26  AI Agents in Clinical Development

Agents can draft text. Whether one runs inside a trial turns on something else: can it operate within a regulated environment at all, preserving data integrity, producing audit-ready traces, and remaining under accountable human oversight? The 2024 EMA reflection paper on AI and risk-based approaches for AI use in drug development formalizes that expectation (Unlearn.AI 2025; European Medicines Agency 2024; National Institute of Standards and Technology 2023; Amershi et al. 2019).

Clinical trials run on a layered digital infrastructure: the core transactional systems the previous chapter described (EDC, CTMS, eTMF, RTSM), integration middleware, and the analytics and automation services that continuously convert operational events into risk signals and decisions. Cloud adoption and regulatory acceptance of technology-enabled trial conduct, including decentralized elements and digital health technologies for remote data acquisition, have accelerated the shift toward that model (U.S. Food and Drug Administration 2024, 2023c). This chapter takes up that top layer, along with the methodological approaches changing how trials are designed and how evidence is generated.

The transition is conceptual as much as technological: from automation as a feature (dashboards, rules engines, isolated machine-learning models) to automation as an operating model, in which software agents can plan, execute, and verify multi-step workflows across the clinical operations stack. This shifts the trial from a linear “pipeline” towards a dynamic, electronic map of activities, where information infrastructure identifies bottlenecks and navigates the “diverse web of failure-prone iterative learning loops” that characterize modern drug development (Wagner et al. 2018). “Agentic AI” here is a design pattern: “compound” systems that combine models with retrieval, tools, and control logic, grounded in computer science work on “generative agents” that simulate human-like behavior and goal-directed action (Park et al. 2023; Berkeley Artificial Intelligence Research 2024). Preliminary findings from 2024-2025 benchmarks suggest agentic systems can improve clinical task performance by up to 60 percentage points over base language models in specific tasks such as evidence retrieval and task planning, though most improvements are more modest and results vary considerably by task and study design (Abou Ali, Dornaika, and Charafeddine 2026). The same shift converges with Digital Twin “Moonshots”, which aim to integrate personalized digital twins directly into medical records to optimize both clinical care and trial participation (Duke Center for Virtual Imaging Trials 2024).

The money has moved accordingly. The clinical trial software market was estimated at about USD 11 billion in 2024 and is projected to grow at a compound annual rate of about 11% through 2033 (Grand View Research 2025), with AI applications growing rapidly within it (Fortune Business Insights 2024). None of that is evidence of clinical benefit on its own, but it explains why sponsors, CROs, and platform vendors are reorganizing workflows around automation and AI-mediated operations.

26.1 Agentic AI in Clinical Workflows

Pull the right protocol amendment from an eTMF. Verify whether a site has an updated IRB approval. Check that a safety narrative is consistent with the underlying case data. Each of these is a retrieval and verification problem under audit constraints, not a “creative writing” task, and none of them is settled by calling a model once.

Building Blocks and System Architecture

The recent resurgence of “agents” is a systems shift. Agentic systems orchestrate multiple calls, state, and tools to complete a task, decomposing goals, retrieving context, taking actions, and iterating on feedback (Cheng et al. 2024; Wang et al. 2024). On the Berkeley AI Research account of “compound AI systems,” reliability gains often come from engineered compositions (retrievers, checkers, constrained tool calls, repeated sampling) rather than from a single monolithic model invocation (Berkeley Artificial Intelligence Research 2024).

Three technical ideas recur. Structured intermediate reasoning, such as chain-of-thought prompting, can improve performance on complex tasks, though it does not guarantee correctness in the presence of missing or stale information (Wei et al. 2022). Retrieval-augmented generation (RAG) externalizes “memory” into a maintained knowledge base and, when paired with citations to retrieved sources, can supply provenance (Lewis et al. 2020; Douze et al. 2024). Coupling reasoning with tool use, where a model decides when to query, calculate, or fetch, also yields more controllable trajectories than purely generative text (Yao et al. 2023).

Agents also need consistent interfaces to external systems, and tool calling standards such as the Model Context Protocol (MCP) aim to standardize how AI applications connect to data sources and actions, turning “integration” into a first-class design surface (Anthropic 2025). In regulated contexts the payoff is not novelty; it is enforceable access controls, a log of every tool invocation, and agent actions a reviewer can follow.

Clinical workflows are naturally multi-role (data management, monitoring, safety, regulatory, sites), and multi-agent frameworks formalize that by allocating roles and coordinating conversations among specialized agents (Wu et al. 2023). Empirical analyses, however, show that such systems fail in predictable ways: specification gaps, coordination failures, and weak verification loops (Cemri et al. 2025). “Agent quality” therefore has to be measured on cost, robustness, and reproducibility alongside task accuracy (Kapoor et al. 2024), which is where evaluation is heading: benchmarks now emphasize realistic multi-step work with tool use and resource constraints, and propose clearer reporting of cost and reproducibility (Chan et al. 2024; Cappello et al. 2025). Human-in-the-loop debugging and steering tools are emerging so that multi-step traces can be inspected and edited (Epperson et al. 2025).

Run agents concurrently across several workflows and the operational concerns become first-order: context windows turn into a managed resource, tools must be scheduled and access-controlled, and traces must be stored in an auditable way. Emerging “agent runtime” proposals make this explicit by separating agent applications from shared services such as scheduling, context management, and access control (Mei et al. 2024). Pipeline frameworks that treat agent workflows as composable graphs add systematic optimization and regression testing across versions, which matters in any environment where changes must be validated rather than “shipped and hoped” (Khattab et al. 2023).

One risk remains whatever the architecture. Agentic systems can create an illusion of competence, in which fluent outputs are mistaken for validated decisions: a known socio-technical risk in scientific work, where productivity gains may coexist with a decline in genuine understanding and critical scrutiny (Messeri and Crockett 2024). Human-AI interaction guidelines answer it by making uncertainty visible, supporting oversight and correction, and ensuring users understand system limitations, principles that map directly to quality management and inspection readiness (Amershi et al. 2019).

26.2 Regulatory Framework for AI in Clinical Trials

Most of what FDA and EMA have said about AI in drug development is guidance: influential, expected of sponsors, and not legally binding. One AI-specific instrument in this section is law, and it is not a medicines regulation at all. The validation and documentation requirements for computerized systems (21 CFR Part 11, EMA’s computerized-systems guideline, Annex 11, ICH E6) are covered in Chapter 25.

The 10 Guiding Principles for Good AI Practice

In January 2026, FDA (CDER and CBER) and EMA jointly published Guiding Principles of Good AI Practice in Drug Development, the first joint regulatory statement establishing foundational expectations for AI across the drug product lifecycle (U.S. Food and Drug Administration and European Medicines Agency 2026). The ten principles represent international consensus on what “good practice” means here.

