From Protocol to Predictive Intelligence: The Future of Clinical Trials

Phase III protocols today generate an estimated 3.5 million data points on average across the study lifecycle ^[16]. Since 2005, according to IQVIA's protocol complexity analysis, the average protocol has grown to include 139% more procedures per patient, 214% more endpoints, and 600% more data points collected per study ^[4]. The science driving that growth is real: precision medicine, biomarker stratification, and decentralized designs represent genuine advances in how trials answer clinical questions. But the infrastructure for designing those protocols has not kept pace with the complexity they require.

The consequences are well documented. Tufts CSDD benchmark research published in 2024 found that three out of four protocols required at least one substantial amendment during the study, with a mean of 3.3 substantial amendments implemented per protocol across Phase II and III trials ^[1]. Implementing a single Phase III amendment now takes an average of 260 days from internal identification to final ethics committee approval, a timeline that has nearly tripled over the past decade ^[2]. The median direct cost of each Phase III amendment stands at $535,000 ^[3].

These figures are not peripheral inefficiencies. They represent execution failures that originate, in most cases, at the design stage: eligibility criteria set too narrowly, visit schedules that underestimate patient burden, endpoints that prove difficult to measure at the sites actually enrolled, feasibility assumptions that collapse on contact with real recruitment data.

Addressing those failures requires more than faster authoring tools. It requires a different model for how protocols are developed, tested, and connected to the operational decisions that follow from them.

That was the core argument of Kitsa's webinar with Tech Mahindra, From Protocol to Predictive Intelligence: Reimagining and Accelerating Clinical Trials. The panel brought together Dr. Greg Licholai on clinical strategy, Kristen Ambrosini from AWS on cloud and AI infrastructure, and Rohit Banga, CTO and co-founder of Kitsa, on applied AI for clinical operations. Their discussion traced a path from the protocol as a static document to the protocol as the starting point of an intelligent, connected development system.

Why Protocol Design Has Become a Financial and Scientific Risk Point

A protocol is not only a scientific plan. It is also the operational blueprint for site activation, the patient experience map governing burden and retention, the regulatory artifact governing submissions, and the financial basis for study budgets. When it is miscalibrated in any of those dimensions, the consequences spread through every downstream function.

The current data make that spread measurable. Nearly 90% of Phase II protocols and 82% of Phase III protocols had at least one substantial amendment in the 2023 Tufts CSDD benchmark cohort ^[2]. Changes to eligibility criteria were among the most common drivers, reflecting designs that could not survive contact with the patient populations actually available at the enrolled sites ^[1]. Protocol design flaws and recruitment difficulties ranked among the top reported reasons for amending protocols in the earlier 2015 Tufts CSDD benchmark study ^[2]. A decade later, the rates are higher.

The financial exposure compounds quickly. When a protocol requires multiple amendments, the 260-day implementation timeline does not simply add once; it compounds per amendment, per geography, per affected site ^[2]. Trials with at least one substantial amendment run materially longer from first patient enrolled to database lock, according to Tufts CSDD ^[1].

Phesi's 2024 analysis of global clinical development found that 31% of Phase II trials were terminated during that year, rising from 29% in 2023 and representing roughly a 50% increase over pre-pandemic averages ^[5]. At the same time, Citeline's phase transition analysis for the decade from 2014 to 2023 found that the average likelihood of approval for a new Phase I drug has fallen to 6.7%, an all-time low, with the Phase II transition emerging as the single largest attrition point across all development stages ^[14]. Poor protocol design is not the only driver of those numbers, but it is a measurable contributor that clinical teams can act on before a trial starts.

The Case for Simulation Before Protocol Lock

Before going further, it helps to distinguish three AI use cases this article covers, because they are often conflated. Protocol simulation uses historical trial data and AI analytics to model how design choices will perform before a trial begins. Digital twins generate individualized AI-based predictions of participant outcomes, which can then serve as prognostic covariates in statistical analysis to improve efficiency. AI-assisted regulatory document generation uses specialized agents to draft, structure, and track protocol sections with human review at each stage. These are different applications with different regulatory considerations, different data requirements, and different points of impact in the trial lifecycle.

