AI-Native Clinical Operations: The Next Operating Model

Clinical research has accumulated a decade of AI tools without yet building an AI organization. The distinction matters. A trial run on point-solution AI still functions like a 2005 trial: documents get written by humans and spot-checked by AI, sites get selected by judgment committees with AI flags, patients get screened by coordinators who receive AI-generated suggestions. These are meaningful improvements. They are not an operating model change.

The industry is now crossing a different threshold. The IQVIA Institute's 2025 Global R&D Trends report documented that inter-trial delays, which peaked at 32 months in 2022, had fallen to 17 months by 2024, with enrollment durations stabilizing for the first time after a multi-year climb ^[1]. That is a measurable shift. But the IQVIA Institute's 2026 follow-up report noted that in 2025 overall development timelines grew longer again, driven by enrollment and inter-trial pressures that point-solution adoption alone has not resolved ^[2]. The efficiency ceiling of layered AI tools is becoming visible. AI-native clinical operations is the argument that there is a higher ceiling, and that reaching it requires rebuilding workflows rather than retrofitting them.

This article examines what AI-native clinical operations means in practice, where the evidence currently sits, and what sponsors, CROs, and sites should understand about the regulatory and organizational conditions that make this transition possible.

Why the Patchwork Model Has a Ceiling

The dominant AI adoption pattern in clinical research is incremental: a recruitment tool added to an existing enrollment workflow, a risk-based monitoring dashboard sitting above a manual data review process, a generative AI assistant that drafts CSR sections for a medical writer to revise. Each tool addresses a specific friction point. The aggregate, however, does not add up to systemic efficiency.

Advarra's 2024 Site-Sponsor/CRO Collaboration survey found that only 29% of site respondents agreed that sponsor and CRO technology solutions delivered on their promised value for integration and efficiency ^[14]. The stated reason was consistent: technologies address isolated pain points rather than the broader operational fabric that sites navigate daily. More tools, in this framing, can increase rather than reduce total workload.

The Tufts Center for the Study of Drug Development quantified this in a 2024-2025 global assessment of 302 respondents from 79 sponsors and CRO organizations. As of late 2024, only 11% of responding companies had fully implemented AI and machine learning to support clinical trial activities, with an additional 22% reporting partial implementation ^[3]. Two-thirds of respondents indicated low confidence in the accuracy and quality of the data underlying their AI solutions. Intellectual property concerns and data governance gaps were the top adoption barriers ^[3]. These are not small problems. They are structural constraints that point-solution deployment, by definition, cannot fix: each new tool introduces its own data environment, its own trust question, its own governance gap.

The clinical trial startup picture reinforces this. An ICON survey conducted in June 2025 among more than 100 principal investigators and senior site personnel found that 55% of respondents reported time from site selection to full activation exceeding five months, with 39% reporting worse timelines than two years prior ^[7]. Contract and budget delays affected 66% of respondents frequently ^[7]. The NCI recommends a 90-day activation target for cancer center trials; 2024 data from NCI-designated centers shows actual activation times ranging from 78 to 313 days ^[9]. According to GlobalData, the share of clinical trials experiencing delayed start dates grew from 4.5% in 2003 to 21.8% in 2024 ^[8]. According to a Tufts CSDD analysis published in 2024, the mean direct cost to conduct a Phase III clinical trial is $55,716 per day in 2023 USD, based on 409 protocols conducted between 2016 and 2021 ^[19]. If site activation delays extend the overall Phase III trial timeline by two months, the direct trial-conduct cost exposure approaches $3.3 million before any downstream enrollment or opportunity costs. Not every delayed site extends the program timeline by two months, but across a multicenter program where multiple sites fall behind simultaneously, the cumulative exposure is computable and large.

None of these delays exist because sponsors lack AI tools. They exist because the operational infrastructure surrounding site activation, protocol design, document generation, and monitoring is still organized around a fundamentally manual model, with AI dropped in at individual steps.

What AI-Native Actually Means

The term gets used loosely. For the purposes of this discussion, AI-native clinical operations has a specific meaning: the operating model in which AI is not a layer added to existing workflows but the backbone through which workflows are designed, executed, and monitored. In a point-solution model, a human runs the process and AI provides inputs. In an AI-native model, AI executes the core workflow steps, with humans providing oversight, judgment on non-routine decisions, and final sign-off on consequential actions.