NoteJoint FDA-EMA Guiding Principles (January 2026)

The following is quoted verbatim from the joint FDA-EMA document:

  1. Human-centric by design The development and use of AI technologies align with ethical and human-centric values.
  2. Risk-based approach The development and use of AI technologies follow a risk-based approach with proportionate validation, risk mitigation, and oversight based on the context of use and determined model risk.
  3. Adherence to standards AI technologies adhere to relevant legal, ethical, technical, scientific, cybersecurity, and regulatory standards, including Good Practices (GxP).
  4. Clear context of use AI technologies have a well-defined context of use (role and scope for why it is being used).
  5. Multidisciplinary expertise Multidisciplinary expertise covering both the AI technology and its context of use are integrated throughout the technology’s life cycle.
  6. Data governance and documentation Data source provenance, processing steps, and analytical decisions are documented in a detailed, traceable, and verifiable manner, in line with GxP requirements. Appropriate governance, including privacy and protection for sensitive data, is maintained throughout the technology’s life cycle.
  7. Model design and development practices The development of AI technologies follows best practices in model and system design and software engineering and leverages data that is fit-for-use, considering interpretability, explainability, and predictive performance. Good model and system development promotes transparency, reliability, generalizability, and robustness for AI technologies contributing to patient safety.
  8. Risk-based performance assessment Risk-based performance assessments evaluate the complete system including human-AI interactions, using fit-for-use data and metrics appropriate for the intended context of use, supported by validation of predictive performance through appropriately designed testing and evaluation methods.
  9. Life cycle management Risk-based quality management systems are implemented throughout the AI technologies’ life cycles, including to support capturing, assessing, and addressing issues. The AI technologies undergo scheduled monitoring and periodic re-evaluation to ensure adequate performance (e.g., to address data drift).
  10. Clear, essential information Plain language is used to present clear, accessible, and contextually relevant information to the intended audience, including users and patients, regarding the AI technology’s context of use, performance, limitations, underlying data, updates, and interpretability or explainability.

All ten operationalize one expectation: AI in drug development must be validated, documented, and maintained with the same rigor as any other regulated activity, plus extra attention to what makes AI different, its data-dependency, its potential opacity, and its tendency to drift over time.

The FDA Risk-Based Credibility Framework

FDA’s January 2025 draft guidance “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products” (U.S. Food and Drug Administration 2025) sets out a seven-step risk-based framework and, though still a draft, represents FDA’s expected approach for evaluating AI-generated evidence.

ImportantScope of FDA AI Guidance

The FDA guidance applies when AI is used to produce information or data intended to support regulatory decision-making regarding safety, effectiveness, or quality. It does not apply to AI used in drug discovery (the in-scope drug product life cycle begins at nonclinical development, so not all pre-IND work is excluded), nor to AI used purely for operational efficiencies such as internal workflows and resource allocation that do not impact patient safety, drug quality, or study reliability.

The Seven-Step Credibility Assessment Process:

  1. Define the Question of Interest: What specific question, decision, or concern is the AI model addressing?
  2. Define the Context of Use (COU): What is the specific role and scope of the AI model? Will other evidence be used alongside it?
  3. Assess Model Risk: Combine model influence (contribution of AI evidence relative to other evidence) with decision consequence (significance of adverse outcomes from incorrect decisions). See Figure 26.1 for how common AI applications map onto this two-dimensional risk space.
  4. Develop a Credibility Assessment Plan: Document model architecture, training data, evaluation methods, and performance metrics, with rigor proportional to model risk.
  5. Execute the Plan: Implement the credibility assessment activities.
  6. Document Results: Prepare a credibility assessment report with any deviations from the plan.
  7. Determine Adequacy: Evaluate whether model credibility is sufficient for the COU.
quadrantChart
    title AI Model Risk Assessment
    x-axis Low Model Influence --> High Model Influence
    y-axis Low Decision Consequence --> High Decision Consequence
    quadrant-1 High Risk
    quadrant-2 Medium Risk
    quadrant-3 Low Risk
    quadrant-4 Medium Risk
    "Patient stratification": [0.60, 0.85]
    "Dosing recommendations": [0.58, 0.66]
    "Eligibility screening": [0.30, 0.40]
    "Data quality triage": [0.46, 0.27]
    "Protocol complexity": [0.22, 0.16]
    "Document classification": [0.12, 0.32]
Figure 26.1: FDA AI Model Risk Matrix: Risk Increases with Model Influence and Decision Consequence

What You CAN and CANNOT Do with AI in Clinical Trials

Table 26.1 synthesizes the enforceable requirements and regulatory expectations from FDA guidance, EMA reflection papers, and ICH E6(R3), ordered from the applications a sponsor can deploy today to the one it cannot deploy at all.

Table 26.1: AI Applications in Clinical Trials: Regulatory Status and Requirements
AI Application Regulatory Status Requirements Limitations
Document classification and TMF filing Permitted with oversight Validate classification accuracy; maintain human review for inspection-critical documents Cannot replace human accountability for TMF completeness
Data quality screening and query generation Permitted with oversight Document AI logic; validate query generation; apply risk-based human review of generated queries No sponsor changes to investigator- or participant-entered data without the investigator’s advance agreement; automated query closure needs a validated, documented process
Patient matching/eligibility pre-screening Permitted as decision support Validate against eligibility criteria; investigator makes final determination Cannot make final eligibility decisions; investigator responsibility under GCP
Protocol complexity scoring Permitted Document methodology; validate predictions against historical data Operational tool only; no regulatory submission required
Site selection and feasibility Permitted Document data sources and model logic Operational tool; sponsor retains responsibility for site qualification
Adverse event case processing Permitted with lifecycle monitoring Monitor model performance; human review of serious/unexpected cases Cannot replace pharmacovigilance qualified person oversight
Statistical analysis endpoints Permitted with pre-specification Pre-specify in SAP; in pivotal trials, freeze model before database lock; prospective validation required for high-impact uses Non-prespecified model changes after unblinding make the analysis post hoc, not confirmatory
Primary endpoint assessment Conditional: high regulatory scrutiny EMA: prospective testing with newly acquired data required; FDA: credibility assessment proportional to risk Model must be frozen and fully documented in SAP
Dosing/treatment assignment decisions Conditional: requires early FDA/EMA engagement Full credibility assessment; human-in-the-loop required; extensive safety monitoring Cannot be sole determinant without validated safety controls
Replacing clinical judgment Not permitted N/A AI supports decisions; investigators and physicians retain accountability under GCP
WarningCritical Principle: Human Accountability Cannot Be Delegated to AI

Under ICH E6 GCP and FDA regulations, the investigator is responsible for medical decisions affecting trial participants, and the sponsor is responsible for trial conduct and data integrity. AI can support these decisions but cannot replace the accountable human. An investigator cannot defend a protocol deviation by stating “the AI said it was acceptable.”

AI Model Lifecycle Maintenance

Unlike static software, AI models may degrade or drift over time as input data distributions change. Both FDA and EMA guidance therefore expect lifecycle maintenance for models deployed over extended periods: in manufacturing (quality control, process optimization), in pharmacovigilance (case classification, signal detection), and in any AI system whose performance may change as new data arrives. Five activities make it up.

  1. Performance monitoring: Define metrics and thresholds; monitor on risk-based frequency.
  2. Drift detection: Identify when input data diverges from training data distribution.
  3. Revalidation triggers: Pre-define conditions requiring model retesting or retraining.
  4. Change management: Evaluate model changes through the applicable change-control process (for manufacturing models, the pharmaceutical quality system).
  5. Regulatory notification: Report changes impacting model performance per applicable requirements (e.g., 21 CFR 314.70 for manufacturing changes under an approved NDA).
TipEMA on Incremental Learning

For pharmacovigilance applications, EMA permits a “more flexible approach” where incremental learning can continuously enhance models. However, the MAH retains responsibility to validate, monitor, and document model performance as part of the pharmacovigilance system.

When to Engage Regulators

Early engagement with FDA or EMA is strongly recommended for AI applications with high regulatory impact or high patient risk; Table 26.2 summarizes the pathways.

