How AI Orchestrates End-to-End Regulatory Workflows

Contents

A completed clinical study report does not arrive independently. It is the downstream product of decisions made in the protocol, information consolidated in the investigator's brochure, risk parameters set in the development safety update report, and consent language derived from eligibility criteria written months or years earlier. When any upstream document changes, every document downstream is potentially compromised. That interdependency is precisely where regulatory workflows break down, and where AI orchestration offers something more than productivity gains: a fundamentally different way of managing the clinical document lifecycle as a connected system rather than a sequence of isolated writing tasks.

AI orchestration

How AI orchestrates the regulatory document lifecycle

Structured protocol source

Machine-readable protocol fields, eligibility criteria, endpoints, procedures, safety rules

↓

Specialized AI agents

Protocol agent, ICF agent, IB cross-reference agent, DSUR agent, CSR agent, compliance-checking agent

↓

Shared structured data layer

Versioned source fields, document dependencies, amendment history, input records, model outputs

↓

Downstream regulatory documents

ICF, IB, DSUR, SAP, CSR, patient materials, regulatory submissions

↓

Human review and approval

Qualified medical writer review, regulatory review, sponsor approval, audit-ready sign-off

Feedback loop

Protocol amendment → structured data layer → affected document sections flagged → human review → approved downstream update

AI orchestration governs the relationships between documents, not just the generation of individual drafts.

Why Document Fragmentation Is a Clinical Trial Operations Problem

The regulatory document burden in clinical development is not simply about volume. It is about consistency across documents that share data but are produced by different teams, in different systems, at different points in the trial timeline.

A 2024 Tufts Center for the Study of Drug Development analysis published in Therapeutic Innovation and Regulatory Science examined 950 protocols and 2,188 amendments across 16 pharmaceutical companies and contract research organizations ^[1]. The results showed that 76% of Phase I-IV trials now require at least one protocol amendment, up from 57% in 2015, and that the mean number of amendments per protocol has risen 60% to 3.3, compared with 2.1 in the prior benchmark period. Each of those amendments propagates through the document set: a change to eligibility criteria in the protocol requires corresponding revisions to the informed consent form, updated language in patient-facing materials, and potentially reconciliation with data already captured. A 2016 Tufts CSDD study in the same journal found that the median direct cost to implement a substantial amendment was $141,000 for Phase II protocols and $535,000 for Phase III protocols ^[2], and those figures do not capture indirect costs from delayed timelines, site disruptions, or regulatory queries.

The volume of amendment submissions gives a sense of how much regulatory overhead this creates in practice. A 2026 analysis published in the British Journal of Clinical Pharmacology evaluated MHRA clinical trial applications between September 2023 and August 2024 and found that the agency received 4,002 substantial amendment submissions against 615 initial applications in the same twelve-month period ^[3]. Amendments outnumbered original submissions by more than six to one. That ratio illustrates how much of the industry's regulatory overhead is spent not on original design but on correcting documents that were inconsistent with each other or with evolving evidence.

A peer-reviewed analysis of 129 FDA warning letters issued to clinical investigators between 2005 and 2010, published in the Indian Journal of Medical Ethics, found that deviation from the investigational plan was the most common violation theme, appearing in 81% of investigator warning letters ^[4]. The relationship between high amendment rates and downstream site-level adherence challenges is discussed across clinical trial performance literature, though direct causal attribution between amendment frequency and specific warning-letter patterns is difficult to establish from available data alone ^[1]^,[3].

What the Research Shows About AI in Regulatory Document Generation

The use of large language models (LLMs) for regulatory document generation has moved from theoretical exploration to empirical evaluation at meaningful scale. The evidence reveals both credible capability and well-defined limitations.

A 2025 study published in the Journal of the American Medical Informatics Association introduced InformBench, a benchmark comprising 900 clinical trial documents, and evaluated LLM performance on transforming trial protocols into informed consent forms ^[5]. Baseline models including GPT-4o achieved only 70-80% compliance with regulatory requirements and exhibited factual errors in 18-43% of cases. The InformGen pipeline, which combined retrieval-augmented generation with human-in-the-loop review, achieved near-100% regulatory compliance and over 90% factual accuracy as validated by five domain-expert annotators.

