This draft guidance focuses on how artificial intelligence and machine learning technologies are being applied in the development of drugs and biological products. Its primary intent is to set expectations and offer clarity for sponsors who plan to integrate AI tools across the lifecycle of drug development—from early research to regulatory submission and beyond.

Importantly, this document is not applicable to AI-based software used as medical devices, which fall under separate regulatory frameworks. Instead, the guidance speaks directly to AI tools used for:

Designing and conducting clinical trials.Interpreting nonclinical data.Enhancing manufacturing and product quality(Chemistry, Manufacturing, and Controls(CMC)applications).Leveraging real-world data and pharmacovigilance.

The guidance does not impose new legal requirements but serves as a framework to align AI use with existing scientific and regulatory principles. By offering this direction, the FDA encourages early communication and proactive planning with regulators, helping ensure that AI-driven methods are trustworthy, scientifically justified, and aligned with public health goals [4][5].

The FDA’s draft guidance outlines several foundational themes that are essential when integrating AI/ML into drug and biological product development. At its core, the document emphasizes transparency, reliability, and proactive planning as critical pillars of responsible AI adoption [7][8].

2.2.1. Transparency in AI Models

Transparency doesn’t just mean explaining the model’s function—it also includes how the model was trained, the nature and origin of the data, its limitations, and any bias that may be embedded in its performance [9][10]. Regulatory submissions involving AI should clearly describe:

The purpose and scope of the AI tool.The development process, including datasets used for training and validation.Measures taken to ensure data integrity and generalizability.

This level of openness allows regulatory reviewers to evaluate the scientific validity of the tool [4][9] and its appropriateness for its intended use.

2.2.2. Risk-Based Thinking

The guidance advocates for a risk-based approach to AI deployment. Sponsors are expected to assess the impact of AI tools on product quality, patient safety, and decision-making. For instance, if an AI tool influences dosing decisions or trial endpoints, the regulatory bar will be higher compared to tools used for internal resource planning or supply chain modeling [11][12] (Figure 2 and Figure 3).

Risk assessments should consider:

The degree of automation versus human oversight.The potential for AI-related errors to affect clinical outcomes.The strategies in place for monitoring AI performance over time.

Figure 2. Risk Matrix for artificial intelligence models used in regulatory decision-making.

Figure 3. Stepwise framework for risk-based credibility assessment of AI models.

2.2.3. Lifecycle Management and Updates

AI models are not static; they often evolve as new data becomes available. The FDA stresses the importance of managing AI tools over their lifecycle—including plans for periodic retraining, performance re-validation, and documentation of updates. This is especially critical in the context of continuous learning systems, where model drift can pose hidden risks [11][13].

Regulators expect sponsors to outline:

How changes to the model will be controlled and documented.What version control mechanisms exist.How post-deployment monitoring will be conducted, particularly in real-world use scenarios.

2.2.4. Early and Ongoing Engagement

Finally, the FDA encourages early communication between sponsors and regulatory agencies. Engaging with the FDA through pre-submission meetings or INTERACT meetings allows for alignment on expectations, especially when novel AI applications are being proposed [4].

This open dialogue helps avoid surprises late in the development process and fosters a shared understanding of how AI tools will support regulatory decisions.

In the discovery phase, AI algorithms are being used to model protein-ligand interactions, screen virtual compound libraries, and predict off-target effects. Deep learning models, especially graph neural networks and transformer-based architectures, are enabling faster lead identification with improved accuracy [1][3]. Moreover, natural language processing (NLP) tools can mine vast datasets—from biomedical literature to patents—to identify novel therapeutic targets or repurpose existing drugs.

In early development, AI is also leveraged for high-content screening, automating image analysis in cell-based assays, and optimizing synthetic pathways.

An AI model is used to integrate in vitro assay data and in vivo toxicology results to predict a safe starting dose for first-in-human studies. The AI output supports selection of the maximum recommended starting dose (MRSD) included in an IND submission. Because this model directly informs human dosing decisions, its performance, training data, and limitations would be subject to FDA scrutiny under a risk-based framework [4][15].

AI supports clinical trial design in several transformative ways:

Patient recruitment and stratification: Machine learning models can mine electronic health records (EHRs) and genomic data to identify eligible patients and stratify them into appropriate trial arms [8].Synthetic control arms: In rare diseases or high-risk indications, AI-generated synthetic control arms—based on historical data—may reduce the need for placebo groups, although this approach still requires rigorous validation [15].Trial monitoring and adaptive protocols: AI can analyze real-time data to flag protocol deviations or recommend adaptive changes to dosing or inclusion criteria.

The FDA guidance encourages early engagement if AI will influence key trial decisions, especially if the tool is tied to primary endpoints or safety assessments.

An AI-based algorithm is used to stratify patients and recommend adaptive modifications to enrollment criteria based on interim efficacy and safety data during a Phase II clinical trial. The model’s outputs inform protocol amendments and influence interpretation of primary endpoints submitted to the FDA. Given its direct impact on trial conduct and regulatory conclusions, this use case represents a high-risk application requiring robust validation, transparency, and defined human oversight [4][15].