Table 26.2: Regulatory Engagement Pathways for AI in Clinical Trials
AI Use Case Recommended Engagement FDA Contact EMA Contact
Novel clinical trial design using AI CDER C3TI or CID Meeting Program CDERclinicaltrialinnovation@fda.hhs.gov Innovation Task Force (ITF)
AI for endpoint evaluation Drug Development Tools (DDT) or ISTAND CDER-BiomarkerQualificationProgram@fda.hhs.gov (biomarkers; clinical outcome assessments and ISTAND have their own contacts) SAWP qualification
AI-enabled digital health technology DHT Program DHTsforDrugDevelopment@fda.hhs.gov N/A
AI in pharmacovigilance Emerging Drug Safety Technology Program AIMLforDrugDevelopment@fda.hhs.gov PRAC interaction
AI in manufacturing Emerging Technology Program (CDER ETP) CDER-ETT@fda.hhs.gov N/A
Model-informed drug development MIDD Paired Meeting Program MIDD@fda.hhs.gov SAWP scientific advice

EMA-Specific Considerations

The EMA reflection paper (September 2024) introduces terminology that differs somewhat from FDA’s. It sorts applications by high patient risk, meaning AI uses affecting patient safety such as dosing or treatment assignment, and by high regulatory impact, meaning AI uses substantially affecting regulatory decisions such as primary endpoint analysis. The rigor of the credibility assessment should be proportional to both. Five positions follow from that framing.

  1. Transparent models preferred: “The use of transparent models is preferred” to strengthen accountability. Black box models may be acceptable if transparent models show unsatisfactory performance, with additional documentation and monitoring requirements.

  2. Frozen models for pivotal trials: “Prior to the database lock and subsequent unblinding…the data pre-processing pipeline and all models should be frozen and documented in a traceable manner in the statistical analysis plan.”

  3. Prospective validation for high-impact uses: “Prior to model deployment in high regulatory impact settings, such as in relation to the primary endpoint, performance should be tested with prospectively generated data (future calendar time) that is acquired in a setting or population representative of the intended context of use.”

  4. No incremental learning in pivotal trials: “Incremental learning approaches are not accepted, and any modification of the model during the trial requires a regulatory interaction.”

  5. Human-in-the-loop for precision medicine: AI-driven indication or posology recommendations are “high patient risk as well as high regulatory impact” and require “fall-back treatment strategies in cases of technical failure.”

The EU AI Act

This is the instrument that is law. The European Union’s AI Act (Regulation (EU) 2024/1689) (European Parliament and Council of the European Union 2024) is the first comprehensive, binding AI statute, and it applies horizontally across sectors rather than being specific to medicines. Having entered into force in August 2024, it applies in phases: prohibitions on a narrow set of unacceptable-risk practices from February 2025, obligations for general-purpose AI models from August 2025, and the core obligations for high-risk systems originally due from August 2026, with a further tranche tied to regulated products (via the Medical Device Regulation and In Vitro Diagnostic Regulation) from August 2027. The Digital Omnibus on AI, provisionally agreed in May 2026 and in force from 27 July 2026, deferred both high-risk tranches: standalone (Annex III) systems now apply from 2 December 2027 and product-embedded (Annex I) systems from 2 August 2028. The Article 50 transparency obligations kept their 2 August 2026 start, except that providers of generative AI systems already on the market before that date have until 2 December 2026 to mark their outputs, in machine-readable form, as artificially generated (European Parliament and Council of the European Union 2026).

Most AI used in trial operations (document filing, query triage, feasibility scoring) is not “high-risk” under the Act’s definitions, so its obligations are light: providers and deployers must take measures to support their staff’s AI literacy, and Article 50’s transparency duties apply only to particular functions, such as systems that interact directly with people or generate synthetic content (European Parliament and Council of the European Union 2024, 2026). AI that functions as, or is a safety component of, a regulated medical device can fall into the high-risk tier, inheriting obligations layered on top of the device regulations. The Act sits alongside, not in place of, the EMA reflection paper and the Clinical Trials Regulation, so sponsors running EU trials should assess where each AI application falls under its classification: a system treated as “decision support” under GCP may still trigger binding obligations under the Act.

Practical Implications: A Decision Framework

Figure 26.2 turns all of this into a path a sponsor can walk before committing to a deployment.

%%{init: {'flowchart': {'nodeSpacing': 15, 'rankSpacing': 30, 'padding': 8, 'diagramPadding': 40, 'subGraphTitleMargin': {'top': 6, 'bottom': 14}}}}%%
flowchart TD
    A[Proposed AI<br/>Application] --> B{Does AI produce<br/>data/information for<br/>regulatory decisions?}
    B -->|"No -<br/>Operational only"| C[Lower regulatory burden<br/>Document for<br/>inspection readiness]
    B -->|Yes| D{What is the decision<br/>consequence if AI<br/>output is incorrect?}
    D -->|Low| F1{What is the AI<br/>model influence?}
    D -->|Medium/High| F2{What is the AI<br/>model influence?}
    F1 -->|"Low - other<br/>evidence also used"| E[Low-risk application<br/>Standard validation<br/>Document methodology]
    F1 -->|"High - AI is primary<br/>or sole evidence"| G[Medium-risk application<br/>Proportional credibility<br/>assessment<br/>Human oversight of outputs]
    F2 -->|"Low - other<br/>evidence also used"| G
    F2 -->|"High - AI is primary<br/>or sole evidence"| H[High-risk application<br/>Full credibility assessment<br/>Early regulatory engagement<br/>Prospective validation<br/>Pre-specified in SAP]

    C --> I[Proceed with GxP-compliant<br/>documentation and oversight]
    E --> I
    G --> J[Develop credibility<br/>assessment plan<br/>Consider regulatory<br/>feedback]
    H --> K[Engage FDA/EMA<br/>before deployment<br/>Full 7-step<br/>credibility process]
Figure 26.2: Decision Framework for AI Deployment in Clinical Trials

Summary: The Regulatory Bottom Line

Five things hold throughout. Human accountability for medical decisions and regulatory compliance cannot be delegated to AI; investigators, sponsors, and MAHs retain responsibility. Rigor of validation, documentation, and oversight should scale with model risk, which combines decision consequence and model influence. In pivotal trials, EMA expects the data pipeline and every model used for confirmatory analysis to be frozen and documented in the statistical analysis plan before database lock and unblinding; non-prespecified changes made after the data are opened make the results post hoc, unsuitable as confirmatory evidence (European Medicines Agency 2024). Models that keep operating over time, in manufacturing and pharmacovigilance, need performance monitoring, drift detection, and change management. For high-risk applications, early consultation with FDA or EMA aligns expectations before a late-stage objection becomes expensive.

26.3 AI across the Development Lifecycle

AI now touches every phase of clinical development. What follows marks where automation delivers measurable value and where human oversight remains essential.

AI in Protocol Design

Benchmarking studies suggest that most clinical trials fail to meet planned enrollment timelines, often because of design choices baked in before the first patient is screened (Lamberti et al. 2024). The tools in Table 26.3 attack the problem where it originates, at the design stage, following the workflow in Figure 26.3.

%%{init: {'flowchart': {'nodeSpacing': 20, 'rankSpacing': 30, 'padding': 8, 'diagramPadding': 30, 'subGraphTitleMargin': {'top': 6, 'bottom': 14}}}}%%
flowchart TB
    subgraph Input["Inputs"]
        RWD[Real-World Data<br/>large cohorts]
        Hist[Historical Trials<br/>Protocol library]
        Reg[Regulatory<br/>Requirements]
    end

    subgraph AI["AI Analysis"]
        Sim[Patient<br/>Simulation]
        Burden[Burden<br/>Scoring]
        Feasibility[Site<br/>Feasibility]
    end

    subgraph Output["Outputs"]
        Protocol[Optimized<br/>Protocol]
        SoA[Schedule of<br/>Activities]
        Sites[Recommended<br/>Sites]
    end

    RWD --> Sim
    Hist --> Burden
    Reg --> Protocol

    Sim --> Feasibility
    Sim --> Protocol
    Burden --> Protocol
    Feasibility --> Sites
    Feasibility ~~~ Protocol

    Protocol --> SoA
Figure 26.3: AI-Driven Protocol Design Workflow
Table 26.3: Protocol Design AI Tools Comparison
Tool Data Source Primary Use Case Key Metric
Faro Health Protocol library + RWD Operational burden prediction Complexity score
Phesi 100M+ patient profiles Enrollment simulation Patient availability
Medidata AI Historical trial data Protocol optimization Predicted enrollment rate
Formation Bio (formerly TrialSpark) Site network data Site selection Per-site enrollment probability

Faro Health exemplifies the approach. Instead of writing a static Word document, study teams design the trial in a structured cloud platform. The AI scores the complexity of the schedule of assessments against real-world data to predict operational burden, shows the team which visits pile too many procedures onto one patient, and then generates the protocol document and Schedule of Activities from the structured design (Faro Health 2024).

