Agentic AI vs AI-Assisted Tools in Clinical Trial Protocols

Contents

Most clinical trial teams have had some version of this experience: a protocol clears internal review, enters pre-submission preparation, and someone notices that the primary endpoint phrasing in Section 3 no longer matches the analysis section written three weeks later. A statistical assumption in the synopsis does not align with the eligibility criteria. The IB reference cited in the background section is two versions out of date. None of this is caught automatically, because the tools in use are competent at individual tasks and blind to everything else.

That is a product of AI-assisted tooling, which is how many clinical development and medical writing teams currently work with artificial intelligence. It is also the specific operational gap that agentic AI is designed to address. The two approaches are not interchangeable names for the same general category of technology. They differ in architecture, scope, and what they can actually do inside the regulated environment of a clinical trial. For sponsors, CROs, and medical writers deciding how to invest in AI infrastructure for protocol development, understanding this distinction is a practical question, not a theoretical one.

Teams looking at AI solutions for regulatory document generation can find background on how Kitsa approaches this at kitsa.ai/protocol-development.

Why This Topic Matters in Clinical Trials

Protocol development sits at the center of every downstream decision in a clinical trial. The eligibility criteria determine who can be enrolled. The endpoints drive the statistical analysis plan. The dosing rationale connects to the investigator's brochure. The risk management section structures monitoring and DSMB review. When those elements are internally inconsistent, or when changes to one section are not reflected throughout the document, the result is often an avoidable protocol amendment or a regulatory query that delays approval.

The scale of this problem is documented. A Tufts Center for the Study of Drug Development (Tufts CSDD) analysis published in Therapeutic Innovation and Regulatory Science in 2024, drawing on data from 950 protocols and 2,188 amendments across 16 pharmaceutical companies and CROs, found that 76% of Phase I through Phase IV trials required at least one amendment, up from 57% in 2015 ^[1]. The cost of each amendment compounds the problem. Tufts CSDD estimates that a single substantial amendment runs between $141,000 for Phase II studies and $535,000 for Phase III trials, figures that exclude CRO contract change orders, site disruption costs, and enrollment delays ^[2]. (For a deeper look at what drives amendment costs, see Why Protocol Amendments Cost Sponsors Millions on the Kitsa blog.)

An IQVIA analysis published in January 2026, examining protocol design trends since 2005, found a 214% rise in the number of endpoints per trial, a 139% increase in the number of procedures, and a 600% surge in the volume of data points collected ^[3]. A Tufts CSDD study published in Therapeutic Innovation and Regulatory Science in 2023, analyzing 187 protocols from 20 major pharmaceutical companies and CROs, confirmed a continuing upward trend across all protocol design variables including endpoints, eligibility criteria, protocol pages, and data points collected ^[4]. When a Phase III protocol now contains more than twice the endpoints and roughly twice the procedures it did twenty years ago, the probability that a single writer or review cycle will catch every internal inconsistency is not a process failure. It is a structural one.

These are not problems that a grammar-aware copilot will resolve.

Current Evidence and State of Research

How Protocol Complexity Has Outpaced Manual Review

The increase in protocol complexity is measurable, not perceived. A 2024 machine learning analysis published in Scientific Reports examined more than 16,000 trials registered on ClinicalTrials.gov and assigned complexity scores based on the number of endpoints, eligibility criteria, and trial arms ^[5]. Complexity scores rose across all trial phases between 2011 and 2022, with the average score increasing by more than 10 percentage points in that period ^[5]. Oncology trials showed the highest absolute complexity, but cardiovascular trials showed the steepest rate of increase, driven in part by wearable device endpoints ^[5].

The eligibility criteria problem specifically has been documented over time. Tufts CSDD data from the Getz 2023 benchmarks study showed continuing increases in eligibility criteria per trial across all phases, and a 2022 JAMA Oncology analysis by investigators at UT Southwestern and three other academic cancer centers concluded that overly restrictive and complex eligibility criteria have created substantial barriers to patient access, hindered trial recruitment, and limited the generalizability of results ^[6]. Overly complex eligibility criteria are also a documented contributor to protocol deviations: according to Tufts CSDD data compiled in the 2025 SCRS Landscape Survey Report, mean deviations per pivotal trial have more than doubled over the prior decade, a trend Tufts CSDD Executive Director Ken Getz directly linked to protocol design complexity ^[7].

