Artificial Intelligence and the Validated State

This article examines the regulatory landscape governing artificial intelligence (AI) in quality-critical systems and distinguishes between algorithms, AI, and a nuanced middle category called smart systems. It explains how organizations can responsibly embrace advanced programmatic solutions while preserving the integrity of validated test methods and safeguarding patient safety.

All inspection technologies have benefited from the advancement of sensory measurement and the associated algorithms to analyze test results. Over the past decade, industrial computing capability has advanced significantly. Today, it is necessary to address how new computing capabilities, including AI, may affect critical inspection technology. A new category of technology is emerging as a compliant pathway for deploying advanced computational methods and AI enabled systems without compromising the validated state. This article will explain how test and measurement solutions in a highly regulated environment can remain compliant with industry requirements while also advancing test and measurement into new realms.

Regulatory Compliance

AI is transforming how organizations approach quality and laboratory management. When integrated thoughtfully and responsibly, AI can support compliance with international standards such as EU GMP Annex 22 “Artificial Intelligence”; International Council for Harmonization of Technical Requirements for Pharmaceuticals for Human Use (ICH) Q2(R2), Q14, and Q9(R1); and ISPE GAMP® 5 Guide: A Risk-Based Approach to Compliant GxP Computerized Systems, Second Edition, and ISPE GAMP® Guide: Artificial Intelligence1^, 2^, 3^, 4^, 5^, 7. Although Annex 22 and ISPE GAMP® Guide: Artificial Intelligence offer specific guidance on AI, the ICH guidance documents offer a fundamental view of validation and risk management that applies broad guidance that invariably must be considered for any system.

Although AI can streamline tasks and aid in analysis, its application raises challenges in transparency, explainability, traceability, and risk management. Generative AI and large language models (LLMs), for example, are not suitable for active critical GMP decision processing. By contrast, static, deterministic approaches are permitted when appropriately validated.

Qualification and Verification: Industry Requirements

Validation ensures that instrumentation operates consistently as intended, producing reliable, repeatable results in compliance with regulatory standards. It protects against drift and performance deviations, which is vital for inspection systems in regulated environments. AI enabled systems introduce the concept and risk that the actual functional code is created that needs to be validated. As systems and algorithms evolve, defining validation boundaries becomes increasingly complex, reinforcing the need for a structured and phased approach as outlined in GAMP 54 (see Figure 1).

Instrument qualification is the foundational first step, providing documented evidence that equipment is correctly installed, functions as expected, and performs within specifications. Completion of installation qualification (IQ), operational qualification (OQ), and performance qualification (PQ) establishes that the system is fit for its intended use and suitable to support subsequent analytical activities.

The core instrument qualification steps include:

• IQ: Verify equipment installation and supporting software environment
• OQ: Confirm instrument functionality, secure data handling, and measurement accuracy
• PQ: Demonstrate reliable operation under representative conditions

Figure 1: Establishing the validated state.

Following successful instrument qualification, test method development and test method validation activities are performed. Test method development establishes a robust, fit-for-purpose approach by defining critical method parameters, optimizing operating conditions, and understanding method limitations and sources of variability. This phase ensures the method is suitable for its intended application prior to formal validation and can be appropriately tailored to specific product package configurations, accounting for variables such as material properties, geometry, product formulation, and sealing characteristics.

Test method validation then demonstrates that the developed analytical method consistently delivers accurate, dependable results under defined conditions. This is especially important when introducing new methods, applying them to specific product package configurations, or transferring them across laboratories.

Once both qualification and method validation are established, ongoing life cycle management ensures sustained system performance4. This includes routine monitoring, change control, periodic review, and maintenance of audit-ready documentation to ensure continued compliance throughout the system’s operational life.

Once qualified, a system should remain in its validated state unless formally altered through change control and requalification. Validated methods on the qualified system similarly depend on preserving this state to ensure reproducibility, with clearly defined acceptance criteria. For example, in container closure integrity testing (CCIT), this includes using quantitative data to reliably distinguish intact from defective primary packaging. The draft Annex 22 requires that qualification data remain fully independent from any training data used to develop an algorithm, a principle that is expected to become binding EU GMP requirement upon final publication7. As with any validation, the person or entity preparing training data should be separated from the entities or people who execute qualification. For an advanced programmatic system, that means there must be a distinct and clear firewall between the data used to develop a test method from the data that is generated for the qualification testing.