Phesi takes simulation further, drawing on data from over 100 million patients to model trial outcomes before the protocol is finalized. Their Digital Patient Profile reduces the likelihood of the “zero-enrollment site” problem mentioned in Chapter 18.

AI in Recruitment and Site Selection

Finding the right patients remains the perennial bottleneck. Benchmarking studies suggest that recruitment and retention often take substantially longer and cost more than planned, and that each day of delay can be economically material for sponsors in some therapeutic areas (Lamberti et al. 2024; Deloitte Centre for Health Solutions 2025). AI tools now scan electronic health records for “needle in the haystack” candidates and, when integrated into workflow, can reduce the manual effort of identifying potentially eligible participants. Figure 26.4 shows the typical pipeline; Table 26.4 profiles the leading vendors.

%%{init: {'flowchart': {'nodeSpacing': 20, 'rankSpacing': 40, 'padding': 8, 'diagramPadding': 30, 'subGraphTitleMargin': {'top': 6, 'bottom': 14}}}}%%
flowchart TB
    subgraph Sources["Patient Data Sources"]
        EHR["EHR<br/>Systems"]
        Claims["Claims<br/>Data"]
        Labs["Lab<br/>Results"]
        Genomics["Genomic<br/>Profiles"]
    end

    subgraph NLP["NLP Processing"]
        Extract["Entity<br/>Extraction"]
        Normalize["Terminology<br/>Normalization"]
        Temporal["Temporal<br/>Reasoning"]
    end

    subgraph Match["Matching Engine"]
        Criteria["I/E Criteria<br/>Parser"]
        Score["Eligibility<br/>Scoring"]
        Rank["Patient<br/>Ranking"]
    end

    subgraph Output["Results"]
        Patients["Matched<br/>Patients"]
        Sites["Optimized<br/>Sites"]
        Alerts["Provider<br/>Alerts"]
    end

    Protocol["Protocol<br/>I/E Criteria"] --> Match

    EHR --> Extract
    Claims --> Extract
    Labs --> Normalize
    Genomics --> Normalize

    Extract --> Normalize
    Normalize --> Temporal

    Temporal --> Score
    Criteria --> Score
    Score --> Rank

    Rank --> Output
    Patients --> Sites
    Patients --> Alerts
Figure 26.4: AI-Powered Patient Matching Pipeline
Table 26.4: Patient Recruitment AI Vendors
Vendor Technology Data Assets Best For
NextTrial.ai NLP + ML matching EHR integration Complex I/E criteria
H1 KOL mapping, investigator analytics Publications, claims, trials Site selection, investigator finding
Deep 6 AI Real-time EHR search Health system partnerships Oncology, rare disease
TriNetX Federated network 300M+ patients (2026) Global feasibility
Komodo Health Healthcare map Claims + RWD Patient path analysis

NextTrial.ai uses natural language processing to bridge the gap between protocol criteria and patient records. The platform ingests unstructured data (clinician notes, pathology reports, and genetic profiles), flags the patients who meet complex inclusion/exclusion criteria, and predicts which investigator sites hold the highest density of them (NextTrial 2024).

H1 maps the global network of Key Opinion Leaders and investigators instead. Analyzing large volumes of data from publications, claims records, and clinical trial databases, it helps sponsors find investigators who actively treat the target patient population, not only those who are academically prominent (H1 2024).

AI in Clinical Operations and Data Quality

Once data starts flowing, it must be cleaned, and wearables generate 1,000+ data points per patient per day. At that volume manual query resolution stops being sustainable. The emerging approach layers several AI techniques across the data pipeline, each suited to a different kind of quality problem; Table 26.5 profiles the leading tools.

Rule-based checks remain the foundation: programmed validations that flag impossible values (negative ages, dates in the future) or logical inconsistencies (randomization before consent). They are deterministic, fast, auditable, and well-understood.

Machine learning anomaly detection adds statistical pattern recognition. Models trained on historical trial data can identify unusual distributions, digit preferences suggestive of fabrication, or sites whose data patterns diverge from comparators, which means they surface problems nobody anticipated during study design.

Large language models (LLMs) parse the free-text fields, adverse event narratives, medical history, concomitant medication entries, for inconsistencies invisible to structured checks: a narrative describing “chest pain radiating to left arm” against a coded adverse event of “headache.”

Cross-domain validation checks EDC entries against central lab results, device uploads against visit schedules, and ePRO responses against clinical observations; discrepancies across domains often indicate transcription errors or protocol deviations.

The output is a set of automated actions: auto-generated queries with draft text for site response, risk flags that prioritize monitoring attention, trend alerts that surface site-level quality patterns, and quality reports for oversight review (Figure 26.5). The goal is not to displace data management staff but to move their time off low-value triage and onto judgment-intensive resolution.

%%{init: {'flowchart': {'nodeSpacing': 20, 'rankSpacing': 30, 'padding': 8, 'diagramPadding': 20, 'subGraphTitleMargin': {'top': 6, 'bottom': 14}}}}%%
flowchart LR
    subgraph Ingest["Data Ingestion"]
        EDC[EDC Data]
        Devices[Device Data]
        Labs[Central Labs]
        ePRO[ePRO Responses]
    end

    subgraph AI["AI Processing"]
        Rules[Rule-Based Checks]
        ML[ML Anomaly Detection]
        LLM[LLM Discrepancy Analysis]
        Cross[Cross-Domain Validation]
    end

    subgraph Actions["Automated Actions"]
        AutoQuery[Auto-Generated Queries]
        Flag[Risk Flags]
        Trend[Trend Alerts]
        Report[Quality Reports]
    end

    EDC --> Rules
    EDC --> ML
    EDC --> LLM
    EDC --> Cross
    Devices --> Rules
    Devices --> ML
    Labs --> Rules
    Labs --> Cross
    ePRO --> Rules
    ePRO --> LLM

    Rules --> AutoQuery
    ML --> Flag
    LLM --> AutoQuery
    Cross --> Trend

    AutoQuery --> Report
    Flag --> Report
    Trend --> Report
Figure 26.5: AI-Powered Data Quality Management
Table 26.5: Data Quality AI Tools
Tool Primary Approach Best Feature Integration Depth
Saama Clinical Command Center Unified EDC/CTMS/eTMF view Deep multi-system
Octozi LLM-based discrepancy detection Natural language queries EDC-focused
Veeva CDB Vault-native data management Single-platform simplicity Veeva ecosystem
Medidata Detect Signal detection algorithms Safety signal identification Rave ecosystem

Saama provides an AI-driven Clinical Command Center, unifying data from EDC, CTMS, and eTMF into an integrated view of trial health. Its models, trained on large proprietary datasets from clinical trial operations, predict site non-compliance and enrollment delays, enabling intervention before problems escalate (Saama Technologies 2024).