In aerospace, financial modeling, and pharmaceutical manufacturing, high-stakes systems are modeled and tested against failure scenarios before they are deployed at scale. Clinical trial protocol design has not followed that pattern. Historically, protocols were reviewed by experts drawing on therapeutic knowledge, regulatory precedent, and operational experience. That expertise has been and remains essential. But it has also been the primary, and in many cases the only, mechanism for catching design risks before they become expensive.

AI-powered simulation changes what is available at that stage. Teams can now test different eligibility criteria configurations against databases of historical enrollment data, model how visit schedule changes affect predicted patient burden, compare endpoint structures across analogous trials, and assess site performance patterns against protocol requirements before those requirements are locked. The question shifts from "Does this design look defensible?" to "What does historical evidence suggest about how this design will perform?"

Digital twin methodology is advancing in parallel. Deli Wang et al., publishing in Alzheimer's & Dementia: Translational Research & Clinical Interventions in 2025, evaluated AI-generated digital twins as prognostic covariates in the AWARE Phase 2 Alzheimer's trial: models trained on harmonized data from 6,736 subjects showed positive correlations with observed cognitive outcomes in AWARE and similar validation performance across three independent trials [11]. The authors reported that using digital twins as prognostic covariates could reduce total sample size by approximately 9 to 15% and control-arm sample size by 17 to 26% in future Alzheimer's trials. It is worth noting that this approach uses AI-generated predictions as statistical covariates within a randomized trial, not as a substitute for the randomized control arm itself.

Applied Clinical Trials, reviewing developments through mid-2024, noted that sponsors are increasingly turning to predictive analytics and scenario modeling in trial planning specifically because Phase II termination rates remain persistently elevated ^[10]. The calculus is not complicated: if a third of Phase II trials are terminated, and each substantial Phase III amendment costs a median of $535,000 and 260 days to implement ^[2][3], the return on investment from better upfront design is not marginal. It is structural.

How Protocol Complexity Is Outpacing Manual Review Alone

Dr. Greg Licholai's contribution to the webinar captured something that clinical teams feel in practice before they can always articulate it: the volume and heterogeneity of evidence now required to design a protocol well has grown faster than the review processes designed to evaluate it. He described a reality that clinical teams across the industry recognize: trials are becoming harder to run, not because sponsors are less capable, but because the design decisions made before execution begins are carrying more downstream weight than the tools used to make them were built to handle.

IQVIA's analysis of protocol complexity trends documents how the accumulation of procedures, endpoints, and eligibility requirements has translated directly into harder execution: more opportunities for sites to deviate, more patient burden driving dropout, and more design gaps that surface as amendments after the trial has started ^[4]. Protocol deviations, recruitment shortfalls, site activation delays, and document backlogs all compound that underlying problem.

Human expert review cannot systematically process all of those interdependencies at design time, not because experts lack judgment, but because no human reviewer can hold the historical enrollment outcomes of comparable trials, the operational performance records of candidate sites, and the patient burden implications of each protocol element in view simultaneously. AI can help with exactly that problem. A statistician reviewing a simulation of how a proposed sample size performs under different enrollment assumptions is making a better-informed decision than one working from experience alone. A medical writer who can compare a proposed eligibility criterion against historical enrollment data for similar criteria can identify restrictiveness that standard review would miss.

The role of clinical expertise does not diminish in this model. What changes is the quality of the inputs that expertise is applied to.

Agentic AI and Regulatory Document Generation

Kristen Ambrosini from AWS framed the infrastructure context for the webinar: AI is now touching the full clinical research workflow, from trial design and site selection through patient identification, multimodal data analysis, document generation, and governance. The value of AI in this environment is not in any single task. It is in connection across tasks that currently operate in fragmented silos.

That connection is where Rohit Banga, CTO and co-founder of Kitsa, focused his part of the discussion. During the webinar, he described the design goal directly: clinical research does not need generic AI. It needs clinical workflow-native AI, systems built around how clinical authoring teams actually organize their work. He described how Kitsa's KScribe Design Studio approaches regulatory document generation using specialized AI agents: separate agents for objectives, endpoints, study design, and eligibility criteria, each designed for its specific domain rather than a single general model attempting to author the entire protocol from a blank starting point. This architecture mirrors the way experienced clinical authoring teams actually work. Medical writers, statisticians, regulatory specialists, and clinical operations leads each own distinct sections of a protocol, with review and approval workflows operating across all of them.

