Biostatistics in the age of AI: Judgment, innovation and the human imperative

A combined position paper for biotech or pharmaceutical sponsor biometrics and clinical development organizations

*Suman Bhattacharya, Ph.D., Melissa Polansky, MPH, and Jeffrey Rieske, MS, MBA*

Abstract

Artificial intelligence is reshaping drug development and the practice of biostatistics. While AI enables gains in speed and scale, it also risks eroding the statistical judgment that underpins scientific credibility, and regulatory confidence. It is essential to leverage human judgment in an increasingly automated environment.

AI does not diminish the need for biostatistical expertise. As routine tasks are automated, the role shifts toward interrogating assumptions, adjudicating ambiguity and translating evidence into clinically meaningful and regulatory-defensible conclusions. These functions cannot be delegated without compromising evidentiary integrity.

This paper presents a unified framework for biostatistics leaders. It addresses the judgment gap in AI-assisted workflows, advances responsible methodological innovation within regulatory constraints, and calls for repositioning methods groups as scientific co-designers. It also defines an expanded competency profile for next-generation biostatisticians.

Biostatistics is not at risk of replacement, but of marginalization if human judgment is undervalued.

Introduction

Artificial intelligence (AI) is reshaping the biostatistics profession at a pivotal moment. The same computational advances that promise to accelerate drug development also risk eroding the statistical judgment that makes that development credible. This position paper synthesizes two complementary perspectives into a unified framework for organizations navigating this transformation; the first addresses the strategic and organizational imperatives for the biostatistics profession, the other examines the evolving human role of the biostatistician within AI-enabled clinical development workflows.

In this paper, we argue that the adoption of AI in clinical development does not diminish the need for statistical expertise. Rather, it necessitates that biostatisticians adapt to a “human-in-the-loop” model, where they retain accountability for statistical judgement, interpretability and regulatory defensibility for tasks such as defining estimands, adjudicating edge cases, stress-testing methodological assumptions and translating complex results into meaningful narratives. Organizations that combine technological ambition with deliberate investment in human judgment will not merely adopt AI; they will define what rigorous, trustworthy biostatistics means in the decades ahead.

PART I: PROBLEM DEFINITION

1. The judgment problem: Building what was once learned by doing

For a generation of biostatisticians, statistical judgement was forged through repetitive, hands-on engagement with data, such as coding in Statistical Analysis System (SAS) from scratch, manually debugging output and proactively understanding why models fail. The next generation of biostatisticians has arrived with powerful tools and broad computational fluency, but far less formative friction. While new tools enable substantial advances in the field of biostatistics, their ease of use may also weaken the intuition needed to proactively detect errors. This gap is compounded by a documented widening of the AI skills shortage across pharma biometrics functions, with organizations reporting difficulty sourcing statisticians who combine deep methodological expertise with AI literacy. As a result, the next generation of biostatisticians requires deliberate cultivation of the statistical judgement that was once built naturally through hands-on practice.

1.1 Where human biostatistical expertise remains non-delegable

Despite rapid advances in AI enablement, several core dimensions of the biostatistician’s role remain inherently human. In clinical development, a human‑in‑the‑loop model means that AI systems may generate analyses, suggestions or documentation, but biostatisticians retain responsibility for methodological choices, interpretation and regulatory accountability. Foundational decisions around study design, estimand definition and overall analysis strategy require expert judgment to balance statistical rigor, clinical relevance, ethical considerations and regulatory expectations [2]. These decisions are context-dependent and often involve ambiguity, incomplete information and trade-offs that exceed the capabilities of automated systems. While AI can surface patterns, options and inconsistencies, ultimately, interpretation and accountability rest with the biostatistician.

1.2 The nature of statistical judgment

Statistical judgment is not a checklist, but rather a pattern recognition capacity built over years of encountering failures, anomalies and edge cases. In AI‑assisted environments, statistical judgment functions as the human‑in‑the‑loop mechanism that evaluates whether automated outputs align with clinical intent, causal logic and regulatory expectations. It includes:

• Recognizing when assumptions are violated even when diagnostics pass

• Knowing which questions to ask before seeing the data

• Identifying when a perfectly valid model produces a clinically meaningless estimate

• Understanding the difference between what is statistically defensible and what a regulator will accept

• Sensing when a sponsor’s framing of a question embeds an analytical choice that advantages a result

1.3 How to build judgment in an AI-assisted environment

As we usher in the new generation of biostatisticians and begin to incorporate AI-assisted tools, we will also transition into an environment that loses the formative friction we could previously rely on to build strong statistical judgement. Statistical judgement must now be formed with more intention across both educational training and professional development. The following approaches are recommended:

Structured retrospective analysis: Require junior statisticians to review and critique historical analyses, including ones that failed, were rejected, or produced counter-intuitive results. Strong statistical judgement requires the ability to identify whether a final product is reliable, reasonable and meaningful. The goal is exposure to the full life cycle of statistical decision-making, not just the polished final product.