The implication is that protocol complexity is creating documentation and execution challenges simultaneously. More endpoints, more criteria, more cross-references between sections mean more surface area for inconsistency. More surface area for inconsistency means more opportunities for avoidable amendments, regulatory queries, and deviation reports.

Two Categories of AI, Not One

The current generation of AI tools in clinical research covers a wide range of functionality, but the operational distinction that matters for protocol development comes down to a specific question: does the system maintain context across the entire workflow, or does it respond to discrete task prompts without persistent memory of what came before?

AI-assisted tools, which include generative AI interfaces, large language model (LLM)-powered copilots, retrieval-augmented generation (RAG) systems, and template-filling automation, all share the same fundamental limitation when used in isolation: they operate on the section or task in front of them. A medical writer using an AI-assisted platform can ask it to draft an eligibility criteria section, summarize a competitor protocol, or generate a synopsis from a full document. The tool responds. The human acts on the output, edits it, moves it into the document, and manually checks it against adjacent sections. Even sophisticated RAG-based systems that can retrieve relevant content from a knowledge base are still producing outputs for discrete queries. The intelligence in the system does not persist across the full document workflow.

Agentic AI operates differently in a specific and meaningful way. As the Drug Information Association's Global Forum described in its August 2025 analysis of next-generation AI in clinical research, agentic AI is not simply about answering prompts ^[8]. Agents can plan sequences of tasks, maintain context across long-running processes, and coordinate with humans and other agents between steps ^[8]. In a protocol development context, this means a governed agentic system could be designed to draft an eligibility criteria section, check it against the IB for consistency, flag a discrepancy with the endpoint section, and surface that discrepancy to the medical writer with a specific document reference, all without the writer issuing a new instruction for each step.

A December 2025 Everest Group analysis of agentic AI in clinical trials, a commercial consulting report, framed this operational distinction in similar terms: agentic AI is designed for autonomous decision-making and active workflow management, while generative AI focuses on content creation and data synthesis for individual tasks ^[9]. Both categories are real and both have genuine clinical research applications. They are not interchangeable.

Operational Implications for Protocol Design

Where AI-Assisted Tools Genuinely Help

Being specific about what AI-assisted tools do well is necessary for an honest account of where agentic AI adds something different.

AI-assisted tools have meaningfully reduced the time required to produce first drafts. A medical writer working with a generative AI platform to draft protocol background and rationale sections can complete that work considerably faster than working from scratch, particularly when the tool has access to relevant trial data, therapeutic area literature, and prior protocols. A 2025 implementation study published in the Journal of Personalized Medicine found that AI chatbot assistance in protocol development improved research engagement and reduced documentation timelines in the assessed setting ^[10]. RAG-based systems that retrieve content from historical protocols have shown improvement in information extraction accuracy over standalone LLMs in published benchmarking work from 2026 ^[11].

The operational ceiling is scope. An AI-assisted tool handles the section in front of it. It does not hold the rest of the protocol in working memory while generating a new subsection. It does not know that the statistical assumptions used in the sample size calculation are inconsistent with the eligibility criteria written earlier in the same session. It cannot check whether a change to the primary endpoint in the synopsis has been reflected in the outcome measures section, the analysis plan, and the data management section. That cross-document work remains manual regardless of how capable the generative model is.

This limitation is not inherent to the sophistication of the underlying model. It is a function of how the system is architected and governed. A long-context model with full document access can identify some inconsistencies in a single pass, and a RAG-based system can retrieve and compare content across sections. What they typically do not do, when deployed as standalone copilots without agentic orchestration and governance infrastructure, is maintain that context persistently across multiple sessions, propagate an approved change through every affected section in a documented sequence, and produce a complete audit trail of who directed what, what was generated, and what a qualified reviewer approved. That combination of persistent orchestration, governed change control, and auditability is what separates an AI-assisted tool from a production-ready agentic workflow in a regulated document context.

A Concrete Example: Endpoint Change Propagation

Consider what happens when a sponsor decides, mid-drafting, to change the primary endpoint from progression-free survival to overall survival in an oncology protocol. In a manual or AI-assisted workflow, that change must be propagated by hand through at least six distinct sections: the objectives, the schedule of assessments, the statistical analysis plan, the informed consent form, the synopsis, and the data management section. Each of those sections was written at a different time, likely by different people, possibly in a different file.