The distinction is not cosmetic. It changes what functions require dedicated headcount, how data flows between trial functions, what quality checkpoints look like, and how organizations can scale across multiple concurrent programs.

Three functional domains define the AI-native transition in clinical operations. The table below summarizes how each model layer differs in practice.

Protocol and Documentation Intelligence

In the current model, a clinical protocol is authored by a team over several months, incorporating inputs from clinical, regulatory, biostatistics, and medical affairs stakeholders. Protocol quality failures are expensive: a Tufts CSDD analysis found that 57% of clinical trial protocols include at least one substantial amendment during conduct, with 45% of those amendments classified as avoidable. The median direct cost of a Phase III protocol amendment is $535,000 ^[16]. An avoidable amendment at that rate is a half-million-dollar design failure.

AI-native protocol development changes the authoring process rather than reviewing the output. Systems trained on historical protocol libraries, outcome data, regulatory precedent, and site feasibility signals can surface design risks before a protocol is finalized, flagging eligibility criteria that historically underperform enrollment targets, inclusion/exclusion combinations that generate amendment patterns, or endpoint selections that have created monitoring and data management complexity in comparable programs. Generative AI then supports document drafting across the full regulatory package: protocol, investigator's brochure, informed consent forms, clinical study reports, and safety documents, all generated with consistency enforcement across documents that have historically been authored in organizational silos.

The Tufts CSDD assessment reported that AI tools supporting regulatory document preparation delivered a mean 63% time reduction in gathering, organizing, and compiling submission documentation, the highest time savings of any use case in the study ^[3]. AI systems supporting clinical study report writing produced approximately 10% time savings off typical durations ^[3]. These are early-adoption figures from a population in which only 11% of companies have fully implemented AI for trial activities; the ceiling is not visible yet. For a more detailed treatment of how protocol amendments cascade across regulatory documents and how AI-native document infrastructure reduces their downstream impact, see Kitsa's analysis of protocol amendments and downstream document cascades.

Site Intelligence and Activation

Site selection in the traditional model is a combination of CRA judgment, historical site performance data, geography, and therapeutic area expertise. The process is labor-intensive and structurally biased toward known relationships. Sites that lack an established track record with a sponsor or CRO are systematically underused, even when their patient population and investigator capabilities match the protocol well.

AI-native site intelligence replaces the manual feasibility survey and relationship-driven shortlist with continuous, signal-based site profiling. Historical trial performance, publication activity, IRB turnaround averages, coordinator staffing levels, therapeutic area specialization, and local patient population characteristics become inputs to a ranked selection model rather than elements in a spreadsheet reviewed by a startup coordinator. IQVIA's 2025 life sciences AI buildout with NVIDIA specifically targeted investigative site selection as one of the highest-investment AI functions, and the Tufts CSDD data confirms this: average AI/ML investment for site selection functions averaged $1.9 million per implementation, the third-highest of any clinical trial AI use case examined ^[3].

Beyond selection, AI-native startup processes address the activation bottleneck directly. The ICON June 2025 survey found that 66% of respondents experienced contract and budget delays frequently, with 92% identifying those areas as where sponsors and CROs most need to improve ^[7]. WCG's 2024 Clinical Research Site Challenges Report, based on a survey of more than 850 clinical research sites worldwide, found that contract and budget issues are the single largest obstacle to study startup, with approximately three-quarters of surveyed sites citing them as a primary barrier ^[11],[15]. Actual median activation times run 140 to 167 days against an NCI benchmark target of 90 days and an industry target of 90 to 120 days ^[9]. ICON's Executive Vice President Brian Mallon called for automation of repetitive administrative tasks and standardization of clinical trial agreements as the highest-priority near-term interventions ^[18]. In an AI-native model, contract language review, budget gap analysis, and IRB document preparation are not manual tasks with AI assists; they are AI-managed workflows with human escalation pathways for non-standard situations. For a deeper analysis of how activation bottlenecks propagate delay across clinical programs, see Kitsa's analysis of protocol delay propagation across clinical operations.