Separately, a study published in the journal Clinical Trials evaluated LLM-generated protocol sections across four dimensions: clinical thinking and logic, transparency and references, medical terminology, and content relevance ^[6]. Retrieval-augmented generation, seeded with regulatory guidance documents and structured trial data, materially outperformed off-the-shelf generation on clinical thinking and logic, though both approaches showed comparable strength in terminology accuracy. The pattern across both studies is consistent: models grounded in structured regulatory content produce more accurate outputs than models generating from general pretraining data alone.

These findings carry an important operational implication. The question is not whether an LLM can generate a regulatory document at all. Under the right conditions, it can produce a credible first draft. The question is whether the generation is governed well enough, and connected tightly enough to source documents, to propagate correctly through the full document set when inputs change.

How Orchestration Changes the Architecture

A standalone AI document tool and an orchestrated AI regulatory system are different things. The distinction is not about sophistication; it is about architecture.

In a standalone tool, a medical writer provides inputs and the system produces a document. Each document is generated independently. If the protocol changes, the writer must identify every downstream document affected and revise each one manually. Cross-document consistency is enforced by human review, which is unreliable at scale and slow under timeline pressure.

In an orchestrated system, specialized agents operate within a defined workflow graph, each handling a bounded task and passing structured outputs to the next stage. A 2025 peer-reviewed article in Federal Practitioner described the general architecture of multiagent healthcare AI systems, in which each agent handles specific aspects of a complex workflow while a coordinating layer routes tasks and aggregates outputs ^[7]. The clinical regulatory document context applies the same architectural principle: a protocol agent, an ICF agent, an IB cross-reference agent, and a compliance-checking agent can each operate on their respective scope while sharing a common structured data layer. The Borkowski and Ben-Ari article supports the general multiagent framework; the specific application to regulatory document production represents an extension of that model into the clinical documentation domain.

When the protocol changes in this design, the orchestrator identifies which downstream documents are affected and triggers regeneration or flagging automatically, rather than waiting for a human to trace every dependency manually. The differences between the two models can be summarized as follows:

Dimension	Standalone AI generation	Orchestrated regulatory workflow
Document scope	Single document per task	Full document set managed as connected system
Source of truth	Writer provides inputs per document	Shared structured data layer queried by all agents
Amendment propagation	Manual: writer identifies and revises each affected document	Automated: orchestrator flags or triggers all derived content
Consistency enforcement	Human review after each draft	Consistency checks embedded throughout the workflow
Audit trail	Document-level version history	Field-level log of every agent call, input, output, and approval
Validation requirement	System/tool validation based on context of use	Pipeline validation proportional to risk and context of use; 21 CFR Part 11 applies where qualifying electronic records or signatures are involved [12]

The audit dimension is equally important. ICH E6(R3) Annex 1 requires that records be organized such that audits and inspections can trace any data point back to its source ^[10]. An orchestrated system designed to log every agent call, every input queried, every output generated, and every human review action produces an inherently traceable record. This field-level logging is not stated verbatim in either 21 CFR Part 11 or ICH E6(R3), but it represents the implementation design pattern that best supports the traceability and source-review expectations both frameworks establish ^[12]^,[10].

A Concrete Example: Eligibility Criteria and the Propagation Problem

Consider what happens when a sponsor narrows an inclusion criterion in the protocol, say, restricting enrollment from adults aged 18 and above to those aged 30 and above, following interim safety findings. In a disconnected document workflow, the protocol amendment is filed, but the informed consent form still contains the original age language, the recruitment materials reference the original population, the statistical analysis plan was written around the broader cohort, and the CSR's enrollment section will later need to reconcile data collected before and after the amendment. Each of these documents may be maintained by a different team using a different template system.

In an orchestrated workflow, the eligibility field in the structured data layer is updated once. The orchestrator identifies every downstream document containing derived age-criterion content and either triggers a revision with the updated value or flags the affected sections for human review. The amendment history is logged at the data layer, so the CSR authoring team can see exactly when the criterion changed and which data records were collected under each criterion version. That traceability supports inspection-readiness when reviewers need to understand which document version governed which data, a question that arises in both routine audits and regulatory inspections. (For a deeper look at amendment costs and prevention strategies, see also: Protocol Amendments: Hidden Costs and How to Prevent Them.)

The Regulatory and Data Standards Enabling Machine-Readable Workflows

Several converging regulatory standards are making orchestrated document workflows technically feasible at an infrastructure level, not merely conceptually possible.