Manual / AI-Assisted

Endpoint changed in Section 3 (Objectives)
↓
Writer manually checks Objectives [checked or missed?]
↓
Writer manually checks Schedule of Assessments [checked or missed?]
↓
Writer manually checks SAP [checked or missed?]
↓
Writer manually checks ICF [checked or missed?]
↓
Writer manually checks Synopsis [checked or missed?]
↓
Writer manually checks Data Management [checked or missed?]
↓
Risk: inconsistency survives to review or submission

Governed Agentic Workflow

Endpoint changed in Section 3 (Objectives)
↓
Agent scans all 6 dependent sections automatically
↓
Agent presents complete change list to writer
↓
Writer reviews and approves each change
↓
System logs review + approval with timestamp
↓
Protocol exits session with complete change history
↓
Outcome: traceable, auditable, consistent

In a governed agentic workflow, the change to the endpoint triggers an automated scan across all sections that reference the affected variable. The agent identifies every downstream document passage containing "progression-free survival," presents the complete list to the medical writer, and drafts updated language for each location consistent with the new endpoint definition. The writer reviews and approves each change. The system logs the review and approval with a timestamp. The protocol leaves the session with a complete change history, not a checklist of items that may or may not have been manually updated.

This is not a speculative scenario. Causaly's January 2026 analysis of AI agentification in pharma R&D described this category of document-level workflow precisely: a governed agentic workflow produces repeatable, reviewable artifacts, whereas a copilot accelerates individuals on discrete tasks ^[12]. A protocol is an artifact. Its production benefits from repeatability and reviewability, not just faster drafting of individual sections.

The Amendment Rate Question

No published clinical evidence currently links agentic AI protocol tooling to lower amendment rates, and this article does not claim otherwise. The 2024 Tufts CSDD amendment benchmarks study does not decompose amendment causes in a way that isolates cross-document inconsistency as a root driver ^[1]. The connection between agentic AI and amendment reduction is a structural argument, not an empirically demonstrated finding.

What is established is that approximately one-quarter of amendments are classified as avoidable at study initiation ^[1]^,[2], and that the categories of avoidable amendments may include issues identifiable at the initial design stage: endpoints or eligibility criteria that were revised without being updated throughout the document, protocols that did not anticipate regulatory feedback requirements, and inconsistencies between the protocol and related documents such as the ICF or SAP. Agentic systems capable of maintaining cross-document context are architecturally better positioned to catch these issues before they trigger amendments. Whether that structural capability translates to meaningful amendment reductions in clinical practice is a question that published evidence has not yet answered.

Regulatory and Documentation Considerations

The September 2025 final FDA guidance on ICH E6(R3) Good Clinical Practice marked a substantive shift in how regulators expect sponsors to think about quality in clinical trial conduct ^[13]. Rather than prescribing specific processes, E6(R3) adopts a quality-by-design framework, requiring sponsors to identify critical-to-quality factors and build oversight proportionate to identified risks. The European Medicines Agency brought Annex 1 of E6(R3) into effect on July 23, 2025 ^[14].

For protocol development specifically, E6(R3)'s emphasis on quality by design means that a sponsor is not simply producing a compliant-looking document. The protocol must be constructed in a way that supports reliable execution, which includes internal consistency, adequate anticipation of operational risks, and documentation that can be reviewed against its source data. When AI tools contribute to protocol drafts, the sponsor retains full accountability for quality.

The EMA's reflection paper on the use of artificial intelligence in the medicinal product lifecycle, adopted by the Committee for Human Medicinal Products (CHMP) and the Committee for Veterinary Medicinal Products (CVMP) in September 2024, emphasizes traceable documentation, human-centric oversight, and GxP-aligned controls for AI and ML use across the medicinal product lifecycle ^[15]. Its framing is consistent with existing GCP principles: AI tools do not transfer accountability from the sponsor or marketing authorisation holder to the system. FDA's January 2025 draft guidance on AI in regulatory decision-making extended this principle to drug and biological product submissions, proposing a seven-step credibility assessment framework requiring sponsors to pre-define the model's context of use, manage risks, and re-evaluate models when updated ^[16].

