定,组织必须确保其政策全面覆盖传统数据管理与新兴 AI/LLM(大语言模型)技术。
Note, Annex 22 covers static models only, Dynamic models present significant compliance risks within GMP environment, and Regulatory agencies are evaluating measures to limit their application in this context.
注意:附录 22 只覆盖静态模型。动态模型在 GMP 环境内带来重大合规风险,监管机构正在评估限制其在此语境下应用的措施。
对比项 | 静态模型 | 动态模型 |
|---|---|---|
输出特征 | 确定性输出,同一输入给同一结果 | 随新数据持续更新,输出会变 |
附录 22 的覆盖 | 草案覆盖 | 草案未覆盖 |
GMP 环境中的定位 | 可用于关键 GMP 系统,须人在回路监督 | 被视为重大合规风险,监管机构正在评估限制措施 |
上表由本文依据原文陈述归纳,原文未展开静态模型与动态模型的定义。
Organizations can follow these key considerations for effective AI Governance policies and procedures:
组织可遵循以下关键考量来制定有效的 AI 治理政策与程序:
· Establish AI Governance Ownership: Maintain detailed records of every AI generated output, whether its document creation, data analysis or quality decisions. It's essential to outline who must review/approve it.
· Clear Roles and Responsibilities with Human in the Loop (HITL) Governance: Define accountability for data integrity across all levels, with explicit emphasis on human oversight of AI or automated systems used.
· Validation and Verification: Establish rigorous validation protocols for electronic systems and AI tools to confirm they function as intended and comply with CGMP requirements.
· Training and Awareness: Regularly train staff on updated AI Governance policies, emphasising the importance of vigilance when working with AI-generated outputs.
· Continuous Review and Improvement: Periodically review and update policies to reflect regulatory changes and lessons learned from internal audits or inspections.
· 确立 AI 治理归属:对每一条 AI 生成输出(无论是文档创建、数据分析还是质量决定)保留详细记录,必须明确谁来审查/批准。
· 清晰的角色与职责,人在回路(HITL)治理:在各层级界定数据完整性问责,明确强调对所用 AI 或自动化系统的人类监督。
· 验证与确认:为电子系统与 AI 工具建立严格的验证规程,确认其按预期运行并符合 CGMP 要求。
· 培训与意识:定期就更新后的 AI 治理政策培训员工,强调处理 AI 生成输出时保持警觉的重要性。
· 持续审查与改进:定期审查并更新政策,以反映监管变化与内部审计/检查的经验教训。
The regulatory expectation is clear, Annex 22 allows static, deterministic models in critical GMP systems such as computer networks that directly impact patient safety, product quality, and data integrity, with HITL oversight. This shifts accountability to subject matter expert i.e. HITL.
监管预期很明确:附录 22 允许在直接影响患者安全、产品质量与数据完整性的关键 GMP 系统(如计算机网络)中使用静态、确定性模型,但须有人在回路监督。这把问责转移到主题专家,即 HITL 身上。
This requires policies and procedures that are comprehensive, transparent, and rigorously enforced. Organisations that embrace this expectation will be prepared for inspections, confident in their compliance readiness, and positioned as industry leaders in responsible AI adoption.
这要求政策与程序全面、透明并被严格执行。拥抱这一预期的组织将为检查做好准备、对合规就绪度有信心,并在负责任的 AI 采用上占据行业领先地位。
图 1 灌装工位:西林瓶经星轮定位后送至灌装针头下方(图源:steriline.it)
三、未来:把 AI 融入验证与 SDLC
The Future - Integrating AI into Validation and SDLC
Understanding these governance expectations is essential, but governance alone is insufficient – organisations must also translate these principles into concrete validation practices that embed AI oversight at every stage of the system lifecycle.
理解这些治理预期是必要的,但仅有治理不够:组织还必须把这些原则转化为具体的验证实践,在系统生命周期每个阶段嵌入 AI 监督。
The shift in validation approach from system level to model level, requires a new mindset.
验证方法从系统级向模型级的转变,需要新的思维方式。
· AI is a highly transformative technology that is changing how we deliver software solutions. However, the basic principles of Software Development Life Cycle (SDLC) still need to be followed. The fundamental principle of validation still applies with additional aspects related to AI technology.
· Traditional computer system validation stays at the system level. Whereas, for AI enabled computer systems, the validation also includes the 'sub-system' that is the model, the data and the performance monitoring.
· The shift is from 'Functionality verification' to 'Functionality verification' plus 'AI model governance, data management, data-driven testing'.
· Transparency and data traceability through the life cycle of the system remains paramount.
· Continuously monitor the AI models for anomalous behaviour and performance issues.