This distinction matters in practice. AI that produces text is not the same as AI that understands protocol structure, regulatory terminology, cross-document dependencies, version control requirements, and submission formatting. The second type has to be built specifically for clinical research workflows. It also has to support the collaborative, multi-role review processes that follow authoring: author handoffs, reviewer comments, approver sign-off, audit trail documentation, and version history. Those are not features that can be retrofitted; they need to be part of the design.

This is the difference between AI as a writing shortcut and AI as clinical workflow infrastructure.

Regulatory and Documentation Considerations

Regulatory agencies are not passive observers of these developments. The FDA issued a non-binding draft guidance, "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products," in January 2025, establishing a risk-based credibility assessment framework for AI applications that generate information supporting safety, effectiveness, and quality decisions ^[6]. The draft guidance recommends early FDA engagement for sponsors using AI in ways that will inform regulatory submissions, and describes human oversight throughout the AI model lifecycle as a baseline expectation.

In January 2026, FDA's Center for Drug Evaluation and Research (CDER) and Center for Biologics Evaluation and Research (CBER), working with the EMA, published ten guiding principles of good AI practice in drug development, covering human-centric design, risk-based approaches, data governance, model development practices, lifecycle performance monitoring, and transparent communication of AI limitations ^[8]. Both agencies described these principles as a shared foundation to be supplemented by additional guidance as specific use cases develop.

The EMA took a concrete step in March 2025, when its CHMP issued the first formal Qualification Opinion for an AI tool in clinical research: AIM-NASH, a PathAI system that helps pathologists measure MASH disease activity and fibrosis in liver biopsies ^[13]. The CHMP found that AI-verified readings confirmed by a single expert pathologist produce less inter-rater variability than three-pathologist consensus, demonstrating how one AI-based endpoint methodology can be formally qualified for a defined context of use.

ICH E6(R3), finalized by the FDA in September 2025 and effective for EU trials as of 23 July 2025, introduces Quality by Design and risk-proportionate quality management as core GCP expectations ^[7][12][20]. The guideline calls for sponsors to identify Critical to Quality factors from protocol inception and to apply oversight proportionate to the risk each element represents. E6(R3) does not establish AI-specific rules; its broader expectations around traceability, documented human review, version control, and accountability apply to AI-assisted workflows in the same way they apply to any other approach to generating clinical data or documents. AI-generated content should meet the same standards as human-authored content, with documented evidence of the review and approval process ^[7][12][20].

As Rohit Banga noted during the webinar, Kitsa's approach integrates HIPAA-compliant deployment infrastructure, change tracking, traceability, and multi-role review workflows into the authoring experience from the start. These are not add-ons to the AI capability. They are the conditions under which AI-generated clinical content can be used responsibly in a regulated environment.

AI-Powered Patient Screening and Site Intelligence

Two areas the webinar addressed directly have a growing evidence base worth examining.

On patient screening: a randomized trial of 4,476 patients published in JAMA in 2025 found that enrollment rates using RECTIFIER, an AI-assisted RAG-based screening tool developed at Mass General Brigham, were nearly double those achieved through traditional manual screening (hazard ratio 1.79), with no statistically significant differences in eligibility or enrollment rates across demographic subgroups ^[9]. The trial was conducted within an ongoing heart failure study at Mass General Brigham, and the authors noted that the tool's eligibility criteria can be adapted to other therapeutic areas. That study built on a 2024 proof-of-concept published in NEJM AI, which found that RECTIFIER identified eligible heart failure trial candidates more accurately than trained research coordinators conducting manual chart review, at a screening cost of approximately $0.11 per patient ^[15][19]. Mass General Brigham reported that the tool has since expanded to more than 20 active use cases spanning cardiology, oncology, gastroenterology, neurology, pathology, and psychiatry ^[18].

On site selection: KScout is designed, as Rohit Banga described in the webinar, to analyze prior enrollment patterns, patient density by geography, and investigator experience with specific trial designs, surfacing site candidates a feasibility questionnaire would miss and flagging sites that look viable on paper but carry structural enrollment risks. Separately, site activation timelines are a persistent bottleneck once sites are selected: a 2018 AACI (Association of American Cancer Institutes) benchmarking survey of 61 cancer center members reported a median activation time of 167 days, and more recent Applied Clinical Trials data from 2024 found ranges of 78 to 313 days across NCI-designated cancer centers, depending on protocol complexity and site capacity [17].