These accountability requirements do not create a barrier to agentic AI adoption. They define what a production-ready agentic system in a regulated setting must be able to demonstrate: traceable output, documented human review at defined steps, and a change history that can survive regulatory inspection. An agentic system designed from the ground up for regulated document workflows can satisfy these requirements. A standalone copilot that generates output into an untracked document workflow does not.

The ICH M11 Clinical Electronic Structured Harmonised Protocol (CeSHarP) reached final guidance status at FDA on May 22, 2026, with the announcement published in the Federal Register ^[17]. The EMA adopted the CeSHarP template under Step 5 in December 2025, with a coming-into-effect date of June 11, 2026 ^[18]. The guidance creates an internationally harmonized, machine-readable format for clinical trial protocols, covering standardized content structure, required and optional components, and technical specifications for interoperable electronic exchange of protocol data across regulatory authorities ^[17].

Its relevance to the AI tooling question is concrete. A protocol structured according to CeSHarP is organized around defined data elements with specified conformance requirements, not freeform text in an unstructured Word document. An agentic system that can parse a CeSHarP-compliant protocol can check consistency at the element level: does the inclusion criterion in field X match the population definition in field Y, and does the primary endpoint in field Z align with the analysis description in field W? That level of structured access is not available in protocols produced as unstructured documents.

The timeline matters: as of June 11, 2025, organizations building agentic AI infrastructure for protocol development should treat CeSHarP alignment as a near-term design consideration rather than a future one. The guidance is nonbinding, and M11 functions as a harmonized international standard rather than a universal legal mandate, but its adoption trajectory across FDA, EMA, and other ICH member authorities makes it a practical reference point for protocol infrastructure decisions in 2026 and beyond.

FDA's 21 CFR Part 11 establishes requirements for electronic records and electronic signatures maintained or submitted under FDA predicate-rule requirements, including electronic records supporting INDs, NDAs, and BLAs ^[19]. Where an AI system generates or modifies content that becomes part of such records, the surrounding workflow should be designed to support applicable Part 11 controls: system validation, defined access controls, tamper-evident audit trails, and record retention procedures. "Part 11-ready" in practical terms means the workflow can produce, on inspection, a complete log of who directed what, what the system produced, who reviewed it, and what was approved. The question is not whether the AI model itself is validated in isolation, but whether the end-to-end workflow producing the regulated record satisfies those controls.

The practical implication is that the bottleneck for agentic AI deployment in protocol development is not primarily the technology. A TCS analysis from May 2026, examining agentic AI deployment in drug development across the industry, found that most AI work in clinical development has not entered core, inspection-ready workflows ^[20]. Pilots multiply; production deployment does not. The gap is governance architecture: validated systems, defined human review checkpoints, audit trail design, and documentation of AI use that can be produced during a regulatory inspection. Organizations that address the governance design as the primary implementation challenge are more likely to reach production deployment than those that start with model selection.

An Honest View of AI Automation in Protocol Development

Dimension	AI-Assisted Tool (Copilot)	Agentic AI (Pilot Stage)	Governed Agentic Workflow (Production)
Scope	Single task / section	Multi-step workflow	End-to-end document lifecycle
Context persistence	Per-session, per-prompt	Within workflow run	Persistent, logged across sessions
Cross-document consistency	Manual	Partial automation	Automated with human review gates
Change propagation	Manual	Agent-driven, inconsistent	Agent-driven, logged, human-approved
Regulatory compliance	Depends on human process	Not yet governed	Designed to support applicable Part 11 controls where electronic records are in scope
Human oversight	User-directed at every step	Required; partially defined	Defined review points with signoff
Deployment readiness	Widely deployed	Early-stage in pharma	Limited production deployments

As market context, Gartner has projected that by 2028, at least 15% of daily work decisions will be made autonomously through agentic AI, compared to near-zero in 2024 ^[21]. In a clinical research context, the relevant question is which decisions fall within that category. Protocol strategy, endpoint selection, scientific rationale, and IRB-specific considerations are not in scope for autonomous decision-making. Document consistency checking, cross-reference validation, CeSHarP template alignment, and change propagation are candidates for agentic automation under appropriate human oversight.

The DIA's August 2025 analysis is candid about maturity: true agentic behavior in clinical trial workflows remains early-stage ^[8]. The technology capability is largely available; the gap is operating discipline, validation frameworks, and the institutional trust required to move from pilot to production in a GCP-regulated environment ^[20]. These are solvable problems, and organizations working on them now will be better positioned than those that wait for the governance frameworks to arrive pre-packaged. Note that references [9], [12], [20], and [21] are commercial or vendor sources used here for market and workflow context, not as clinical or regulatory evidence.

