The Future of AI-Native Clinical Operations

At the 2026 SCOPE Summit, the annual gathering for clinical operations executives, which drew more than 5,000 attendees in Orlando this February ^[23], the conversations about artificial intelligence had shifted register. Event coverage described a field where AI had moved from research question to baseline operational expectation ^[16], with industry commentary suggesting discussions had narrowed to a more specific problem: why do most organizations' AI deployments remain isolated rather than integrated, and what separates sponsors with coherent AI infrastructure from those still running disconnected pilots?

That question has a structural answer. The AI-based clinical trial market reached $9.17 billion in 2025, growing from $7.73 billion in 2024, with analyst projections placing it at $21.79 billion by 2030 at a compound annual growth rate approaching 19% ^[1]. Those figures describe accelerating adoption, though market size reflects technology spend and adoption intent, not validated operational outcomes, which remain a separate and less settled question. What the market numbers do not capture is the gap between organizations deploying AI tools across isolated workflows and those rebuilding their operational architecture around AI as a continuous layer. Closing that gap is what the shift toward AI-native clinical operations actually requires.

This article examines what that architecture looks like in practice, what the evidence says about its effects, and what regulatory and data infrastructure changes are shaping the next several years of clinical development.

Why the Existing Model Has Run Out of Room

The financial and operational pressures on clinical development are not new, but they have grown more acute. A 2024 empirical study by the Tufts Center for the Study of Drug Development, published in Therapeutic Innovation and Regulatory Science, analyzed 409 clinical trial budgets and found that the average direct cost of a Phase III trial day is $55,716 in 2023 USD, nearly 55% higher than the $36,000-per-day figure the industry had used since the prior decade ^[2]. Every week a Phase III site remains unopened costs the sponsoring organization roughly $390,000 in direct operational overhead alone.

Site activation has become a particular constraint. A Tufts CSDD analysis presented at the 2024 SCOPE Summit found that decentralized trial components have increased site burden by more than 25%, and site activation rates in North America have fallen to approximately 62% ^[3]. Contract negotiations average approximately 230 days per site and cost sponsors an estimated $500,000 per day in unrealized drug sales during that period, based on Tufts CSDD data reported in industry literature ^[4]. Protocol amendments compound these pressures further: each amendment costs between $141,000 and $535,000 and typically adds three months to development timelines, also per Tufts CSDD figures ^[4].

At the same time, trials have grown structurally more demanding. Industry analysis suggests clinical operations teams now work with roughly seven times more data points per trial than they managed in prior years ^[19], without a corresponding expansion in operational capacity. Against that backdrop, a Tufts CSDD analysis found that organizations applying AI to clinical trial activity achieved time savings of 18% ^[5]. That number is meaningful, but it reflects AI bolted onto existing workflows, not AI embedded in the operational architecture. The distinction is what the rest of this article addresses.

AI-Native vs. AI-Augmented: A Distinction That Matters

Most clinical organizations that have adopted AI tools have done so in augmentation mode: AI assists at specific points in a workflow while everything else operates as before. Recruitment screening tools flag eligible patients. Predictive models score site feasibility. Natural language processing extracts signals from adverse event reports. Each tool generates incremental gains within its domain.

An AI-native architecture differs in a specific way. Applied Clinical Trials described the divide emerging in the industry in early 2026: organizations building AI fluency into every layer of the clinical process on one side; legacy operators still running isolated pilots on the other, with their long-term competitive position increasingly at risk ^[16]. In an augmented model, AI handles tasks. In an AI-native model, AI structures the operational decision flow, and humans provide oversight, judgment, and regulatory accountability.

Agentic AI, autonomous systems capable of decomposing complex objectives, coordinating across multiple systems, and adapting based on intermediate outcomes, is the technical mechanism making that architecture viable ^[22]. Unlike conventional AI tools, which require human input at each decision point, agentic systems plan and execute across multi-step tasks. A DIA Global Forum analysis in August 2025 described the practical implication for trial operations: an agentic system acting with goal-directed autonomy could continuously monitor trial data in real time, identify underperforming enrollment trends, suggest adjustments, and execute them under predefined rules, rather than waiting for a human to act on a report ^[21].