Kitsa's KScout is designed around that kind of connected intelligence, matching site selection to the specific patient profile encoded in the protocol rather than evaluating feasibility as a generic site attribute. KScreener applies FHIR-connected EHR data to match individual patient records against trial eligibility criteria before manual chart review is needed.

How Kitsa Fits Into This Problem

Kitsa's approach, as Rohit Banga described in the webinar, is to build AI infrastructure that spans the full trial workflow. KScribe supports regulatory document generation through specialized AI agents and structured human review. KScout connects protocol intent to site selection intelligence. KScreener applies FHIR-based eligibility matching to patient pre-screening. Kitsa's product materials describe HIPAA-compliant, SOC 2 Type II, and ISO 27001-certified infrastructure, with KScribe supporting version control, traceability, collaborative review, approvals, and audit readiness across regulatory documents.

The value is not any individual feature in isolation. It is the connection between design decisions and their operational consequences, visible and manageable from a connected platform rather than scattered across tools that do not share context. When the protocol changes, the eligibility matching updates. When a site is selected, its patient population characteristics inform the screening configuration. When a document is generated, its version history and reviewer annotations are preserved in a format that meets regulatory accountability expectations.

This is what the webinar meant by the shift from protocol to predictive intelligence: not a more powerful document editor, but a connected system where the intelligence that exists at one stage informs every stage that follows.

Key Takeaways

Note: Several statistics below draw on IQVIA, Phesi, and Norstella/Citeline analyses. These are widely cited industry benchmarks rather than peer-reviewed or regulatory evidence and should be read with that context.

• Three out of four protocols required at least one substantial amendment in recent Tufts CSDD benchmarking, with a mean of 3.3 amendments per protocol. Each Phase III amendment carries a median direct cost of $535,000 and takes an average of 260 days to implement ^[1][2][3].
• Since 2005, protocol complexity has grown sharply: 600% more data collected per study, 214% more endpoints, and 139% more procedures per patient ^[4].
• Phase II clinical trial termination reached 31% in 2024, roughly 50% above pre-pandemic averages, according to Phesi's 2024 commercial analytics analysis, with poor design decisions among the contributing factors ^[5].
• AI-generated digital twins have shown validated performance as prognostic covariates in Alzheimer's disease trials, with projected total sample size reductions of 9 to 15% and control-arm reductions of 17 to 26%, and the EMA issued the first formal Qualification Opinion for an AI tool (AIM-NASH, an AI-based histology scoring system for MASH liver biopsies) in March 2025 ^[11][13].
• ICH E6(R3), finalized in 2025, introduces Quality by Design requirements and risk-proportionate oversight with direct implications for how AI-generated protocol content should be documented, reviewed, and approved ^[7][12].
• The FDA's January 2025 non-binding draft guidance and the FDA-EMA joint guiding principles (January 2026) both establish human oversight, context-specific validation, and lifecycle monitoring as baseline expectations for AI in regulatory decision-making ^[6][8].
• A randomized JAMA trial of nearly 4,500 patients found that AI-assisted screening nearly doubled enrollment rates compared to traditional manual methods, demonstrating a measurable operational benefit from connected pre-screening workflows ^[9].