How Kitsa Fits Into This Problem

KScribe, Kitsa's regulatory document generation product, is built on the operational distinction described in this article. Rather than functioning as a copilot that assists with individual protocol sections in isolation, KScribe is designed as an agentic infrastructure layer for regulatory document generation. According to Kitsa's product materials, it produces and checks protocol content, ICFs, IBs, DSURs, and CSRs through an architecture oriented around cross-document consistency, with version tracking and human review steps across the full submission package. That design is intended to address the cross-document coherence problem that task-level AI tooling leaves to manual review. More at kitsa.ai/protocol-development.

Key Takeaways

•Tufts CSDD data from 2024 (950 protocols, 2,188 amendments across 16 sponsors and CROs) shows that 76% of trials now require at least one amendment, up from 57% in 2015. Each substantial amendment costs between $141,000 and $535,000.
•AI-assisted tools and agentic AI are architecturally distinct. The former responds to discrete task prompts; the latter maintains context across a workflow and can act autonomously between steps. Both categories require human oversight in regulated environments.
•The operational limitation of AI-assisted copilots is not the quality of the underlying model. It is the absence of persistent workflow context and governance infrastructure. Even long-context or RAG-based systems operating as copilots do not propagate changes across a protocol document or generate an audit trail designed to support applicable Part 11 controls.
•There is currently no published clinical evidence directly linking agentic AI protocol tooling to lower amendment rates. The structural argument is sound, but it is prospective. Organizations should not treat amendment reduction as a proven outcome.
•ICH E6(R3) (FDA final guidance September 2025, EMA Annex 1 effective July 23, 2025) reinforces that sponsors remain accountable for all AI-generated content in submissions. Traceability, human review, and audit trails are not optional design choices.
•ICH M11 CeSHarP reached final guidance status at FDA on May 22, 2026, and the EMA template is scheduled to take effect on June 11, 2026. Machine-readable, structured protocols substantially expand what agentic systems can do with protocol content.
•The primary barrier to production deployment of agentic AI in protocol workflows is governance architecture, not technology capability. Validated systems, defined human review points, and audit trails designed to support applicable Part 11 controls are the gap between a pilot and a production-ready workflow.

KScribe · AI Regulatory Document Generation

Built as an agentic infrastructure layer for regulated document workflows, cross-document consistency, version tracking, and human review steps across the full protocol and submission package.

Explore KScribe

FAQ

What is the practical difference between an AI-assisted copilot and an agentic AI system in protocol development?

An AI-assisted copilot responds to user prompts for specific tasks: drafting a section, summarizing a document, or retrieving relevant content. The human user directs every step and manually integrates outputs into the working document. An agentic system can plan a sequence of tasks, maintain context across the full document, act between steps without a new prompt for each one, and coordinate actions across related documents. In regulated environments, both require human review before outputs are used in submissions. The distinction is in what the system does between those review steps.

Can agentic AI reduce protocol amendment rates?

There is no published clinical evidence establishing a direct link between agentic AI protocol tooling and lower amendment rates. The structural argument is that agentic systems, by maintaining cross-document context, are better positioned to catch the category of avoidable amendments that stem from inconsistencies identifiable at the design stage. Tufts CSDD data indicates that roughly one-quarter of amendments are avoidable at initiation ^[1]^,[2]. Whether agentic architecture translates that structural advantage into measured amendment reductions in real trials is an open empirical question. Organizations should treat this as a hypothesis to test, not a proven outcome to claim in procurement decisions.

How does ICH E6(R3) affect the use of AI-generated content in clinical trial protocols?

ICH E6(R3), finalized by the FDA in September 2025 and effective at EMA for Annex 1 on July 23, 2025 ^[13]^,[14], adopts a quality-by-design framework requiring sponsors to identify critical-to-quality factors and build oversight proportionate to those risks. It does not restrict AI use in protocol development, but it reinforces that sponsors remain accountable for all documentation quality. The EMA's September 2024 reflection paper on AI in the medicinal product lifecycle ^[15] emphasizes traceable documentation, human-centric oversight, and GxP-aligned controls for AI and ML use. FDA's January 2025 draft guidance on AI in regulatory submissions ^[16] proposes that AI models used to produce information or data supporting regulatory decision-making should be evaluated through a risk-based credibility framework tied to the model's defined context of use. Agentic systems used in protocol development should be designed to produce auditable outputs and to incorporate defined human review points.