Any algorithm that operates outside the validated state but influences it presents a significant compliance and patient safety risk. Such influence undermines documented evidence, compromises data integrity, and may necessitate full or partial revalidation. Maintaining strict boundaries around the validated state and rigorously assessing the effect of changes are therefore essential to safeguarding regulatory trust and ensuring reliable test outcomes. Validation processes are generally established internally by manufacturing quality control departments. These departments retain responsibility for defining, justifying, and documenting their individual validation requirements, particularly regarding change management procedures for qualified systems.

Explainability and clarity of AI algorithms are critical steps preceding validation. Annex 22 requires that algorithms used in GMP-critical systems must be deterministic, transparent in their operation, and accompanied by evidence explaining how outputs are derived7. Models must provide confidence scores and thresholds that can be linked to deterministic acceptance criteria or human review. GAMP reinforces this by emphasizing that AI systems should include interpretable documentation such as models and should feature importance or explainability reports that allow quality, regulatory, and technical staff to understand why a decision was made5.

Together, these expectations ensure that algorithms are not “black boxes” influencing validated states. Instead, they remain within the boundaries of documented, reviewable, and explainable logic. In this way, they preserve the integrity of validation and enable accountable change management when models are updated or replaced. Industry standards can help navigate complex requirements when introducing and deploying new technologies. Comprehensive documentation, particularly when integrating advanced technologies such as AI, is mandatory to ensure thorough compliance amid an evolving technological landscape.

Algorithms, Smart Systems, and AI

Algorithms and models are pervasive in test and measurement. Signal processing algorithms are often included inside detection sensors, sometimes hidden from the trained validation eye. Any regression model, smoothing technique, or other signal-filtering method that may affect a measurement system must be explained and justified, and it must provide some level of reliable and repeatable performance. For some advanced methods, a system may require a reference dataset or inputs to the test analysis. In these cases, this reference data must be defined before beginning a new test method validation (TMV).

A validation cannot begin with varying test inputs other than the test conditions. Establishing reference datasets may take place as part of a recipe-development process, or this may simply be a predefined set of limits that the system uses to normalize a measurement. In any of these cases, the algorithm’s model and the test inputs beyond the test condition must be predefined and may not change once the system enters validation.

Although AI can be considered a broad spectrum of computational approaches, this article will use two common terms to categorize advanced computational systems, and a third nuanced category that champions the space in between.

Algorithms

Algorithms are a finite sequence of explicitly defined steps that, when executed, convert specified input into the intended output. They are persistent and deterministic6. Their processes do not change, only the data the processes are applied to. Most test and measurement systems employ algorithms, which can range from basic to highly complex.

AI

AI has multiple definitions and interpretations. For this article and measurement technologies, AI refers to neural networks and machine learning approaches in which the algorithm or model being applied to the analysis can evolve or adapt.

Smart Systems

There is a nuanced position in which a system can lean on AI for certain critical functions, then lock that dynamic system so it deploys as a static system. In this article, these will be referred to as a smart system, which can be static or adaptive, but never both, based on the system’s status. It is a category of technology that deploys intelligent, machine-guided analysis tools to reach a validated state and ensure full explainability and certainty after validation6 (see Figure 2).

Pre-validation role

Smart systems can use dynamic inputs, complex data analysis, and even limited AI or LLM support to guide method or process development. They may compile multiple datasets, perform meta-analyses, and suggest optimized parameters. Machine guidance accelerates the path to a robust, reliable method.

Transition to validation

Once a method is selected, the smart system fixes its algorithmic pathway, locking all parameters and clarifying how results are generated6. The system no longer adapts dynamically; the method is now deterministic, reviewable, and subject to qualification and requalification requirements.

Distinction from AI

Unlike adaptive AI models, smart systems do not continue learning or evolving during operation. They apply intelligence upfront, but once validated, they function as transparent, reproducible algorithms. Any further changes require formal change control and revalidation.