Octozi applies Large Language Models to data review, scanning for inconsistencies such as “male patient listed as pregnant” or “adverse event date before informed consent.” By surfacing these issues immediately, Octozi narrows the line-by-line manual review that traditionally consumed data management resources, leaving reviewers to adjudicate the flagged records rather than scan every listing (Octozi 2024).

26.4 Emerging Methodologies and Operational Solutions

Innovation in clinical trials is not limited to software; it includes new methodological frameworks that use real-world data, federated learning, and digital simulations to rethink the evidence-generation process.

Solving Logistical Challenges with IT

Modern trials face significant logistical friction (see Section 20.6.1). The challenges are structural: supply chains that must deliver temperature-sensitive products to thousands of endpoints, sites overwhelmed by redundant data entry across disconnected systems, protocols too complex for manual compliance tracking, and global operations fragmented across time zones and organizations. These are coordination problems, and coordination problems respond to better information infrastructure, as Table 26.6 illustrates.

Supply chain complexity has intensified with decentralized trials and direct-to-patient distribution. Traditional approaches relied on spreadsheets and email to track inventory and shipments; delays were discovered after the fact. Modern systems integrate IoT sensors that monitor temperature, location, and chain of custody in real time. AI-based demand forecasting uses enrollment trajectories and visit schedules to predict resupply needs before stockouts occur. The result is reduced waste (fewer expired products) and fewer patient visits disrupted by supply failures.

Site burden accumulates when coordinators must enter the same data into multiple systems: pulling information from the electronic health record, transcribing it to the EDC, and reconciling discrepancies later. EHR-to-EDC integration automates the flow of structured data (lab values, vital signs, demographics) from the source system to the trial database, reducing transcription errors and freeing coordinator time for patient-facing work. This integration requires careful validation and mapping, but when implemented well, it addresses one of the most persistent complaints from clinical sites.

Protocol deviations often result from complexity: too many visits, too many procedures, too many eligibility criteria for staff to track manually. Real-time nudge systems monitor visit windows and upcoming assessments, alerting coordinators before a deviation occurs rather than flagging it retrospectively. Detection systems analyze patterns across sites to identify systematic compliance failures that may indicate training gaps or protocol design problems.

Global operations compound all of these challenges. When a trial spans 25 countries, 80 sites, and multiple CRO partners, fragmented regional teams operating from different data sources create reconciliation overhead and conflicting reports. Unified command centers aggregate operational data into a single platform, providing consistent metrics, standardized escalation pathways, and a shared view of trial status across all stakeholders.

Table 26.6: IT Solutions for Logistical Challenges (reported impacts vary by implementation)
Challenge Traditional Approach Modern IT Solution Reported Impact
Supply Chain Spreadsheets, email IoT sensors, AI forecasting Reduced waste, fewer stockouts
Site Burden Redundant data entry EHR-to-EDC integration Less transcription time
Protocol Deviations Retrospective monitoring Real-time nudge engines Fewer major deviations
Global Operations Fragmented regional teams Unified command centers Single source of truth

Novel Methodological Approaches

This section covers a set of approaches that challenge traditional assumptions about how evidence is generated and compared: using observational data to support decisions (RWE), using historical or synthetic comparators to reduce concurrent control enrollment, and using models to predict counterfactual outcomes. All remain subjects of active regulatory and scientific debate.

Clinical research data are fragmented across institutions, each with privacy regulations, competitive interests, and technical barriers to sharing. Federated learning addresses this by training models collaboratively without centralizing data: algorithms travel to data sources, train locally, and share only model updates (gradients or weights) rather than patient records. Owkin has applied federated learning to construct external control arms from real-world data distributed across hospital networks. Their 2025 publication in Nature Communications described the method, FedECA, and deployed it across cancer centers in France, Spain, and the United States without pooling patient data; the authors note that the shared updates could in theory still be exploited to recover patient information, and they tested differential privacy as a safeguard (Ogier du Terrail et al. 2025). The practical value is clearest in rare diseases and oncology, where recruiting concurrent controls may be infeasible or ethically questionable. However, the approach requires sophisticated infrastructure, standardized data formats across institutions, and careful attention to the comparability of patient populations.

Digital twin is borrowed from engineering, where virtual models of physical systems enable simulation and optimization (Laubenbacher, Sluka, and Glazier 2021). In clinical research the term means a computational model that predicts an individual patient’s disease trajectory under control conditions, based on baseline characteristics and historical data from similar patients (Sel et al. 2025; Ross et al. 2024).

Unlearn.AI’s PROCOVA methodology is the worked example. Models trained on historical patient data generate, for each enrolled patient, a prognostic score forecasting that patient’s likely outcome on the control arm; the score then enters the primary analysis as a baseline covariate, reducing residual variance and allowing smaller control groups while preserving unbiased treatment effect estimates and Type I error control (European Medicines Agency, Committee for Medicinal Products for Human Use (CHMP) 2022; Ross et al. 2024). The European Medicines Agency issued a favorable qualification opinion in September 2022 (European Medicines Agency, Committee for Medicinal Products for Human Use (CHMP) 2022). FDA has not qualified the method; Unlearn reports that the agency’s Center for Drug Evaluation and Research, responding to its submission, found the approach already covered by FDA’s guidance on covariate adjustment in randomized trials (Fisher 2024).

What EMA qualified was not “twin replaces a randomized patient.” It qualified PROCOVA as a covariate adjustment method, analogous to using baseline HbA1c or age as a covariate in an ANCOVA, except that the covariate is itself the forecast of a deep neural network trained on historical data. The randomized control arm is still required; what the method saves is sample size, not concurrent randomized controls.

A team evaluating PROCOVA therefore has one question to answer. Standard covariate adjustment using well-chosen baseline predictors (baseline disease score, amyloid level, MMSE) can achieve much of the same variance reduction without a machine learning model, so the value of PROCOVA turns on whether the deep neural network predicts the 18-month outcome better than those covariates do. Indication by indication, that is not settled.

ImportantAI in regulated trials: proprietary models require transparency

If the prognostic score driving the analysis comes from a proprietary black-box model, FDA and EMA cannot independently verify the model’s behavior, and sponsors cannot audit whether the prediction is stable across the training data and population. Any AI or machine learning model used as part of a primary statistical analysis in a regulated trial should be written down in enough detail that a third party can reprogram it from the specification and improve it as new data emerge. A score delivered without the model behind it is a governance risk even when regulators have accepted the method class, because acceptance of a method class is not acceptance of any specific model implementation.

Real-World Evidence and Regulatory Considerations

The randomized controlled trial remains the gold standard for establishing causal treatment effects, but real-world evidence (RWE) is playing an expanding role in drug development and regulatory decision-making. Real-world data (RWD) is data on patient health and health care delivery collected routinely rather than for a traditional trial: electronic health records, claims databases, disease registries, wearable devices, and patient-generated data. Analyzed appropriately, in observational studies or within randomized designs such as pragmatic trials, it yields evidence about drug safety, effectiveness, and utilization patterns, and the distinction matters: RWD is the raw data, RWE the clinical evidence derived from it through rigorous analysis (U.S. Food and Drug Administration 2018).

For safety, RWE has long been used in post-marketing surveillance, detecting rare adverse events that trials were not powered to identify. For effectiveness, FDA acceptance is more cautious but expanding: the 21st Century Cures Act directed the agency to evaluate RWE for approving new indications for existing drugs and for satisfying post-marketing study requirements, and FDA’s 2023 final guidance on RWE for regulatory decisions emphasized that data relevance, reliability, and analytic rigor determine acceptability, not the mere existence of a large dataset (U.S. Food and Drug Administration 2023b).