What is ICH M11 CeSHarP and why does it matter for AI-supported protocol workflows?

ICH M11 is the Clinical Electronic Structured Harmonised Protocol template, which reached final FDA guidance status on May 22, 2026 ^[17] and is scheduled to come into effect at EMA on June 11, 2026 ^[18]. It establishes a harmonized, machine-readable format for clinical trial protocols, with standardized data elements, technical specifications, and interoperable exchange formats acceptable across ICH regulatory regions. A protocol structured according to CeSHarP is substantially more accessible to automated consistency checking than a freeform Word document, because the data elements are defined and machine-parseable. Organizations building agentic AI infrastructure for protocol development should treat CeSHarP alignment as a practical design consideration for both the protocol workflow and the underlying AI system, noting that M11 is a harmonized standard rather than a universal legal mandate.

What does "governed agentic workflow" mean in a regulated document context?

A governed agentic workflow is an agentic AI system operating under the validation, audit trail, and human oversight requirements that apply to GxP-relevant electronic records. This means the system has been validated per a defined context of use, generates a complete, tamper-evident log of every automated action, requires qualified human review and approval at defined steps before outputs become part of the official record, and produces documentation adequate for regulatory inspection. This is distinct from an agentic system deployed informally as a productivity tool, where outputs are generated without governance controls. The governance design, not the model capability, is what separates a pilot from a production-ready workflow.

Will agentic AI replace medical writers in protocol development?

No. The use cases where agentic AI adds the most value in protocol development are consistency checking, change propagation, template compliance validation, and structured first-draft generation against defined formats. These are high-volume, attention-dependent tasks. Protocol strategy, endpoint selection, scientific justification, patient population reasoning, and regulatory negotiation require clinical and scientific expertise that is outside the scope of current agentic systems. Medical writers working with governed agentic infrastructure spend less time on structural and cross-reference tasks and more time on the judgment-dependent work that actually requires their expertise.