ICH M11 CeSHarP: Adopted under Step 4 by the ICH Assembly on November 19, 2025, and finalized in the Federal Register on May 22, 2026, the ICH M11 Clinical Electronic Structured Harmonised Protocol (CeSHarP) establishes a machine-readable, internationally harmonized standard for the content and electronic exchange of clinical trial protocol information ^[8]. (For a detailed treatment of M11's structure and implications, see: ICH M11 CeSHarP and the Future of Protocol Standardization.) The technical specification, issued in draft form in June 2025, recommends an open, non-proprietary standard that enables electronic exchange of protocol content across regulatory authorities, sponsors, ethics bodies, and investigators ^[9]. For AI orchestration, M11 is foundational: a structured, field-level protocol representation allows downstream document generation systems to ingest protocol data programmatically rather than parsing unstructured narrative text. M11 enables the structured exchange; whether downstream documents remain consistent as the protocol evolves still depends on how deliberately the orchestration system is designed to consume, version, and propagate those fields.

ICH E6(R3): Finalized at Step 4 on January 6, 2025, and published as FDA guidance in September 2025, with EMA bringing Annex 1 into effect on July 23, 2025, ICH E6(R3) Good Clinical Practice addresses computerized systems, electronic records, and data governance in a technology-neutral framework that applies equally to digitally-enabled and AI-assisted workflows ^[10]^,[11]. The updated guideline introduces formal requirements for computerized system validation, audit trail integrity, and risk-based quality management across all trial stages. Annex 1 requires sponsors to maintain a comprehensive system inventory covering each system's purpose, validation status, access controls, and management responsibilities. FDA guidance documents, including E6(R3), represent the agency's current thinking and recommendations; they are not legally binding unless specific regulatory or statutory requirements are cited. (See also: AI Audit Trails in Clinical Trials: What E6(R3) Actually Requires.)

21 CFR Part 11: 21 CFR Part 11 establishes FDA requirements for electronic records and electronic signatures that are used to satisfy FDA-regulated record-keeping requirements, specifying that such records must be accurate, reliable, and protected by secure, computer-generated, time-stamped audit trails ^[12]. The regulation applies when a sponsor uses electronic means to create, modify, maintain, archive, or transmit records that FDA requires in paper or electronic form. For AI-generated documents to enter the regulatory record, they must be produced by validated systems that meet these controls. The validation requirement under 21 CFR 11.10(a) covers accuracy, reliability, and consistent intended performance, as well as the ability to discern invalid or altered records.

FDA Draft Guidance on AI in Regulatory Decision-Making (FDA-2024-D-4689): Published in January 2025, this draft guidance introduced a risk-based credibility assessment framework for AI used to generate information supporting regulatory decisions on drug safety, effectiveness, and quality ^[13]. The framework is explicitly context-of-use specific: the level of validation evidence required scales with the regulatory weight of the output and the risk consequences of an error. As of the date of this article, the guidance remains in draft form; FDA has not publicly announced a finalization date.

FDA-EMA Joint Guiding Principles on AI (January 2026): On January 14, 2026, the FDA and EMA jointly published ten guiding principles for good AI practice in drug development, spanning clinical trials, manufacturing, and post-market surveillance ^[14]. The principles call for AI systems to be high quality, human-centric, and compliant with applicable ethical, scientific, regulatory, and cybersecurity standards. These principles are not formal binding guidance; they represent both agencies' current thinking and are intended to inform more detailed future guidance.

Operational Mechanics of End-to-End Workflow Orchestration

What end-to-end orchestration looks like in practice, across the clinical document lifecycle, is more concrete than the concept suggests.

At the upstream end, the protocol represents the source of truth. In an orchestrated workflow, the approved protocol, ideally structured according to ICH M11 CeSHarP, populates a structured data layer that downstream agents can query. The ICF generation agent draws from this layer to produce consent language that accurately reflects the study's eligibility criteria, procedures, and risks, rather than relying on a medical writer to manually transpose protocol content into patient-facing language. The InformBench research demonstrated that retrieval-based generation from structured protocol sources achieves materially higher factual accuracy than generation from general context ^[5].

At the mid-trial level, the orchestrator monitors document state against incoming data. When a protocol amendment is filed, the system identifies every field in the structured data layer that has changed and flags or triggers revision of downstream documents containing derived content. Human review remains a required step for regulated content. The orchestrator narrows the review burden to specific sections that changed, rather than requiring a complete re-review of every document in the set.