Single-arm trials with external controls compare treated patients against matched historical or concurrent observational cohorts; they are most defensible in diseases with well-characterized natural history and no effective standard of care. Hybrid designs randomize a reduced control arm and borrow information from external data to increase statistical precision. Post-marketing effectiveness studies use RWD to ask whether efficacy demonstrated under controlled conditions survives contact with real-world populations carrying the comorbidities and concomitant medications that pivotal trials excluded.

The core challenge is confounding: in observational data, treatment selection is not random. Patients who receive a therapy may differ systematically from those who do not, and these differences, rather than the treatment itself, may explain observed outcomes. Propensity score methods, instrumental variables, and target trial emulation frameworks attempt to address confounding, but none can fully substitute for randomization. Unmeasured confounders remain a fundamental limitation.

Nonrandomized RWE is most credible when the outcome is objective and reliably captured in routine care, the comparison is against natural history rather than an active comparator, the patient population in the RWD source is demonstrably similar to the trial population, and sensitivity analyses show robustness to plausible unmeasured confounding. Even then, regulators typically treat such evidence as supportive rather than dispositive for efficacy claims: it strengthens a submission that already contains randomized evidence.

FDA’s 2023 draft guidance on externally controlled trials sets out how to build an external control arm from patient-level data from other clinical trials or from real-world data, addressing data quality, comparability, and bias mitigation (U.S. Food and Drug Administration 2023a). It emphasizes that external controls are most appropriate when concurrent randomization is infeasible or unethical, and when the disease has a well-characterized natural history with reliable outcome measurement in routine care.

In 2020, FDA accepted a hybrid synthetic control arm (SCA) design for Medicenna Therapeutics’ planned Phase III trial in recurrent glioblastoma, a precedent-setting acceptance of a hybrid external control in a registrational trial for an indication that had previously relied on conventional randomized controls. In that design, matched external controls make up two-thirds of the pooled control arm, cutting prospective control enrollment by approximately two-thirds relative to conventional 1:1 randomization (Medicenna Therapeutics Corp. 2020; Medidata Solutions 2020).

An SCA constructs the external comparator from patient-level data in historical trials rather than recruiting new control patients. Medidata’s platform draws on a database spanning over 38,000 trials and 12 million patients to statistically match historical controls to current trial populations. The approach is most applicable in rare or life-threatening diseases with inadequate standard-of-care, where historical control data are plentiful and disease progression is well-characterized; hybrid designs, which add historical controls to a smaller concurrent control arm, are one way to extend it to later-stage trials.

Legitimate skepticism attaches to all of this, and it is specific. Verification, validation, and uncertainty quantification (VVUQ) remain incompletely standardized. A 2025 review in npj Digital Medicine called VVUQ frameworks essential for bringing digital twins into clinical practice, argued that VVUQ processes should be standardized as far as possible for quality assurance and risk management, and noted that continually updated digital twins pose a regulatory challenge that FDA’s existing medical-device framework does not fully address (Sel et al. 2025). Uncertainty is the piece most often underemphasized: the same review observes that uncertainty quantification is often lacking in modeling work, even though it is what attaches confidence bounds to a model’s predictions (Sel et al. 2025). A single predicted outcome per patient is therefore inadequate; a digital twin used as a counterfactual should report a distribution of plausible outcomes, not a point estimate.

Model opacity carries a cost that is easy to overlook. PROCOVA is a special case of classical ANCOVA: the adjustment covariate is a prognostic score from a model trained on independent historical data, in Unlearn’s implementation a proprietary machine-learning model (Ross et al. 2024; European Medicines Agency, Committee for Medicinal Products for Human Use (CHMP) 2022). EMA qualified the adjustment procedure but not any procedure for building that model, which leaves each sponsor to develop and validate the prognostic model for its own trial, and it expects sponsors to justify the advantage PROCOVA can attain over ANCOVA adjusting for individual covariates (European Medicines Agency, Committee for Medicinal Products for Human Use (CHMP) 2022). When model details are withheld, sponsors and regulators cannot interrogate, replicate, or improve the methodology, and EMA’s reflection paper on artificial intelligence in the medicinal product lifecycle expects models that influence regulatory decision-making to be fully documented, traceable, and open to independent assessment (European Medicines Agency 2024).

Data quality and completeness constrain all approaches. In oncology, the data needed to model tumor dynamics are often noisy, incomplete, and subject to collection burden (Venkatesh, Raza, and Kvedar 2022), and statistical models layer their own uncertainty on top, particularly when predictions must generalize across populations and disease subtypes.

Comparability assumptions underpin every external control approach. If the historical population differs systematically from the current trial population (because standard of care, patient selection, or measurement practices have changed), treatment effect estimates may be biased regardless of the sophistication of the matching algorithm.

In a JAMA Special Communication published online in October 2024 (January 2025 issue), then-FDA Commissioner Robert Califf and senior FDA officials listed digital twins that generate simulated clinical records to predict placebo-group response among the current and potential uses of AI in clinical research, alongside matching individuals to trials and harmonizing data cleaning, and called for flexible, risk-based regulatory schemes (Warraich, Tazbaz, and Califf 2025). FDA’s January 2025 draft guidance on AI in drug development makes that risk basis concrete: the oversight and credibility evidence expected for a model scale with its influence on a regulatory decision and the consequence of an incorrect decision, and operational uses that do not affect patient safety, drug quality, or the reliability of study results fall outside its scope (U.S. Food and Drug Administration 2025).

Opaque, proprietary methods may therefore face higher evidentiary bars than interpretable, well-documented ones, with sponsors expected to show statistical validity, model transparency, reproducibility, and appropriate characterization of uncertainty. Table 26.7 summarizes where each approach currently stands.

Table 26.7: Emerging Clinical Trial Technologies: Status and Limitations
Approach How It Works Regulatory Status Key Limitation
Federated Learning Models train locally; only updates shared FDA/EMA engagement; no formal qualification Infrastructure complexity
Digital Twins (PROCOVA) Prognostic scores as covariates EMA qualified (2022) (European Medicines Agency, Committee for Medicinal Products for Human Use (CHMP) 2022); FDA: covered by existing guidance, per Unlearn (2024) (Fisher 2024) Model opacity; uncertainty
Synthetic Control Arms Historical data matched to trial FDA accepted hybrid Phase III design (2020) Comparability assumptions

26.5 Future Impact and Economics

The technologies in this chapter bear on the cost drivers of clinical development: shorter cycle times, smaller samples, and less coordination labor, set against new operational complexity.

AI Impact on Trial Operations

Recall the economics from Chapter 6: $2.7 billion to bring a drug to market (Deloitte’s 2025 cohort), $40,000 per day to operate a Phase II/III trial, and, on average, roughly $800,000 per day in opportunity cost from delays, rising to $1.3-1.4 million per day in the most affected therapeutic areas (Z. P. Smith, DiMasi, and Getz 2024; Z. Smith, DiMasi, and Getz 2024; Deloitte Centre for Health Solutions 2026). AI-enabled tools are being applied to each of these cost drivers (see Table 26.8), though the realized impact depends on implementation quality, governance, and the specific operational context.

Table 26.8: Potential AI Impact on Clinical Trial Cost Structure (industry projections; actual results vary)
Current Cost Structure AI Interventions Impact Estimates
\(\approx\) $2-3B to market (est.) Protocol Optimization AI Reduced Development Cost
High attrition (\(\approx\) 90%) Digital Twins & Synthetic Arms Smaller Trials, Similar Power
High daily burn (\(\approx\) $40K/day) Autonomous Agents Lower coordination overhead
CRO gross margins (\(\approx\) 18-33% of revenue) Direct Automation Margin Pressure
Time to first patient (\(\approx\) 166d avg) Parallel Processing AI Faster Timelines

Cost to market decomposes into a small set of dominant drivers: late-stage attrition, enrollment delays, protocol amendments, monitoring intensity, and the long tail of data cleaning. Automation cannot eliminate scientific uncertainty; the question is whether it can shift those operational drivers enough to change the capitalized cost of development in a meaningful way. Table 26.9 sets the main levers discussed in this chapter against the drivers they target.