AI/ML tools have growing utility in process development,process control, and manufacturing operations. Examples include:

Multivariate process monitoring: AI can detect subtle shifts in critical process parameters before they result in deviations or out-of-specification batches.Predictive maintenance: ML-based predictive maintenance systems reduce downtime by anticipating equipment failures.Release testing automation: Computer vision and machine learning are also enhancing the speed and objectivity of visual inspections, dissolution profile assessments, and impurity tracking [16]-[18].

An AI model is deployed to evaluate multivariate process data and determine whether a drug product batch meets predefined quality specifications for release. The AI output supports batch disposition decisions documented in regulatory submissions. Because erroneous outputs could directly affect product quality and patient safety, this application carries high regulatory risk and must meet validation and lifecycle management expectations equivalent to conventional release testing methods.

The FDA guidance stresses that if AI tools are used to release or reject product batches, they must meet the same validation and reliability expectations as conventional methods [4][18][19].

AI can assist sponsors in compiling submission-ready documentation, such as electronic Common Technical Document (eCTD) modules, by organizing large volumes of data, identifying inconsistencies, and suggesting summaries. In addition, NLP can help regulators navigate previous precedent, advisory committee minutes, and similar product reviews.

However, the FDA remains cautious about fully automating regulatory decisions. The guidance clarifies that AI tools used to inform regulatory pathways or submissions must be well-documented and not replace human expert judgment [4].

Post-approval, AI systems are being applied to analyze real-world data for pharmacovigilance. By processing EHRs, insurance claims, and social media signals, AI can detect adverse event signals faster and more sensitively than manual methods [6][20][21].

An AI-based signal detection system is used to analyze real-world data and identify potential safety signals that trigger expedited adverse event reporting and labeling updates. The AI output supports regulatory safety reporting obligations and post-marketing risk management decisions. As these outputs may prompt regulatory action, the model’s data sources, performance characteristics, and bias controls are subject to regulatory evaluation.

AI is also key to evaluating real-world effectiveness, as it enables the integration of heterogeneous data sources to assess how therapies perform across broader patient populations [20][21].

One of the most significant hurdles is the “black-box” nature of many AI models—particularly deep learning approaches. While these models may demonstrate high performance, they often lack interpretability. This opacity can hinder scientific understanding, undermine trust, and complicate regulatory evaluations [9][10].

FDA emphasizes that sponsors should ensure their AI models are accompanied by documentation that explains how they were developed, trained, validated, and how they operate. This includes a clear rationale for the model’s architecture, input data, and any preprocessing steps. Tools that inform regulatory decisions must allow for meaningful interpretation and justification of their outputs [4] (Figure 5 and Figure 6).

Figure 5. Regulatory and technical challenges in the use of AI for drug development.

Figure 6. FDA-Led initiatives and cross-sector collaborations supporting the use of artificial intelligence in drug development.

AI models are only as reliable as the data they are trained on. If datasets are incomplete, unbalanced, or poorly curated, the resulting models can perpetuate or even amplify biases. This has implications for patient safety, especially when AI is used for patient selection, dose prediction, or safety signal detection.

The FDA draft guidance calls for transparency regarding data sources, preprocessing, and handling of missing data. Sponsors are expected to assess and mitigate potential sources of bias, and to justify the representativeness of their training and testing data, especially in diverse populations [10][21][22].

Across nonclinical, clinical, manufacturing, and post-market contexts, regulatory risk is driven not by the presence of AI itself, but by the specific decision being supported and its potential impact on patient safety, product quality, or regulatory outcomes [4].

Unlike traditional software, AI models can evolve over time—particularly when adaptive learning or real-time updating is involved. This raises critical questions about version control, reproducibility, and change management.

FDA urges sponsors to follow a total product lifecycle approach for AI tools. This includes rigorous model validation using independent test datasets, stress-testing across use cases, and a post-deployment monitoring plan. If a model will evolve over time (e.g., through continued learning), predefined protocols must be in place to manage such updates without compromising regulatory compliance [4][11][12].

For AI tools that support regulatory-relevant decisions or data included in INDs, NDAs, or BLAs, sponsors should assemble a credibility/validation package addressing the following minimum elements:

4.3.1. Data Provenance and Suitability

Clear description of data sources, collection context, and preprocessing steps.Justification of data relevance to the intended use (population, process, or setting).Assessment of data completeness, representativeness, and potential bias.

4.3.2. Training, Validation, and Testing Strategy

Description of data partitioning (training/validation/test splits).Use of independent test sets; external validation where feasible.Rationale when external validation is not possible.

4.3.3. Performance Metrics and Acceptance Criteria

Pre-specified performance metrics aligned with the regulatory context of use.Quantitative acceptance criteria defined before model evaluation.Comparison to baseline or conventional methods, where applicable.

4.3.4. Failure-Mode and Limitation Analysis

Identification of known failure modes and conditions where performance degrades.Sensitivity analyses for edge cases or out-of-distribution inputs.Risk mitigation strategies and defined role of human oversight when failures occur.