Patient Matching and Dynamic Monitoring

Enrollment failure is the single largest contributor to trial delay. An AI-native patient matching infrastructure connects eligibility criteria directly to structured and unstructured EHR data through FHIR APIs, enabling continuous pre-screening across patient populations rather than episodic chart review. A 2022 study published in the European Heart Journal - Digital Health described a FHIR R4-based recruitment support system deployed in a cardiology department that enabled on-demand screening of all hospitalized and ambulatory patients against trial eligibility criteria, replacing manual coordinator review for that task ^[20]. A comprehensive review published in ScienceDirect in October 2025 across PubMed, Embase, IEEE Xplore, and Google Scholar from 2015 to 2024 found that AI patient recruitment tools improved enrollment rates by 65% in reviewed studies ^[10].

The Cleveland Clinic deployed an AI-based eligibility screening system across its unified EHR in August 2024. Published results show 96.2% accuracy in determining trial eligibility from structured and unstructured EHR data, with investigators noting substantially faster identification compared to manual chart review ^[12]. A study presented at the 2025 American Society of Hematology Annual Meeting reported, in conference-stage findings, that an LLM-based system embedded in the EMR identified seven times more eligible patients for a polycythemia vera trial than standard workflows, achieving 100% positive predictive value after research staff verification ^[22]. The operational implication: enrollment bottlenecks caused by manual identification of candidates are addressable with current technology. The constraint is data infrastructure and implementation, not AI capability. KScreener, Kitsa's FHIR-based patient matching product, applies this approach within Kitsa's clinical research network.

On the monitoring side, the same comprehensive review found that AI digital biomarkers enabled continuous adverse event detection with 90% sensitivity ^[10]. The Tufts CSDD data reported that AI-enabled patient monitoring delivered a 75% mean time reduction across assessed use cases, the highest of any clinical trial execution domain ^[3].

The Agentic Layer: From Analysis to Action

Point-solution AI analyzes and recommends. Agentic AI acts. The distinction is significant enough to warrant its own section, because it is the mechanism through which AI-native operations differs most sharply from AI-augmented operations.

Agentic AI systems decompose complex multi-step tasks into sequences of actions, coordinate across data sources and operational systems, adapt based on intermediate results, and maintain an audit trail of reasoning. In clinical research, agentic workflows are already operating in trial master file management, informed consent generation, data quality review, and site communication coordination. IQVIA's pharmaphorum case report from January 2026 describes agentic AI managing eTMF workflows through trial closeout, with autonomous handling of document indexing, inspection-readiness checks, and quality control processes ^[13].

The pace of institutional adoption is accelerating across the life sciences sector. On December 1, 2025, the FDA deployed agentic AI capabilities for all agency employees, enabling complex multi-step workflows for pre-market reviews, post-market surveillance, inspections, and administrative tasks ^[23]. The deployment is voluntary and built within a GovCloud environment in which models do not train on regulated industry data ^[23]. IQVIA has active agentic workflow programs across eTMF and data quality functions ^[13], and Fortrea unveiled its Fortrea Intelligent Technology (FIT) suite in April 2026, an AI-enhanced platform designed to automate workflows and deliver near-real-time clinical trial insights across sponsor and site teams ^[24]. For clinical operations teams, the relevant observation is not that AI agents are a future capability but that the organizations that process the most trial data at scale are actively investing in and deploying them.

The practical application in clinical operations is not a general AI assistant but specialized agents with defined roles and data access. A site activation agent monitors contract status, flags outstanding items, drafts response correspondence, and surfaces regulatory submission completeness gaps. A data review agent continuously monitors incoming trial data against protocol specifications, prioritizes queries by clinical significance, and escalates safety signals on a defined schedule. A document consistency agent cross-references the protocol, ICF, IB, and CSR to surface definitional discrepancies before they reach a regulatory reviewer. None of these agents requires a human to initiate each task. Each maintains a reasoning trace that satisfies audit requirements.

This is what distinguishes an AI-native operating model from an AI-augmented one: the organization is not running trials with AI tools; it is running trials through AI infrastructure with human governance.

Regulatory Conditions That Enable the Transition

The regulatory environment has moved materially toward enabling AI-native operations over the past 18 months. Two developments are foundational.