At the reporting end, the CSR is by definition a synthesis document. Its efficacy and safety summaries must reconcile with the statistical analysis plan, the clinical data listings, the IB risk characterization, and the protocol-specified endpoints. In a disconnected workflow, these reconciliations are performed manually during authoring, often revealing discrepancies late in the process. In an orchestrated workflow, consistency checks run against structured data throughout the trial lifecycle, surfacing potential conflicts while there is still time to address them before submission.

Regulatory Considerations and What AI Orchestration Does Not Eliminate

The regulatory frameworks described above set a clear expectation: AI can assist with regulatory document production, but human accountability cannot be delegated to the system. FDA-2024-D-4689 specifies a risk-based credibility assessment framework in which validation evidence requirements scale with the context of use and the risk consequences of an incorrect output ^[13]. The FDA-EMA joint principles explicitly frame AI as a support for human decision-making, not a replacement for qualified review and authorship accountability ^[14].

Sponsors introducing AI orchestration into document production should address at least the following elements before deployment, each grounded in existing regulatory requirements:

Regulatory deployment checklist

Seven controls for AI orchestration in regulated document production

[1] Source traceability

Every field in a generated document should be traceable to a specific input record in the structured data layer, with version history capturing when each input changed, satisfying ICH E6(R3)'s requirement that any data point can be traced back to its source [10].

[2] Model versioning and reproducibility

The AI model configuration used to generate each document version should be recorded and locked. Regulatory workflows require consistent outputs; model updates must go through documented change control.

[3] Human approval workflow

Qualified medical writers and regulatory affairs professionals retain authorship and approval responsibility. Formal sign-off steps should be captured in the audit log.

[4] Audit log completeness

System logs should capture every agent call, every input queried, and every output produced in a secure, time-stamped record satisfying 21 CFR 11.10 controls [12].

[5] Regulatory rule testing

Domain-specific compliance rules, such as the 18 regulatory rules used in the InformBench evaluation framework [5], should be codified as testable assertions and run as part of acceptance testing for each system release.

[6] Change-propagation validation

Sponsors should validate not only the accuracy of generated document content but also the logic by which changes in one document trigger updates in derived documents. An amendment that propagates incorrectly is as hazardous as one that is not propagated at all.

[7] Privacy and data protection

Clinical trial protocols and patient-derived data are sensitive. AI systems processing this content must operate within validated, access-controlled environments with documented data handling procedures consistent with the sponsor's data governance framework and applicable privacy regulations.

Sponsors evaluating AI orchestration vendors should, at minimum, request documentation of the vendor's validation approach under 21 CFR Part 11, the architecture of their audit trail, how model versioning and change control are managed, and whether the system produces source-traceable outputs at the field level rather than only at the document level.

How Kitsa Supports Connected Regulatory Document Production

Kitsa's KScribe product is designed for the workflow problem described above. According to Kitsa, KScribe is built to share structured document data across the clinical document set, so that downstream documents reflect upstream source content with greater consistency than manual workflows maintain at scale. Kitsa states that the system is designed to include human-review controls and audit capabilities intended to align with 21 CFR Part 11 and ICH E6(R3) expectations. Sponsors can review Kitsa's product positioning at kitsa.ai/regulatory-document-generation and should request vendor validation documentation directly when evaluating the platform for regulated workflows. As with any AI system used in a submission context, validation status should be assessed against the sponsor's specific intended context of use, as FDA-2024-D-4689 makes clear that validation requirements scale with the regulatory weight of the output ^[13].

KScribe · AI Regulatory Workflow Orchestration

End-to-end regulatory workflows do not fail because teams cannot write documents. They fail because protocols, ICFs, IBs, DSURs, SAPs, CSRs, amendments, review actions, and audit records are managed as disconnected artifacts. KScribe is designed to support connected regulatory document production across the clinical document lifecycle, helping sponsors manage source traceability, cross-document consistency, human review, and audit-ready workflow records.