The published evidence on agentic AI in clinical settings is still early. A widely cited paper in npj Digital Medicine noted that as of 2024, only 86 randomized controlled trials of machine learning interventions had been conducted and published worldwide ^[18], a small number relative to the breadth of AI deployment in research contexts, reflecting the gap between adoption and prospective validation. The available evidence is instructive, but generalizability across different trial types, indications, and institutional settings requires further prospective study.

What the Evidence Shows

A 2025 narrative review published in the International Journal of Medical Informatics, covering studies from January 2015 through December 2024 across PubMed, Embase, IEEE Xplore, and Google Scholar, found AI benefits across the clinical trial lifecycle: patient recruitment tools improved enrollment rates by up to 65%, predictive analytics models achieved 85% accuracy in forecasting trial outcomes, and AI integration was associated with 30 to 50% acceleration in trial timelines and cost reductions up to 40% across the included literature ^[6]. These are ranges drawn from heterogeneous studies across different indications, AI approaches, and data environments, and should not be treated as generalizable benchmarks for any specific program. The review itself describes substantial variation.

More tractable evidence comes from prospective validations in defined institutional contexts. A randomized, blinded controlled trial of nearly 4,500 patients at Mass General Brigham, published in JAMA in 2025 ^[7], tested a retrieval-augmented generation tool called RECTIFIER (RAG-Enabled Clinical Trial Infrastructure for Inclusion Exclusion Review) against manual chart screening for a heart failure trial. AI-assisted screening identified 458 protocol-eligible patients during the study period, compared to 284 by manual review. Of those screened, 35 patients ultimately enrolled through the AI arm versus 19 in the manual arm, a rate of approximately 1.6% versus 0.9%, representing enrollment nearly double that of standard practice, without significant differences across race, gender, or ethnicity ^[7]. A 2024 proof-of-concept study in NEJM AI had also found RECTIFIER more accurate than manual screening at identifying eligible patients for the same heart failure trial, at a cost of approximately two cents per patient screened compared to the substantially higher cost of manual review ^[8].

At Cleveland Clinic, an AI system was deployed in 2024 within a unified EHR covering hospitals and clinics across Ohio and Florida to assess patient eligibility for clinical trials. The system achieved 96% accuracy across both structured and unstructured EHR data, including clinical notes and reports ^[9]. That result derives from a 100-patient physician-review sample with human-in-the-loop validation, important context for interpreting accuracy in broader deployment settings.

A McKinsey analysis of a multi-agent architecture deployed with a large pharmaceutical company for clinical study report generation found that the system autonomously coordinated data extraction, performed analyses, and generated report drafts with minimal user input, reducing drafting errors by 50% and compressing the time from database lock to report finalization from approximately twelve weeks to six ^[17]. The McKinsey report additionally described how contracting agents generating fair-market-value agreements combined with coordinated site outreach could potentially double site activation rates while reducing staffing requirements by 30 to 50% ^[17]. These are projections from a consulting engagement, not outcomes from a peer-reviewed prospective trial, and warrant interpretation accordingly.

Protocol Infrastructure: The Shift from PDF to Structured Data

The clinical protocol is the trial's operational spine. In traditional operations, it exists as a PDF. Amendments require formal regulatory notification, IRB or ethics committee review, site training, and manual updates to the Informed Consent Form, Investigator Brochure, case report form specifications, IRT parameters, and other dependent documents. Each of those cascades consumes time and resources that compound across complex trials. See also our analysis of protocol amendments and downstream document cascades.

ICH M11, the Clinical Electronic Structured Harmonised Protocol (CeSHarP), addresses the foundational data layer beneath that problem. Adopted at Step 4 by the ICH Assembly on November 19, 2025, the guideline establishes a standardized, machine-readable protocol format accepted across FDA, EMA, PMDA, and other ICH regulatory authorities ^[14]. The Technical Specification defines the data elements and technical attributes enabling interoperable electronic exchange of protocol content through open, non-proprietary standards ^[15]. The M11 template will come into effect on June 11, 2026 ^[14], as of this article's publication date, an upcoming milestone that sponsors integrating protocol infrastructure should plan against now.