ICH E6(R3), adopted in final form in January 2025 and incorporated into FDA guidance in September 2025, replaced the prescriptive site-visit monitoring model of E6(R2) with a risk-based, proportionate quality management framework ^[5],[6]. The EU/EMA implemented Principles and Annex 1 effective July 23, 2025 ^[21]. E6(R3) explicitly provides for innovations in trial design, technology, and operational approaches. It requires sponsors to proactively design quality into trials and to identify factors critical to trial quality rather than inspect for compliance after the fact. This is directionally aligned with how AI-native operating models are designed: quality built earlier into workflow execution, rather than applied primarily as retrospective inspection ^[5].

The FDA's January 2025 draft guidance on the use of AI to support regulatory decision-making for drug and biological products introduced a risk-based credibility assessment framework covering AI models used across the clinical, nonclinical, post-marketing, and manufacturing product lifecycle phases ^[4]. According to the FDA, the guidance was informed in part by CDER's experience reviewing more than 500 drug and biological product submissions containing AI components since 2016, a volume that the agency characterized as reflecting exponential growth in AI use in regulatory submissions ^[17]. The framework is non-binding in draft form but establishes a clear interpretive standard: AI models used in regulatory contexts require documented validation, defined context of use, and an evaluation process proportionate to the decision risk ^[4].

These two frameworks together have materially clarified the regulatory conditions under which AI can support trial conduct under GCP. E6(R3)'s risk-based quality management principles and the FDA's AI credibility framework do not mandate AI-native operations, but they create a workable path: AI tools that are properly validated, that produce auditable reasoning, and that operate within a defined context of use are clearly within scope of both frameworks. The remaining question is whether sponsors and CROs have the data infrastructure, governance models, and organizational design to execute within those conditions. It is worth separating two distinct questions here: whether the regulatory framework permits AI-supported trial workflows (yes, with validation and defined context of use), and whether any specific AI-native implementation has been formally reviewed or accepted by a regulatory authority (that determination is implementation-specific, not pre-approved by E6(R3) or the draft AI guidance).

The Implementation Gap and the Trust Problem

A growing body of empirical evidence suggests measurable AI efficiency gains across multiple clinical trial domains. But the Tufts CSDD data tells a more complex story about where the industry actually sits. Only 11% of responding companies had fully implemented AI for clinical trial activities as of late 2024 ^[3]. Two-thirds reported low confidence in the accuracy and quality of the data underlying their AI tools ^[3]. These are not technology problems. They are data quality, governance, and organizational trust problems.

AI-native clinical operations requires a data substrate that most organizations have not yet built. Sponsor trial data sits in multiple systems, often with inconsistent terminology and limited interoperability. EHR data available for patient matching varies in structure, completeness, and update frequency across health systems. Site performance history is often stored in proprietary CRO databases that sponsors cannot access. Protocol libraries are documents in a file system rather than structured, machine-readable assets.

None of this is insurmountable. It is, however, a meaningful prerequisite. Organizations that have invested in federated data models, standardized terminology (CDISC, OMOP, HL7 FHIR), and governed AI validation workflows are materially closer to AI-native operations than those that have purchased AI tools on top of fragmented data. The Tufts CSDD data suggests the current investment level reflects this: average AI/ML implementation investment for clinical trial activities is $1.1 million, with execution-stage activities requiring three to four times more investment than planning and design activities ^[3]. Data quality and data cleaning functions attracted the highest average AI investment of any use case, at $3.2 million ^[3]. Organizations spending the most on AI data quality are the ones building the substrate that makes everything else possible.

The second barrier is organizational. AI-native clinical operations shifts the clinical research professional's role from doing the work to governing the system that does it. That is not a minor job description update. It requires new competency frameworks, new accountability structures, and organizational leaders who understand both the capabilities and the limitations of the AI systems they are overseeing. The trust problem documented in the Tufts CSDD assessment (two-thirds of organizations reporting low confidence in AI data quality) is partly a technical problem and partly a governance design problem. Organizations that have implemented AI with clear validation standards, human review protocols, and documented escalation pathways report meaningfully higher adoption rates than those that deployed tools without governance architecture ^[3].