Table 26.9: Potential AI Impact on Major Trial Cost Drivers (estimates based on industry reports; actual results depend on implementation) (Lamberti et al. 2024; Getz et al. 2016; Anagnostopoulos et al. 2023; Sertkaya et al. 2016)
Cost Driver Current State AI Approach Potential Impact
Failed Programs ~90% of drugs do not reach market Protocol simulation, patient selection AI Earlier go/no-go decisions; potentially fewer late-stage failures
Enrollment Delays Most trials miss enrollment targets NLP-based patient matching, predictive site selection Faster enrollment (magnitude varies by indication)
Data Cleaning Manual query resolution at $50-100/query ML-assisted triage and draft resolutions Reduced query burden
Monitoring Costs Site monitoring \(\approx\) 9-14% of itemized per-study costs Risk-based + AI-detected anomalies Reduced on-site visit frequency
Protocol Amendments $141K (Phase II) to $535K (Phase III) median direct cost per substantial amendment Design simulation before finalization Potentially fewer amendments

The largest listed CROs reported fiscal 2025 gross margins of roughly 18-33% of revenue (revenue running about 23-50% above direct costs) and operating margins ranging from a loss to about 21% (see Chapter 6). AI-enabled tools are increasingly capable of performing tasks that have traditionally justified those margins: document processing, query management, routine monitoring, and medical coding. As sponsors develop their own AI capabilities, they may demand either lower CRO fees or differentiated services that require domain expertise AI cannot replicate. The speed and extent of this shift remain uncertain.

The “daily burn” of operating a trial is largely a labor-and-coordination cost: monitors reviewing data, coordinators entering and reconciling information, medical writers preparing narrative sections, and project managers coordinating activities across teams and vendors. Some of that work is amenable to partial automation, such as drafting or triage; the rest is inherently judgment- and accountability-driven. Table 26.10 decomposes the spend by role and task category.

Table 26.10: Clinical Operations Roles and Illustrative AI Automation Potential (the percentages and timelines are the authors’ illustrative judgments, not measured benchmarks)
Role Current Cost Contribution AI Automation Potential Timeline
Data Manager Query generation, cleaning 80% automatable Now - 2027
Clinical Monitor Source verification, oversight 50% automatable (remote SDV) 2025 - 2028
Medical Writer Narratives, CSR sections 60% draft automation Now - 2026
Project Manager Status tracking, reporting 40% automatable 2026 - 2029
Regulatory Affairs Submission compilation 70% automatable 2025 - 2027

Where automation reduces repetitive coordination work and shortens the reconciliation tail, the operating cost per day could decline, by a margin that depends on protocol complexity, data sources, and the quality of implementation and oversight.

Delays rarely arise from a single missing document or a single late shipment. They arise from cascades: a contract delay postpones site activation, which shifts the enrollment curve, which changes the drug demand forecast, which raises the risk of stockouts or wastage. That is what makes the logistics challenges described in Chapter 20 (multi-country execution, dozens of sites, long startup timelines, and tightly constrained supply chains) coordination problems in a precise sense: progress depends on many interdependent tasks distributed across organizations, time zones, and systems of record. Oversight imposes its own constraints, since actions must be traceable, exceptions must be reviewable, and accountability cannot be delegated to an opaque process.

In that environment, “agentic” automation is most defensible when it behaves like structured orchestration rather than free-form autonomy (Figure 26.6). Its practical contributions are to maintain state across workflows, to triage and route exceptions, to generate standardized artifacts (draft correspondence, document metadata), and to propose next actions that a responsible owner can approve. Used this way, orchestration turns scattered operational signals into a prioritized work queue with clear provenance, while high-stakes decisions (changes that affect participant safety, protocol interpretation, or regulatory commitments) stay under explicit human control.

%%{init: {'flowchart': {'nodeSpacing': 25, 'rankSpacing': 30, 'padding': 8, 'diagramPadding': 30, 'subGraphTitleMargin': {'top': 6, 'bottom': 14}}}}%%
flowchart LR
    subgraph Orchestrator["Trial<br/>Orchestration<br/>Agent"]
        Master["Master<br/>Coordinator"]
    end

    subgraph Functional["Functional Agents"]
        Site["Site Activation<br/>Agent"]
        Enroll["Enrollment<br/>Agent"]
        Supply["Supply Chain<br/>Agent"]
        Data["Data Quality<br/>Agent"]
        Safety["Safety Monitoring<br/>Agent"]
        Doc["Document<br/>Agent"]
    end

    subgraph Actions["Proposed Actions"]
        A1["Draft site<br/>contracts"]
        A2["Pre-screen<br/>patient matches"]
        A3["Predict<br/>inventory needs"]
        A4["Draft query<br/>responses"]
        A5["Draft safety<br/>narratives"]
        A6["Classify & tag<br/>TMF docs"]
    end

    Master --> Site
    Master --> Enroll
    Master --> Supply
    Master --> Data
    Master --> Safety
    Master --> Doc

    Site --> A1
    Enroll --> A2
    Supply --> A3
    Data --> A4
    Safety --> A5
    Doc --> A6

    A1 --> HITL
    A2 --> HITL
    A3 --> HITL
    A4 --> HITL
    A5 --> HITL
    A6 --> HITL
    HITL{{"Human review<br/>& approval"}} --> Exec["Execute after<br/>sign-off"]
Figure 26.6: Structured Agent Orchestration in Clinical Trials (Human-in-the-Loop)

Study start-up (from final protocol to first patient first visit) averaged 166 days for Phase III protocols in the Tufts CSDD benchmark (Getz, Smith, and Kravet 2023). The delay is distributed across feasibility assessment, contract negotiation, IRB/EC submission, site training, and investigational product shipment. Each step has its own bottleneck, and the steps are typically processed one after another.

AI-enabled tools attack each bottleneck. Predictive site scoring trained on historical enrollment data can replace the manual feasibility questionnaire, contract generation from standardized templates can shorten legal negotiation cycles, and auto-populated submission packages can accelerate IRB preparation. Adaptive e-learning modules can replace in-person training for routine content, and predictive supply chain systems can coordinate IP shipment with site readiness.

The bigger shift is from sequential to parallel processing. Software can execute some activation workstreams simultaneously (assembling submission-ready document sets, preparing templated contract packets) while review, signature, and site-specific customization still run in series. Parallelization can compress elapsed time, but the magnitude depends on integration quality, local ethics processes, contracting norms, and the extent to which sites and sponsors adopt common platforms.

Decentralized trials (DCTs) create a logistics challenge of much greater scale (see Table 26.11): instead of shipping investigational product to a limited number of sites, sponsors may ship to thousands of patient homes and coordinate services across a wider set of vendors and geographies (U.S. Food and Drug Administration 2024). Routine coordination and exception handling at that volume can overwhelm operational teams, which is what pushes organizations toward greater automation and analytics, with the same division of labor: standardized, traceable workflows for the routine components, accountable human review for anything safety- or compliance-critical.

Table 26.11: AI Solutions for Decentralized Trial Logistics
DCT Challenge Traditional Approach AI-Enabled Approach
Cold-Chain Monitoring Periodic temperature logs Real-time IoT + predictive intervention
Home Nursing Coordination Manual scheduling AI-assisted routing and scheduling
Patient Adherence Site calls, diaries Wearable + app AI detecting non-adherence patterns
Document Collection Email reminders, faxes TMF agents that chase, collect, and file automatically
Regulatory Compliance Per-country manual review Multi-jurisdiction compliance AI

Economics and Business Models

Sponsors face a strategic choice: build AI capabilities internally, buy them from vendors, or partner with AI-native CROs. Table 26.12 sets out what each path costs, what it risks, and which sponsors it suits.