When protocol content exists as structured data rather than a PDF, certain downstream systems, eCRF configurations, IRT builds, eligibility screening logic, can in principle be populated or updated from the protocol data object rather than through manual transcription. Applied Clinical Trials described this as the trajectory toward "living protocols": dynamic, machine-readable documents capable of continuous simulation, validation, and adaptation as trial data accrues ^[16]. IQVIA, writing in April 2026, framed digital protocols as a single source of truth enabling interoperability across clinical systems and as the foundation for agentic AI workflows ^[20]. For regulatory document teams already managing the cascade between protocol, ICF, IB, DSUR, and CSR, this structural shift has direct implications for how cross-document consistency is maintained, a workflow that tools like KScribe address through automated cross-document dependency tracking.

The important qualification is that M11 does not reduce the regulatory obligations that accompany protocol amendments. IRB/EC review, regulatory notification, site training, and downstream document validation remain mandatory whenever protocol content changes in ways that trigger amendment requirements. What structured data can reduce is the manual work of propagating those changes consistently across clinical systems, a bounded but real improvement over the current state.

Active industry standardization work, including TransCelerate BioPharma's Digital Data Flow initiative, developed in collaboration with CDISC, is working to define the interoperability specifications that will determine how widely M11 is adopted across clinical systems in practice ^[20].

Site Selection and Activation

Site selection has historically relied on investigator relationships, historical enrollment data from prior trials, and site questionnaires of variable quality. The available information is often stale and unevenly distributed across a sponsor's institutional memory.

AI changes the inputs and speed of those decisions. Historical data from ClinicalTrials.gov can be analyzed at scale to assess site-level enrollment performance by therapeutic area and phase. Where sites participate in FHIR-connected health systems, protocol-eligible patient population estimates can be generated against specific eligibility criteria before a site visit occurs. The practical constraint is that FHIR connectivity is not universal, interoperability depends on health system participation in specific data exchange networks, and data quality varies substantially across institutions.

The McKinsey analysis cited above suggests that agentic contracting and outreach agents could double site activation rates and reduce staffing requirements by 30 to 50%, given that the average site contract negotiation currently runs approximately 230 days at a cost of $500,000 per day in unrealized drug sales ^[17]. That figure is significant if achievable in practice, but prospective validation of those projections in clinical development settings has not yet been published.

Patient Recruitment: The Most Developed Evidence Base

Of all the domains in which AI is being applied in clinical operations, patient recruitment has accumulated the most developed evidence base. The reasons are partly structural: enrollment failure is the most common cause of trial extension and abandonment, the cost of underperformance is directly measurable, and recruitment tools operate on defined inputs and outputs that make prospective validation tractable.

The JAMA RECTIFIER trial is the strongest prospective evidence currently published for AI-assisted screening ^[7], and it comes with important context. It was conducted at a single health system in a heart failure indication, using a RAG-based approach on an EHR with relatively mature data infrastructure. The enrollment advantage reflects both the AI tool's performance and the deployment environment. Generalizability across health systems with different EHR architectures, data governance models, and indication profiles will require additional prospective study.

The Cleveland Clinic 96% accuracy result ^[9] adds evidence that AI can assess eligibility with high precision in a well-defined EHR context, and similarly illustrates the specificity of the validation environment.

EHR-based recruitment infrastructure has improved significantly as HL7 FHIR adoption has expanded. FDA's established real-world evidence framework allows EHR-derived data to support regulatory submissions in appropriate contexts ^[24], and growing FHIR interoperability has expanded the technical infrastructure for structured EHR data access. Accessing patient populations across health system networks in practice, however, still requires institution-specific data-sharing agreements, governance arrangements, and technical integration. FHIR adoption improves the underlying standard; it does not automatically eliminate the contracting and operational work that cross-institutional data access requires.

Regulatory and Documentation Framework

The regulatory environment governing AI in clinical operations has developed substantially in the past eighteen months.

FDA issued draft guidance in January 2025 titled "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products," establishing a risk-based credibility assessment framework for AI models intended to generate data or analysis supporting regulatory decisions about drug safety, effectiveness, or quality ^[10]. The guidance specifies that AI outputs must support regulatory decisions without substituting for human judgment. The public comment period closed in April 2025; FDA has not publicly specified a timeline for finalizing the guidance.

EMA adopted its Reflection Paper on the Use of Artificial Intelligence in the Medicinal Product Lifecycle in September 2024, following a public consultation that generated over 1,300 stakeholder comments ^[11]. The paper outlines principles for AI and machine learning at any step of a medicine's lifecycle, from drug discovery through post-authorization, and calls for a risk-based approach to the development, deployment, and performance monitoring of AI tools in regulated contexts. It reflects a convergence with FDA's framing around risk-proportionate governance for AI.