Explore KScribe

Key Takeaways

The average Phase I-IV protocol now requires 3.3 amendments, a 60% increase since 2015, with direct implementation costs reaching $141,000 for Phase II and $535,000 for Phase III protocols per amendment in Tufts CSDD analysis ^[1]^,[2].
AI orchestration differs from standalone document tools in that it governs the relationships between documents, not merely the generation of individual documents, enabling automated propagation of changes and consistency checking across the full set.
Retrieval-augmented generation from structured protocol sources achieves near-100% regulatory compliance in ICF generation, compared to 70-80% for unstructured baseline LLMs, underscoring that document grounding and architecture matter as much as model capability ^[5].
ICH M11 CeSHarP, finalized in the Federal Register in May 2026, provides the machine-readable protocol standard that makes structured downstream document generation technically feasible at scale ^[8].
ICH E6(R3), effective in the EU from July 23, 2025, and published as FDA guidance in September 2025, establishes computerized system validation and audit trail requirements in a technology-neutral framework that AI-assisted document pipelines must satisfy ^[10]^,[11].
The FDA-EMA joint guiding principles published January 14, 2026 are not formally binding guidance, but they explicitly frame AI as a support for human decision-making and lay the foundation for future detailed guidance from both agencies ^[14].
End-to-end orchestration does not eliminate expert human judgment; it concentrates that judgment on scientifically meaningful decisions rather than mechanical consistency checking and document transposition.

Frequently Asked Questions

What is AI regulatory workflow orchestration in clinical trials?

AI regulatory workflow orchestration refers to the use of interconnected AI agents, governed by a coordinating layer, to manage the generation, cross-referencing, and consistency checking of clinical regulatory documents across the trial lifecycle. Unlike a standalone AI writing tool, an orchestrated system shares structured data across documents such as the protocol, ICF, IB, DSUR, and CSR, so that when one document changes, the system identifies downstream impacts automatically rather than relying on manual review to catch discrepancies.

Does FDA allow AI-generated content in regulatory submissions?

FDA's January 2025 draft guidance FDA-2024-D-4689 provides a risk-based credibility assessment framework for AI used to generate information supporting regulatory decisions on drug safety, effectiveness, and quality ^[13]. The guidance does not prohibit AI-generated content; it establishes the validation and documentation requirements that apply when sponsors use AI to produce submission-relevant data. The guidance remains in draft form as of mid-2026. FDA guidance documents represent the agency's current thinking and are not legally binding unless specific regulatory or statutory requirements are cited.

What does ICH E6(R3) require for computerized systems used in document production?

ICH E6(R3), finalized at Step 4 on January 6, 2025, effective in the EU from July 23, 2025, and published as FDA guidance in September 2025, addresses computerized systems through requirements for system validation, audit trail integrity, and risk-based quality management ^[10]^,[11]. Annex 1 requires sponsors to maintain a comprehensive system inventory covering each system's purpose, validation status, access controls, security measures, and management responsibilities. The guideline is technology-neutral and does not create AI-specific document production requirements; it establishes standards that apply to any digital system handling trial records.

What is ICH M11 CeSHarP and why does it matter for AI document orchestration?

ICH M11 CeSHarP is the internationally harmonized standard for the content and electronic exchange of clinical trial protocol information, adopted by the ICH Assembly on November 19, 2025, and published in the Federal Register on May 22, 2026 ^[8]. It defines standardized data fields, harmonized terminologies, and a machine-readable format for protocol content. For AI orchestration, M11 is foundational because it allows downstream document generation systems to ingest protocol data programmatically, replacing the error-prone process of manually parsing narrative protocol text to populate consent forms, study reports, and other derived documents.

What are the main limitations of AI for regulatory document workflows?

Current LLMs without domain-specific grounding exhibit factual error rates of 18-43% in ICF generation tasks and achieve only 70-80% regulatory compliance on baseline evaluation ^[5]. These limitations are addressable through retrieval-augmented generation and structured human review controls, but they require deliberate system design. AI document systems in regulated environments must also address output reproducibility across model versions, validation under 21 CFR Part 11, and the explicit human oversight requirements of FDA-2024-D-4689 ^[13]. Orchestration narrows the manual burden substantially but does not replace qualified authorship or regulatory accountability.

Are the FDA-EMA joint AI principles from January 2026 binding?

No. The ten guiding principles of good AI practice published jointly by the FDA and EMA on January 14, 2026 are not formal binding guidance ^[14]. They represent both agencies' current thinking and are explicitly intended to inform more detailed future guidance, standards, and harmonized regulatory expectations. The principles establish expectations around transparency, traceability, validation, and human-centric design. Sponsors already meeting the requirements of 21 CFR Part 11, FDA-2024-D-4689, and ICH E6(R3) are substantially aligned with the principles' intent.