Governance, Validation, and the Limits of Current Evidence

AI-native operations does not reduce the obligation for rigorous human oversight; it reorganizes where that oversight is applied. Several important limitations bear stating explicitly. First, most published evidence for AI efficiency gains in clinical trial operations comes from single-institution pilots, vendor-sponsored case studies, or narrative reviews rather than randomized or controlled comparisons across programs. The Tufts CSDD assessment's 18% mean cycle time reduction, while significant, aggregates across a range of use cases and organizations at varying stages of implementation maturity ^[3]. Second, AI models used in regulatory contexts require documented validation, defined context of use, and credibility assessments proportionate to the decision risk, per the FDA's January 2025 draft guidance ^[4]. Organizations adopting AI-native operations should expect this documentation to be an ongoing requirement, not a one-time exercise. Third, data quality issues, which two-thirds of Tufts CSDD respondents cited as a top concern, can propagate rather than disappear in AI-native models; AI systems that process poor-quality input data may produce confident-sounding errors faster than a human reviewer would ^[3]. Clinical operations teams building toward AI-native infrastructure should treat data quality investment as a prerequisite, not a parallel workstream. Fourth, vendor selection and AI model qualification under GCP require the same rigorous evaluation applied to any computerized system under 21 CFR Part 11 and its EU counterparts. The presence of an AI system does not waive supplier qualification obligations; it adds a layer of model-specific validation to an existing compliance framework.

How Kitsa Approaches AI-Native Infrastructure

For sponsors and CROs building toward an AI-native operating model, the practical challenge is connecting AI capability to the regulatory document and site network workflows that trials actually run on. KScribe, Kitsa's AI-powered regulatory document generation platform, addresses the document backbone of clinical operations: drafting and maintaining cross-document consistency across protocols, ICFs, investigator's brochures, DSURs, and CSRs. Kitsa's infrastructure includes SOC 2 and ISO 27001 certifications and HIPAA-aligned controls ^[25], as detailed at kitsa.ai/regulatory-document-generation.

A Practical Starting Point for Sponsors and CROs

Moving toward an AI-native operating model does not require a single large transformation program. Organizations that have made the most progress have generally worked through five stages, in roughly this order.

Key Takeaways

• •The gap between AI tool adoption and AI-native operations is organizational, not technological. Most sponsors and CROs have added AI to existing workflows; AI-native operations means redesigning workflows around AI execution with human governance.
• •The early efficiency data is increasingly concrete: AI-enabled patient monitoring delivered a 75% mean time reduction in Tufts CSDD's 2025 assessment; regulatory document compilation delivered 63% time savings; AI patient recruitment tools improved enrollment rates by 65% in a comprehensive October 2025 peer-reviewed review. Most of this evidence comes from pilots or narrative reviews rather than controlled comparisons, and results will vary by organizational maturity and data quality.
• •Site activation remains the most visible operational failure point. A June 2025 ICON survey found 55% of sites reporting more than five months from site selection to full activation. With Phase III trial costs running at $55,716 per day, the cost of slow startup is computable and large.
• •ICH E6(R3) (effective in the EU July 2025; FDA final guidance September 2025) and the FDA's January 2025 draft guidance on AI in regulatory decision-making clarify the regulatory conditions for validated AI-supported workflows. Both reward quality-by-design and risk-proportionate oversight, principles that AI-native models are designed to put into practice from the start, though each specific AI implementation still requires its own validation, credibility documentation, and vendor qualification under GCP.
• •Agentic AI has moved from concept to active deployment. IQVIA and Fortrea are deploying AI-enabled operational platforms designed to automate clinical trial workflows ^[13],[24], and the FDA explicitly deployed agentic AI capabilities for all agency employees in December 2025 ^[23]. Organizations that have not begun evaluation should treat this as a priority-tier decision, not a watch-and-see item.
• •The primary barrier to AI-native adoption is not AI capability. It is data quality, governance architecture, and organizational readiness. The Tufts CSDD data makes this explicit: two-thirds of organizations report low confidence in their AI data quality.
• •AI-native clinical operations does not remove human expertise from the trial process. It relocates it: from doing tasks to governing systems, setting quality standards, resolving exceptions, and bearing scientific accountability for outcomes.

Frequently Asked Questions

References

Related Articles