· AI 是一项正在改变软件交付方式的高度变革性技术。但软件开发生命周期(SDLC)的基本原则仍需遵循;验证的基本原则依然适用,只是增加了与 AI 技术相关的方面。
· 传统计算机系统验证停留在系统级。而对于 AI 赋能的计算机系统,验证还要覆盖“子系统”,即模型、数据与性能监测。
· 转变是从“功能验证”走向“功能验证 + AI 模型治理、数据管理、数据驱动测试”。
· 系统全生命周期的透明度与数据可追溯性仍然至高无上。
· 持续监测 AI 模型的异常行为与性能问题。
验证范围的两层
| 系统层 传统计算机系统验证
| ||
↓ | ||
AI 赋能系统在系统级之下再加一层子系统 | ||
| 模型
| 数据
| 性能
|
图 2 验证范围的两层:系统级之下的模型、数据与性能监测三个子域(依据原文第三节绘制,非官方发布)
This shift - from system level to model level validation - demands that organizations embed Personnel expertise, rigorous Documentation, and proportionate Quality Risk Management at every stage of the AI lifecycle, ensuring that transparency, accountability, and vigilance become operational realities, not regulatory afterthoughts.
这一从系统级到模型级验证的转变,要求组织在 AI 生命周期每个阶段都嵌入人员专业性、严格文件与成比例的质量风险管理,确保透明度、问责与警觉成为运营现实,而不是监管事后补丁。
四、AI 实施该做与不该做
AI Implementation Do's and Don'ts
Do's / 该做 | Don'ts / 不该做 |
|---|---|
实施持续性能监测与再验证。 | 假设 AI 生成的输出是准确的。 |
建立严格的 HITL 治理。 | 把合规问责委托给 AI 系统。 |
面向监管理解做验证,而不仅是功能。 | 期望 AI 独立理解复杂的监管场景。 |
实施 AI 专属风险管理。 | 在没有 AI 专属风险管理的情况下实施 AI。 |
明确治理归属与问责。 | 把 AI 治理当作可选项或一次性事件。 |
监测模型漂移与性能退化。 | 忽视持续监测与性能指标。 |
记录预期用途、局限与范围边界。 | 让范围边界含糊不清。 |
五、结论
Conclusion
Annex 22 is not a burden; it is the future. Organisations that embed its three core principles – Personnel accountability, rigorous documentation and proportionate Quality Risk Management – will move beyond compliance to build genuine confidence in their AI enabled systems.
附录 22 不是负担,它就是未来。把三项核心原则(人员问责、严格文件与成比例的质量风险管理)嵌入其中的组织,将超越合规,在其 AI 赋能系统上建立真正的信心。
The regulatory expectation is clear – AI can accelerate validation and data integrity efforts but only when certain conditions are met. First, human oversight is non-negotiable – AI generates insights, but humans make compliance decisions. Second, every AI generated output must be traceable, documented and reviewable – transparency is key. Third, validation must shift from system level to the model level, encompassing data quality and continually monitoring for drift.
监管预期很明确:AI 可以加速验证与数据完整性工作,但只有满足特定条件才行。第一,人类监督不可谈判:AI 产生洞察,人类做出合规决定。第二,每条 AI 生成输出必须可追溯、有文件、可复核:透明是关键。第三,验证必须从系统级转向模型级,涵盖数据质量并持续监测漂移。
The FDA inspection underscores a fundamental principle – accountability for compliance cannot be delegated to AI. Ultimate responsibility rests with the organisation.
FDA 检查强调了一条基本原则:合规问责不能委托给 AI。最终责任在组织。
Organisations must move beyond mere compliance. With the EU set to roll out Annex 22 focused on AI, now is the time to ask - Is your organisation truly confident in its AI Governance policies?
组织必须超越单纯的合规。随着欧盟即将推出聚焦 AI 的附录 22,现在该问一句:你的组织真的对自己的 AI 治理政策有信心吗?
资料来源与说明:
Deloitte Ireland,Annex 22: From AI adoption to GMP Assurance,2026-11-10,deloitte.com/ie;本文为英文原文与中文译文逐段对照,译文供学习交流,如与原文有出入以英文原文为准。
附录 22 为 EudraLex 第 4 卷《GMP 人用药品与兽用药品》的新增附录,2025-07-07 发布征求意见稿,预计 2026 年底定稿;本文引用的是德勤对该草案的解读,不是法规原文。
题图与正文图 1 取自 steriline.it 产品页图库;图 2 为依据原文第三节绘制的示意图,非官方发布。
来源:小南AI实验室 · mp.weixin.qq.com