References

[1]Getz K, Smith Z, Botto E, Murphy E, Dauchy A. "New Benchmarks on Protocol Amendment Practices, Trends and Their Impact on Clinical Trial Performance." Ther Innov Regul Sci. 2024;58(3):539-548. doi:10.1007/s43441-024-00622-9. PMID: 38438658. https://pubmed.ncbi.nlm.nih.gov/38438658/
[2]Getz K, Stergiopoulos S, Short M et al. "The Impact of Protocol Amendments on Clinical Trial Performance and Cost." Ther Innov Regul Sci. 2016;50(4):436-441. doi:10.1177/2168479016632271. https://link.springer.com/article/10.1177/2168479016632271
[3]Manfrin A et al. "Evaluation of the Medicines and Healthcare Products Regulatory Agency's Introduction of a Risk-Proportionate Approach for Clinical Trials: An Analysis of 4,617 Applications Assessed Between September 2023 and August 2024." British Journal of Clinical Pharmacology. 2026;92(3):822-829. doi:10.1002/bcp.70308. PMC12930005. https://pmc.ncbi.nlm.nih.gov/articles/PMC12930005/
[4]Gogtay NJ, Doshi BM, Kannan S, Thatte U. "A Study of Warning Letters Issued to Clinical Investigators and Institutional Review Boards by the United States Food and Drug Administration." Indian Journal of Medical Ethics. 2011 Oct. doi:10.20529/IJME.2011.082. PMID: 22106658. https://pubmed.ncbi.nlm.nih.gov/22106658/
[5]Wang Z, Gao J, Danek B, Theodorou B, Shaik R, Thati S, Won S, Sun J. "Compliance and Factuality of Large Language Models for Clinical Research Document Generation." Journal of the American Medical Informatics Association. 2025 Oct 27. doi:10.1093/jamia/ocaf174. PMID: 41144289. https://pubmed.ncbi.nlm.nih.gov/41144289/
[6]Markey N, El-Mansouri I, Rensonnet G, Langen C, Meier C. "From RAGs to Riches: Using Large Language Models to Write Documents for Clinical Trials." Clinical Trials. 2025. arXiv:2402.16406. https://arxiv.org/abs/2402.16406
[7]Borkowski AA, Ben-Ari A. "Multiagent AI Systems in Health Care: Envisioning Next-Generation Intelligence." Federal Practitioner. 2025 May;42(5):188-194. doi:10.12788/fp.0589. PMC12360800. https://pmc.ncbi.nlm.nih.gov/articles/PMC12360800/
[8]U.S. Food and Drug Administration / Federal Register. "M11 Clinical Electronic Structured Harmonised Protocol (CeSHarP); International Council for Harmonisation; Guidance for Industry; Availability." Federal Register. May 22, 2026. https://www.federalregister.gov/documents/2026/05/22/2026-10295/m11-clinical-electronic-structured-harmonised-protocol-cesharp-international-council-for
[9]U.S. Food and Drug Administration / Federal Register. "M11 Technical Specification: Clinical Electronic Structured Harmonised Protocol." Federal Register. June 6, 2025 (Draft). https://www.federalregister.gov/documents/2025/06/06/2025-10359/m11-technical-specification-clinical-electronic-structured-harmonised-protocol-international-council
[10]International Council for Harmonisation. "Guideline for Good Clinical Practice E6(R3)." ICH Harmonised Guideline, Step 4. January 6, 2025. https://database.ich.org/sites/default/files/ICH_E6(R3)_Step4_FinalGuideline_2025_0106.pdf
[11]U.S. Food and Drug Administration. "E6(R3) Good Clinical Practice." FDA Guidance for Industry, September 2025. CDER/CBER. Docket No. FDA-2023-D-1955. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e6r3-good-clinical-practice-gcp
[12]U.S. Food and Drug Administration. "21 CFR Part 11: Electronic Records; Electronic Signatures." Code of Federal Regulations. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-A/part-11
[13]U.S. Food and Drug Administration. "Considerations for the Use of Artificial Intelligence To Support Regulatory Decision Making for Drug and Biological Products." FDA Draft Guidance for Industry. Docket No. FDA-2024-D-4689. January 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological
[14]U.S. Food and Drug Administration. "Guiding Principles of Good AI Practice in Drug Development." CDER/CBER in collaboration with the European Medicines Agency. January 2026. https://www.fda.gov/about-fda/artificial-intelligence-drug-development/guiding-principles-good-ai-practice-drug-development