In April 2026, FDA issued a Request for Information seeking public input on a proposed pilot program to assess how AI-enabled technologies can improve efficiency, speed, and decision quality in early-phase clinical trials ^[12]. The comment period for that RFI has been extended to June 29, 2026 ^[12]. FDA also deployed agentic AI capabilities for all agency employees in December 2025, available on a voluntary basis for meeting management, pre-market reviews, post-market surveillance, inspections, and administrative functions, built within a GovCloud environment that does not train on staff inputs or regulatory submissions ^[13]. Legal analysts have noted that FDA has not yet publicly specified validation parameters or governance frameworks for these internal agentic deployments, raising questions about oversight that the agency will need to address as adoption expands ^[13].

For sponsors and CROs deploying AI in clinical operations, the GxP requirements are well-established even if AI-specific guidance is still evolving. AI systems operating in GxP-regulated environments require Computer System Validation consistent with 21 CFR Part 11 and applicable GxP guidelines, including audit trails, access controls, and validation documentation ^[26]. The specific challenge for continuously learning systems is defining what constitutes a model change requiring re-validation, a question FDA's draft guidance addresses conceptually but that operational implementation continues to work through.

Governance: The Infrastructure Behind the Technology

One of the consistent findings across published assessments of AI in clinical development is that operational gains are not a property of the technology alone. They depend on governance structures that most clinical organizations have not yet built.

The governance elements most frequently cited in regulatory guidance and published operational analyses include:

Organizations that build these governance elements alongside technology deployment will find that the gains documented in the literature are more consistently reproducible. Those that treat governance as a compliance obligation to be fulfilled after deployment will find their implementation rates low and their outcomes variable, and will face additional revision work when regulators eventually require retroactive documentation.

How Kitsa Fits Into This Problem

The following section describes Kitsa's product approach.

Kitsa is designed as AI-native clinical research infrastructure rather than a point-solution toolkit. KScribe, Kitsa's regulatory document generation system, automates the production of protocols, ICFs, Investigator Brochures, DSURs, and Clinical Study Reports with cross-document consistency, so that content relationships between documents are maintained as documents evolve. KScreener applies FHIR-based EHR connectivity to patient pre-screening against protocol eligibility criteria. KScout supports site feasibility and selection workflows ^[25]. The architecture is built to operate within the regulatory and data security requirements that clinical operations teams work under. For sponsors and CROs exploring how AI-native document generation fits into their operations, further detail is at kitsa.ai/regulatory-document-generation.

Key Takeaways

• •Analyst estimates place the AI-based clinical trials market at $9.17 billion in 2025 and project growth to $21.79 billion by 2030, driven by structural adoption across sponsors, CROs, and sites rather than early-stage experimentation.
• •Tufts CSDD's 2024 empirical analysis sets Phase III direct trial costs at $55,716 per day in 2023 USD, more than 50% higher than the prior industry benchmark, raising the operational cost of every delay and the financial case for AI that reduces avoidable delays.
• •AI-native operations differs from AI-augmented operations in that AI functions as a continuous operational layer with human oversight embedded throughout, rather than as discrete tools added to existing workflows. Building that architecture requires data infrastructure investment, governance design, and workforce adaptation alongside technology.
• •The most developed prospective evidence for AI in clinical operations comes from patient recruitment: the JAMA MAPS-LLM randomized trial at Mass General Brigham found AI-assisted screening produced enrollment rates of approximately 1.6% versus 0.9% in the manual arm, with AI-screened sites identifying substantially more eligible patients across a study of nearly 4,500 individuals.
• •ICH M11 (CeSHarP), adopted at Step 4 in November 2025 and coming into effect June 11, 2026, establishes the machine-readable, structured protocol standard that will underpin digital protocol infrastructure. It does not change regulatory amendment obligations, IRB/EC review requirements, or site training mandates.
• •FDA's January 2025 draft guidance establishes a risk-based credibility assessment framework for AI in regulatory submissions; EMA's September 2024 Reflection Paper provides parallel European framing. Both require AI to support, not replace, human regulatory judgment.
• •Governance elements, context of use documentation, human oversight, audit trails, change control, and performance monitoring, determine whether AI's documented gains are reproducible in practice. They are as important as the technology itself.

FAQ

References

Related Articles