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." Therapeutic Innovation and Regulatory Science. 2024 May;58(3):539-548. DOI: 10.1007/s43441-024-00622-9. https://pubmed.ncbi.nlm.nih.gov/38438658/
[2] ICON plc. "Controlling Complexity for Regulator-Ready Protocol." ICON Insights Blog. April 9, 2025. Source data: Tufts CSDD Impact Report. Vol. 25; No. 3. May/June 2023. https://www.iconplc.com/insights/blog/2025/04/09/controlling-complexity-regulator-ready-protocol (Commercial source citing Tufts CSDD data)
[3] IQVIA. "Assessing Protocol Complexity and its Impact on Trial Outcomes." IQVIA Blog. January 2026. https://www.iqvia.com/blogs/2026/01/assessing-protocol-complexity-and-its-impact-on-trial-outcomes (Commercial analysis citing Protocol Design Variables study, TIRS 2022)
[4] Getz K, Smith Z, Kravet M. "Protocol Design and Performance Benchmarks by Phase and by Oncology and Rare Disease Subgroups." Therapeutic Innovation and Regulatory Science. 2023;57(1):49-56. DOI: 10.1007/s43441-022-00438-5. https://link.springer.com/article/10.1007/s43441-022-00438-5
[5] Markey N et al. "Clinical Trials Are Becoming More Complex: A Machine Learning Analysis of Data from Over 16,000 Trials." Scientific Reports. 2024;14(1):3514. DOI: 10.1038/s41598-024-53211-z. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10861486/
[6] Gerber DE et al. "Standardizing Eligibility Criteria: A New Approach to Simplifying and Harmonizing Cancer Clinical Trials." JAMA Oncology. 2022 Sep 1;8(9):1333-1339. DOI: 10.1001/jamaoncol.2022.1664. https://pmc.ncbi.nlm.nih.gov/articles/PMC9934063/
[7] Clinical Leader / Bechtel J, Getz K. "Endpoints, Procedures, And Deviations Are On The Rise: Why?" Reporting on findings from the 2025 SCRS Landscape Survey Report, compiled with Tufts CSDD research. January 20, 2026. https://www.clinicalleader.com/doc/endpoints-procedures-and-deviations-are-on-the-rise-why-0001 (Industry media reporting Tufts CSDD/SCRS data)
[8] DIA Global Forum. "Next Steps in Artificial Intelligence: Agentic AI." August 2025. https://globalforum.diaglobal.org/issue/august-2025/next-steps-in-artificial-intelligence-agentic-ai/
[9] Everest Group. "Agentic Artificial Intelligence (AI): A Game Changer in Clinical Trials." December 8, 2025. https://www.everestgrp.com/blog/agentic-artificial-intelligence-ai-a-game-changer-in-clinical-trials.html (Commercial consulting report)
[10] Weerasinghe K, Olawade DB, Teke J, Msiska M, Boussios S. "The Impact of AI-Driven Chatbot Assistance on Protocol Development and Clinical Research Engagement: An Implementation Report." Journal of Personalized Medicine. 2025;15(7):269. DOI: 10.3390/jpm15070269. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12300609/
[11] Babaeipour R, Charest F, Wright M. "AI-assisted Protocol Information Extraction for Improved Accuracy and Efficiency in Clinical Trial Workflows." Banting Health AI. arXiv preprint. February 3, 2026. https://arxiv.org/pdf/2602.00052 (Preprint; not peer-reviewed)
[12] Causaly. "From Copilots to Agentic Workflows: How AI in R&D Changes in 2026." January 6, 2026. https://www.causaly.com/blog/from-copilots-to-agentic-workflows-how-ai-in-r-d-changes-in-2026 (Vendor thought leadership)
[13] U.S. Food and Drug Administration. "E6(R3) Good Clinical Practice." Final guidance. September 2025. Accessed June 11, 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e6r3-good-clinical-practice-gcp
[14] European Medicines Agency. "ICH E6 Good Clinical Practice." Scientific guideline. Principles and Annex 1 effective July 23, 2025. Accessed June 11, 2025. https://www.ema.europa.eu/en/ich-e6-good-clinical-practice-scientific-guideline
[15] European Medicines Agency. "Reflection Paper on the Use of Artificial Intelligence (AI) in the Medicinal Product Lifecycle." EMA/CHMP/CVMP/83833/2023. Adopted by CHMP and CVMP September 2024. Published September 30, 2024. Accessed June 11, 2025. https://www.ema.europa.eu/en/use-artificial-intelligence-ai-medicinal-product-lifecycle-scientific-guideline
[16] U.S. Food and Drug Administration. "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products." Draft guidance. January 2025. Accessed June 11, 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological
[17] U.S. Food and Drug Administration / Federal Register. "M11 Clinical Electronic Structured Harmonised Protocol (CeSHarP); International Council for Harmonisation; Guidance for Industry; Availability." Final guidance. Federal Register. Published May 22, 2026. Accessed June 11, 2025. https://www.federalregister.gov/documents/2026/05/22/2026-10295/m11-clinical-electronic-structured-harmonised-protocol-cesharp-international-council-for
[18] European Medicines Agency. "ICH M11 Clinical Electronic Structured Harmonised Protocol (CeSHarP): Template." Step 5. Final adoption by CHMP December 11, 2025. Scheduled to come into effect June 11, 2026. EMA/CHMP/ICH/778801/2022. Accessed June 11, 2025. https://www.ema.europa.eu/en/documents/template-form/ich-m11-clinical-electronic-structured-harmonised-protocol-cesharp-template-step-5_en.pdf
[19] U.S. Food and Drug Administration. "Part 11, Electronic Records; Electronic Signatures: Scope and Application." Guidance for Industry. August 2003. Accessed June 11, 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/part-11-electronic-records-electronic-signatures-scope-and-application
[20] Tata Consultancy Services. "Making Agentic AI in Drug Development Real and Scalable." TCS Insights Blog. May 2026. https://www.tcs.com/insights/blogs/making-agentic-ai-drug-development-real-scalable (Commercial thought leadership)
[21] Medable. "Building Blocks: Agentic AI Is Transforming Trial Design, Management, and Outcomes." Citing Gartner prediction on agentic AI decision-making by 2028. https://www.medable.com/knowledge-center/guides-building-blocks-agentic-ai-is-transforming-trial-design-management-and-outcomes (Commercial; Gartner prediction cited secondhand)