FAQ

References

1. [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 & Regulatory Science, 2024 May;58(3):539-548. DOI: 10.1007/s43441-024-00622-9. PMID: 38438658. https://pubmed.ncbi.nlm.nih.gov/38438658/
2. [2]Getz K. "Shining a Light on the Inefficiencies in Amendment Implementation." Applied Clinical Trials Online, 2023. https://www.appliedclinicaltrialsonline.com/view/shining-a-light-on-the-inefficiencies-in-amendment-implementation
3. [3]Getz KA, Stergiopoulos S, Short M, Surgeon L, Krauss R, Pretorius S, Desmond J, Kaitin KI. "The Impact of Protocol Amendments on Clinical Trial Performance and Cost." Therapeutic Innovation & Regulatory Science, 2016;50(4):436-441. DOI: 10.1177/2168479016632271. https://link.springer.com/article/10.1177/2168479016632271
4. [4]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
5. [5]Phesi. "2024 Most Studied Diseases: Phase II Attrition Analysis." Phesi, January 2025. https://www.phesi.com/news/2024-most-studied-diseases/
6. [6]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, January 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological
7. [7]U.S. Food and Drug Administration. "E6(R3) Good Clinical Practice (GCP)." FDA, September 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e6r3-good-clinical-practice-gcp
8. [8]U.S. Food and Drug Administration (FDA) / European Medicines Agency (EMA). "Guiding Principles of Good AI Practice in Drug Development." FDA/EMA Joint Document, January 2026. https://www.fda.gov/about-fda/artificial-intelligence-drug-development/guiding-principles-good-ai-practice-drug-development
9. [9]Unlu O, et al. "Manual vs AI-Assisted Prescreening for Trial Eligibility Using Large Language Models: A Randomized Clinical Trial." JAMA, 2025. DOI: 10.1001/jama.2024.28047. https://jamanetwork.com/journals/jama/fullarticle/10.1001/jama.2024.28047
10. [10]Markey N, et al. "A New Regulatory Road in Clinical Trials: Digital Twins." Applied Clinical Trials Online, 2024. https://www.appliedclinicaltrialsonline.com/view/new-regulatory-road-clinical-trials-digital-twins
11. [11]Wang D (Deli Wang), et al. "Using AI-Generated Digital Twins to Boost Clinical Trial Efficiency in Alzheimer's Disease." Alzheimer's & Dementia: Translational Research & Clinical Interventions, 2025. DOI: 10.1002/trc2.70181. PMC12639399. https://pmc.ncbi.nlm.nih.gov/articles/PMC12639399/
12. [12]ACRP. "FDA Publishes ICH E6(R3): What It Means for U.S. Clinical Trials." ACRP, September 2025. https://acrpnet.org/2025/09/16/fda-publishes-ich-e6r3-what-it-means-for-u-s-clinical-trials
13. [13]European Medicines Agency (EMA). "EMA Qualifies First Artificial Intelligence Tool to Diagnose Inflammatory Liver Disease (MASH) in Biopsy Samples." EMA News, March 2025. https://www.ema.europa.eu/en/news/ema-qualifies-first-artificial-intelligence-tool-diagnose-inflammatory-liver-disease-mash-biopsy-samples
14. [14]Norstella / Citeline. "Why Are Clinical Development Success Rates Falling?" Norstella, 2024. https://www.norstella.com/why-clinical-development-success-rates-falling/
15. [15]Unlu O, et al. "Retrieval Augmented Generation-Enabled GPT-4 for Clinical Trial Screening." NEJM AI, June 2024. DOI: 10.1056/AIoa2400181. https://ai.nejm.org/doi/abs/10.1056/AIoa2400181
16. [16]IQVIA. "Phase IIb and III Clinical Trial Solutions." IQVIA Product Page, accessed 2026. https://www.iqvia.com/solutions/research-and-development/clinical-trials/phase-iibiii-trials
17. [17]Smith S. "Accelerating Clinical Trial Activation." Applied Clinical Trials Online, June 21, 2024. https://www.appliedclinicaltrialsonline.com/view/accelerating-clinical-trial-activation
18. [18]Mass General Brigham. "AIwithCare: Mass General Brigham Announces New AI Company to Accelerate Clinical Trial Screening and Patient Recruitment." Press Release, December 2025. https://www.massgeneralbrigham.org/en/about/newsroom/press-releases/aiwithcare-mass-general-brigham-spinout-new-company
19. [19]Mass General Brigham. "Artificial Intelligence Accurately Screens Heart Failure Patients for Clinical Trial Eligibility." Mass General Brigham News, June 2024. https://www.massgeneralbrigham.org/en/about/newsroom/articles/ai-screens-heart-failure-patients-for-clinical-trial-eligibility
20. [20]European Medicines Agency (EMA). "ICH E6 Good Clinical Practice: Scientific Guideline." EMA, effective 23 July 2025. https://www.ema.europa.eu/en/ich-e6-good-clinical-practice-scientific-guideline

Related Articles

• How AI Is Transforming Clinical TrialsRead article
• 11 Early-Phase Trials and Design Frameworks to Reduce Design RiskRead article
• 4 AI Misconceptions in Clinical TrialsRead article

Kitsa is an AI-native clinical research infrastructure company. Its products, KScribe, KScout, and KScreener, are designed to support regulatory document generation, site selection intelligence, and patient pre-screening workflows across the clinical trial lifecycle. Learn more at kitsa.ai.