Regulatory alignment

In this way, smart systems meet the explainability and determinism requirements of Annex 227, while following the risk-based life-cycle approach in the ISPE GAMP® Guide: Artificial Intelligence guidance. They provide the benefits of advanced computational analysis without compromising the clarity and stability required in validated GMP systems. A method that is validated has established operating parameters, a digital twin, that allows for reestablishing the validated state under any circumstance.

Figure 2: Smart systems.

AI and the Validated State

When considering AI in regulated environments, there is a bold line it cannot cross: the validated state. Validation depends on traceability, scientific clarity, and stasis, qualities incompatible with adaptive, continuously evolving models. True AI that allows algorithms to evolve introduces uncertainty that cannot be reconciled within a validated system. AI algorithms can be complex and capable of learning, but any influence they exert cannot contaminate the validated state. Within qualified systems, only approaches that are deterministic, repeatable, and explainable can be permitted.

This distinction is often described in terms of black box vs. open-box architectures.

A black-box algorithm hides its internal decision-making, making it impossible to fully explain or validate how results are produced. Inputs and outputs may be visible, but the path between them remains opaque. Such algorithms cannot assure repeatability or reproducibility6.

An open-box algorithm, by contrast, offers transparency. Its structure and logic are accessible, auditable, and explainable step by step6. This clarity certifies that results are consistent and reproducible, making it acceptable in a validated state.

To meet regulatory expectations, all algorithms within validated equipment must be fully documented, explainable, and deterministic. All quality-critical test methods must rely on open-box algorithms.

For smart systems, quality risk management (QRM) should go beyond simply enforcing a firewall between AI and the validated state. It should systematically evaluate risks in how a smart system generates, locks, and operates test methods. By identifying hazards, assessing their effect, and defining appropriate controls (such as segregated data, explainability evidence, and operator review), QRM ensures smart systems can be deployed confidently6 while still preserving the determinism and transparency required in validated GMP systems.

Example of Dynamic Leak Detection

Traditionally, developing a CCIT method required a comprehensive design of experiments (DoE) to define input conditions, run positive and negative controls, and establish reference parameters. Method qualification then relied on demonstrating a clear signal-to-noise ratio with reproducibility across test samples. Although effective, this process demanded heavy operator involvement and produced methods limited by the variance in conventional vacuum decay technology. Newer technologies that apply the smart system approach dynamically process significantly larger volume of raw data during the initial recipe creation. Instead of relying solely on operator-led DoE, a smart system automatically explores input ranges, identifies optimal test conditions, and establishes robust pass/fail thresholds. The outcome can be a test method that requires far less manual intervention and greater test capability.

Once the smart system certifies the method:

• The algorithm must be locked, with all variables fixed.
• The mathematical pathway to the outcome must be transparent and explainable.
• Documentation captures the rationale, data, and logic leading to the method.
• The method’s digital twin can be deployed to re-create the validated solution.

The outcome is an optimized and validated test system. Operators benefit from simplified workflows; manufacturers gain more robust and reproducible data. Importantly, once validated, the method retains the power of the advanced algorithm but now functions as a fully deterministic open-box solution that satisfies regulatory and global operational expectations.

Deploying AI Around the Validated State

Higher-level AI, including neural networks, makes it hard to retrace how outcomes are generated. Although these models can adapt to changing inputs and produce favorable results, they do so without transparent reasoning. In GMP contexts, in which data traceability and explainability are essential, AI cannot operate within a validated system’s measurement or analysis processes.

AI may be deployed around the validated state, for example, to support method development or user interfaces, provided its outputs are clearly explainable, repeatable, and converted into fixed, deterministic logic before validation. Without this certainty, it is impossible to demonstrate that a method or system performs to required expectations.

Conclusion

The field of test and measurement can confidently embrace innovation with proper controls. Deterministic systems represent a significant improvement over subjective approaches, and the smart systems framework described in this article offers a structured path to deploy them successfully: AI and advanced algorithms drive method development and optimization, but all parameters are locked before validation begins, producing a fully deterministic, open-box solution that meets regulatory expectations. The objective is not to choose between innovation and compliance, but to apply the right framework in which one does not inhibit the other. This approach is critical to the advancement of the life sciences industry, improving performance for manufacturers and safety for patients. The industry must apply the right level of control to ensure the validated state and patient safety remain protected.