Deliberate constraint exercises: Just as airplane pilots train without autopilot, biostatisticians should periodically work without AI assistance. This can include assigning analyses to be completed without AI-assisted code generation or automated diagnostics. The purpose is not to simulate the past but to build a mental model of what the machine is doing.

Assumption audit culture: Create a team culture of explicitly stating and challenging every assumption in every analysis. This is not about distrust of AI; it is about the discipline of articulating the causal reasoning that no algorithm can supply. With practice, biostatisticians will learn to maintain a critical eye of AI-generated statistical output and sharpen their statistical judgement.

AI output review protocols: Establish formal review checkpoints where AI-generated code, model selection and outputs are interrogated by senior statisticians. Document the questions asked, not just the answers accepted. This creates a visible trail of judgment that also serves as training material.

2. AI across the biostatistical lifecycle

AI augmentation now spans the full lifecycle of biostatistical leadership in clinical trials, supporting both scientific rigor and regulatory confidence. Understanding where AI adds value and where human expertise remains irreplaceable is foundational to strategic deployment.

2.1 Where AI augments biostatistical work

Trial design and decision optimization

AI-enabled simulation of sample size assumptions, adaptive and innovative designs and endpoint sensitivity analyses allows biostatisticians to more rapidly evaluate trade-offs, anticipate risks and optimize designs aligned to clinical and regulatory objectives. The World Economic Forum has identified generative AI’s capacity to accelerate trial design as one of five transformative levers for clinical development, noting that the journey from lab to market, which currently costs over $2.5 billion and takes approximately 8–12 years, demands precisely this kind of design-phase efficiency. Similarly, emerging LLM-based tools can generate and evaluate randomized controlled trial designs, including eligibility criteria and intervention arms, with accuracy validated against data from ClinicalTrials.gov.

Data quality oversight and signal awareness

Continuous AI-driven surveillance supports earlier detection of anomalies, emerging safety or efficacy signals, and data integrity issues. This enables biostatisticians to focus on root-cause assessment and proactive trial steering rather than reactive data cleanup.

Analysis execution with statistical accountability

While automation accelerates routine analysis execution and output generation, biostatisticians retain ownership of analytical decisions, ensuring results remain interpretable, defensible and fit for purpose. Leveraging agentic AI systems may enable biopharma organizations to run twice as many trials with the same resources by automating various workflows.

Interpretation, storytelling and cross-functional translation

AI-assisted visualizations and plain-language summaries help biostatisticians translate complex statistical results into clinically meaningful insights for clinicians, development teams and governance bodies, strengthening evidence-based decision-making.

Submission readiness and regulatory traceability

AI supports preparation of submission artifacts through consistency checks, traceable metadata and structured drafting of statistical sections, allowing biostatisticians to concentrate on scientific narrative coherence and regulatory scrutiny. McKinsey and Merck have co-developed an AI-powered platform that reduced first-draft clinical study report writing time from 180 to 80 hours while cutting errors by 50%, demonstrating the scalability of AI-assisted regulatory authoring.

2.3 Data foundations as a prerequisite for credible AI

For biostatisticians, the value of AI is inseparable from data quality, traceability and standardization. Robust adherence to CDISC standards, transparent data lineage and interoperable trial systems are essential to maintaining confidence in AI-assisted analyses, particularly in regulatory and submission contexts [2]. Without a strong statistical and data governance backbone, AI risks accelerating noise rather than insight. Achieving sustained value requires moving beyond static data environments toward continuous, AI-ready data flow, where data can be ingested, structured and analyzed in near real time across the clinical lifecycle. This shift enables a “Data at the Speed of Light” paradigm, where decision-making is no longer constrained by fragmented systems or delayed data availability.

PART II: IMPLEMENTATION

3. A three-phase evolution of AI-enabled biostatistics

Organizational adoption of AI in biostatistics is not a single transformation but a progression across three distinct phases. Understanding this trajectory enables leaders to sequence investments, manage change and set realistic expectations.

Phase 1: Assisted productivity

Focus on automating well-defined tasks such as SAP drafting assistance, code scaffolding and structured validation checks. Statisticians build AI fluency while maintaining full oversight. Expected outcomes: 15–25% efficiency gains in programming and documentation workflows.