This package should be maintained within the sponsor’s quality system and referenced in regulatory submissions to support transparency, reproducibility, and confidence in AI-supported conclusion.

To operationalize a total product lifecycle approach, sponsors should manage AI models using established change-control principles. AI models used to support regulatory-relevant decisions should be versioned and traceable to training data and intended use. Sponsors should predefine triggers for model updates, including performance drift or changes in data or use conditions, and conduct risk-appropriate revalidation to demonstrate continued fitness for use. Documentation of model changes and performance monitoring should be maintained within the quality system. FDA interaction is generally appropriate when changes affect intended use or regulatory conclusions.

AI tools can support—but not replace—human expertise in regulatory decision-making. Even when models demonstrate high accuracy, final decisions must rest with qualified professionals who understand the AI tool’s context and limitations.

To ensure clear decision accountability in regulated environments, sponsors should define minimum human-oversight controls for AI-assisted activities. Each AI tool used to support regulatory-relevant decisions should have a named accountable role (e.g., study lead, manufacturing quality lead, or safety physician) responsible for approving use, interpreting outputs, and authorizing final decisions.

AI-assisted decisions should be supported by review and audit trails documenting the model version used, input data, outputs generated, and the human rationale for accepting or rejecting those outputs. When AI outputs conflict with expert judgment or predefined expectations, escalation pathways should be defined, including secondary expert review and documentation of resolution prior to regulatory submission or operational action.

Human-oversight practices should be embedded within existing standard operating procedures (SOPs) and quality systems to ensure consistency with established governance, training, deviation management, and inspection readiness expectations. These controls reinforce that AI outputs serve as decision support rather than autonomous decision-makers

The guidance is explicit: AI outputs used in regulatory submissions should be viewed as supportive evidence, not sole arbiters [4][23]. Sponsors must define the role of human judgment in interpreting and acting on AI-driven insights.

AI tools often require access to sensitive clinical, genomic, or manufacturing data. Ensuring the security and confidentiality of these datasets is essential—especially when models are hosted in cloud environments or integrated across multiple systems.

Sponsors must demonstrate that appropriate safeguards are in place for data storage, access, and transmission. Additionally, privacy-preserving techniques such as federated learning or differential privacy may be explored when handling patient-level data [24][25].

Finally, integrating AI into drug development processes will likely require early and frequent interaction with regulatory authorities. The FDA encourages sponsors to engage through existing pathways such as pre-Investigational New Drug (IND) meetings, Type C meetings, or the Innovation and Science Technology Approaches for New Drugs (ISTAND) pilot program when proposing novel methodologies.

Documentation must clearly state the AI tool’s intended use, performance metrics, limitations, and any associated risk controls. Submissions should include technical files (e.g., source code summaries, version histories) and scientific justification for how AI outputs support the drug’s safety, efficacy, or quality [4].

Traditional regulatory frameworks were built around fixed, well-characterized processes. AI, by contrast, is dynamic. Models may evolve through continuous learning, retraining, or integration of real-world evidence. This demands a more adaptive regulatory model—one that recognizes and supports controlled evolution over time without compromising quality or safety.

Regulators are exploring approaches such as predetermined change control plans (PCCPs), model updates within defined limits, and real-time monitoring. These mechanisms can allow innovation to proceed while keeping oversight tight where it matters most.

As AI becomes integral to the drug development value chain—from preclinical modeling to manufacturing analytics—future success will hinge on deep collaboration between pharmaceutical scientists, data scientists, regulatory experts, and clinicians. No single discipline can navigate these complexities alone.

This will also require upskilling the workforce across both industry and regulatory agencies. Regulatory scientists will need fluency in AI methods, while developers must build systems with compliance, explainability, and risk control in mind from day one [8].

Inconsistent expectations across global regulators can slow down innovation and create duplication of effort. Initiatives like the International Council for Harmonization of Technical Requirements for Pharmaceuticals for Human Use (ICH) work on digital technologies and FDA’s collaboration with international health authorities signal a move toward harmonization.

As industry adoption grows, we may also see the emergence of shared AI validation frameworks, model cards, and documentation templates. These can reduce review burden, improve transparency, and set industry-wide expectations for responsible use.

So far, much attention has focused on AI’s use in early-phase discovery or limited clinical trial support. But its impact on manufacturing, supply chain optimization, pharmacovigilance, and post-market surveillance is only beginning to unfold.

FDA has signaled openness to AI-enabled tools across all stages, provided that their role is clearly defined and their limitations understood. The future may include AI-driven process control in real-time release testing, or safety signal detection across massive real-world datasets. Each application will come with unique regulatory nuances that demand tailored solutions.

Ultimately, AI in pharma must serve patients. This means striking a balance between innovation and caution, speed and scrutiny. Ethical considerations—such as fairness, accountability, and data privacy—must be embedded in both technology and policy design.

Regulators like FDA are playing a critical role in shaping a culture of responsible innovation. Their guidance encourages early engagement, continuous learning, and transparent communication—principles that will underpin the next generation of digital transformation in healthcare [8][26]-[29] (Figure 8).

Figure 8. Future vision for the responsible adoption of artificial intelligence in drug development.