Table 26.12: AI Adoption Strategies and Economic Trade-offs (illustrative author estimates; actual budgets vary widely by sponsor size and scope)
Strategy Upfront Investment Ongoing Cost Risk Best For
Build Internal AI $10-50M+ High (talent, compute) Technology obsolescence Top 20 pharma
Buy Platform AI $1-5M licensing Medium (per-trial fees) Vendor lock-in Mid-size sponsors
Partner with AI-Native CRO Minimal Higher per-trial cost Dependency on partner Biotech, small sponsors
Hybrid Model $5-20M Medium Complexity Most large sponsors

The 2030 Trial: A Vision

By 2030, a Phase III trial could look different in organizations that successfully integrate automation into validated workflows. Decision-making does not become “fully autonomous.” What changes is that a larger fraction of routine coordination work becomes standardized, traceable, and partially automated: documents assembled from structured sources instead of rewritten, operational signals reconciled across systems of record rather than re-keyed into decks, exceptions surfaced with clear provenance instead of discovered late through periodic reconciliation.

That operating model needs three things at once: data capture and integration mature enough that key events (screen failures, visit-window risk, temperature excursions, missing essential documents) arrive as machine-readable signals; validation, access controls, and audit trails good enough that automated steps stay inspectable and accountable under GCP and 21 CFR Part 11 expectations; and sites and vendors that actually adopt the workflow, so that automation reduces burden instead of shifting it into more portals and alerts. The scenario in Figure 26.7 and the metrics in Table 26.13 are aspirational on purpose: they show where cycle time and coordination cost might fall, not a single inevitable trajectory.

%%{init: {'flowchart': {'nodeSpacing': 20, 'rankSpacing': 30, 'padding': 8, 'diagramPadding': 30, 'wrappingWidth': 400}}}%%
flowchart TB
    subgraph r1[" "]
        direction LR
        S1["<b>Design (weeks)</b><br/><i>Protocol AI</i><br/>• Simulates large<br/>virtual cohorts<br/>• Optimizes I/E criteria<br/>automatically<br/>• Generates protocol<br/>document"] --> S2["<b>Activation (weeks)</b><br/><i>Site AI</i><br/>• Identifies<br/>high-performing sites<br/>• Generates contracts,<br/>submits to IRBs in parallel<br/>• Ships IP with<br/>predictive inventory"] --> S3["<b>Enrollment (months)</b><br/><i>Matching AI</i><br/>• Scans large EHR cohorts<br/>• Alerts physicians of<br/>eligible patients<br/>• Patients consent via<br/>eConsent platform"]
    end

    subgraph r2[" "]
        direction LR
        S4["<b>Conduct (months)</b><br/><i>Operations AI</i><br/>• Monitors data<br/>in real-time<br/>• Generates and resolves<br/>queries automatically<br/>• Predicts and prevents<br/>protocol deviations"] --> S5["<b>Close (weeks)</b><br/><i>Reporting AI</i><br/>• Locks database with<br/>automated reconciliation<br/>• Generates statistical<br/>outputs<br/>• Drafts Clinical Study<br/>Report"]
    end

    r1 ~~~ r2
    style r1 fill:none,stroke:none
    style r2 fill:none,stroke:none
Figure 26.7: The AI-Native Clinical Trial of 2030
Table 26.13: Scenario trial metrics in an AI-native operating model (illustrative; not a forecast)
Metric 2025 Benchmark 2030 Projection Change
Time to First Patient 166 days 60 days -64%
Enrollment Duration 500+ days 200 days -60%
Daily Operational Cost $40,000 $18,000 -55%
Total Phase III Cost $100M $45M -55%
Human FTEs per Trial 50-100 15-30 -70%
Data Quality Issues 15-20% query rate 3-5% query rate -75%

Industry and Workforce Implications

The same efficiency gain is a saving for one party and a revenue loss for another. Costs, responsibilities, and bargaining power are unevenly distributed across the clinical trial enterprise: organizations bearing the largest absolute operational costs capture most of the direct savings, while organizations earning revenue from the labor-intensive activities that automation targets find the same improvement economically disruptive. Regulatory exposure cuts the same way: actors responsible for compliance, inspection readiness, and patient protection face higher requirements for validation, audit trails, and accountability, which can slow adoption and increase the fixed costs of implementation.

Where automation is deployed matters too (site workflows, sponsor oversight, CRO service delivery, or platform infrastructure), as does whether a new tool reduces burden or simply shifts it across an organizational boundary. Table 26.14 summarizes the resulting incentives.

Table 26.14: Potential AI Transition Impact by Stakeholder (author assessment)
Stakeholder Likely Position Rationale
Large Sponsors Positioned to benefit Can afford AI investment; may capture efficiency gains if governance is effective
Biotech Mixed May benefit from lower trial costs, but less capital to invest in AI infrastructure
Traditional CROs Facing pressure Core services may be commoditized; business models may need to evolve
AI-Native CROs Potentially advantaged Built for automation; lower cost structure if quality can be demonstrated
Clinical Sites Mixed Less manual work, but also potentially less revenue per patient
Patients Potentially benefit May see faster access to trials, less burden, more decentralized options
Regulators Adapting Must develop AI validation frameworks while maintaining safety standards

The clinical research workforce (coordinators, monitors, data managers) faces significant disruption. Automation may reduce some routine tasks while work shifts toward system implementation, validation, and oversight; preserving domain expertise through that transition will take deliberate reskilling and role redesign.

The clinical trial technology ecosystem is moving in two directions at once. Core systems are consolidating around a small number of platforms that can support regulated, end-to-end execution across EDC, CTMS, eTMF, and supply, while sponsors experiment with a growing set of analytics and automation capabilities that sit “on top” of those systems and reshape day-to-day work in feasibility, monitoring, data cleaning, and documentation. That is a gradual reallocation of investment rather than a single wave of disruption: stability in the transactional backbone, faster iteration at the analytics and workflow layer (Table 26.15) (Tufts Center for the Study of Drug Development 2024; Fortune Business Insights 2024):

Table 26.15: Clinical Trial Technology Market Projections (analyst estimates; subject to revision; the AI 2030 value is interpolated from the source’s 2034 forecast at its stated CAGR) (Tufts Center for the Study of Drug Development 2024; Fortune Business Insights 2024; MarketsandMarkets 2024)
Segment 2025 Market Size 2030 Projection CAGR Key Trend
CTMS Platforms $2.4B $4.9B 15% Cloud consolidation
eTMF Systems $1.4B $2.5B 13% AI-powered automation
AI in Clinical Trials $3.8B $20B 39% Rapid adoption
EDC Systems $1.2B $2.1B 12% EHR integration

The IT stack has evolved from a passive repository to an active intelligence layer. Sponsors who integrate design, recruitment, and operations data in near real time can often detect operational risks earlier, and they may be able to run trials faster and at lower operational cost than fragmented legacy workflows allow, by a margin set by disease area, protocol complexity, and the quality of implementation and governance.

Three questions cut through most of what gets called “innovation” in this space. What decision does it change, and what evidence shows that the change improves patient safety, data integrity, or development efficiency in the target context? What are the failure modes, and how are they detected, documented, and corrected under GCP and inspection expectations? How does it integrate into real workflows, including incentives across sponsors, CROs, and sites, so that it reduces burden instead of shifting it? Tools that answer those questions well are more likely to be adopted, defended, and sustained in regulated clinical research.