Phase 2: Integrated agentic workflows

AI becomes embedded across end‑to‑end biostatistical and cross‑functional workflows, coordinating multiple tasks such as continuous data surveillance, automated generation of analysis‑ready datasets, document drafting and interim reporting. Multi‑agent systems handle routine coordination and execution across functions. Human oversight shifts from step‑by‑step review to monitoring, exception handling and interpretation when outputs deviate from expectations. Expected outcomes: 20–30% reduction in rework and faster interim analysis readiness.

Phase 3: Enterprise statistical intelligence

AI evolves from workflow execution to enterprise‑level pattern recognition and foresight. Systems synthesize information across studies, programs and historical trials to surface emerging safety trends, cross‑study signals and predictive indicators of design or submission risk. Biostatisticians no longer supervise tasks; they lead governance, methodological strategy and judgment‑heavy adjudication of AI‑generated insights. Decisions increasingly focus on prioritization, risk trade‑offs and scientific meaning rather than operational execution. Expected outcomes: accelerated database lock, earlier insight generation and scalable study throughput [1].

4. Driving methodological innovation under regulatory constraint

Biostatistics innovation operates in a paradoxical environment: the profession is technically capable of moving faster than ever, while regulatory frameworks, institutional risk aversion and trial conservatism create powerful headwinds. A strategic priority is understanding how innovation happens and how to accelerate it responsibly.

4.1 Active frontiers of innovation

Across the pharmaceutical and biotech industries, AI and machine learning are reshaping methodological practice on multiple fronts simultaneously:

• Estimands Framework (ICH E9 R1): Regulators now expect explicit estimand specification. AI is enabling rapid interrogation of historical trials and protocol scenarios to stress-test estimand definitions, increasing precision and consistency in aligning clinical questions with analysis strategies.

• Bayesian Adaptive Designs: FDA and EMA guidance has expanded. AI-enabled simulation and optimization allow more efficient exploration of adaptive design scenarios, accelerating feasibility assessment and improving operating characteristics, particularly in early development, rare diseases, oncology and pediatric trials.

• External Controls and Synthetic Arms: The use of real-world data as external controls is methodologically rich and regulatorily contested. AI is expanding the ability to identify, match and weight external cohorts at scale, while simultaneously introducing new risks around bias amplification that require rigorous statistical oversight.

• AI-Assisted Safety Signal Detection: Machine learning applied to pharmacovigilance is advancing rapidly. AI enables continuous, large-scale surveillance of safety data and earlier signal detection, but translating these signals into regulatory-grade evidence requires new validation frameworks and statistical rigor.

• Causal Inference Methods: Causal DAGs, g-estimation and target trial emulation are moving from academic literature to applied trial design. AI is accelerating variable selection, hypothesis generation and large-scale observational analyses, making these approaches more operationally feasible in hybrid and real-world evidence contexts.

• LLM-Assisted Protocol and SAP Development: Emerging evidence shows that large language models can generate randomized controlled trial design elements, including eligibility criteria, recruitment strategies and outcome definitions with accuracy approaching expert performance [2]. This shifts effort from drafting to critical review, validation and methodological refinement.

4.2 A Tiered framework for responsible innovation

Not all methodological innovations carry equivalent risk or regulatory readiness. A four-tier approach is recommended:

• Tier 1 — Exploratory: Build the evidence base with no regulatory risk through internal pilot studies, simulation work, academic publication.

• Tier 2 — Pre-competitive: Develop standards without exposing individual sponsors via industry consortia and collaborative methods development.

• Tier 3 — Regulatory Engagement: Validate novel methods with agency input before pivotal use using pre-submission meetings and qualification procedures.

• Tier 4 — Confirmatory: Demonstrate method maturity through primary endpoint application in Phase III.

Many organizations struggle because they try to move directly from early experimentation to full-scale use (Tier 1 to Tier 4). Sustainable innovation requires investment in Tiers 2 and 3, which are under-resourced in most pharma biostatistics groups. Organizations spreading AI capabilities beyond data science teams into R&D and commercial functions are better positioned to realize innovation at scale.

4.3 Engaging regulators as innovation partners

The FDA’s Complex Innovative Trial Designs (CID) program and EMA’s Innovation Task Force represent genuine openings for methodological dialogue. The FDA’s 2025 draft guidance on AI in regulatory decision-making establishes a risk-based framework for AI model credibility, emphasizing transparency, validation and data governance as prerequisites for regulatory acceptance of AI-assisted methods [2]. Recommendations for engagement:

• Participate in public comment periods on draft guidance documents

• Submit methodological papers to journals

• Attend and present at joint regulator-industry scientific conferences (e.g. DIA, ASA Biopharmaceutical Section, FDA-industry workshop)

• Request Type B or C meetings to discuss novel methodology before committing to pivotal use

PART III: INSIGHTS FOR LEADERS

5. Building high-functioning methods groups

A methods group, whether formally constituted or informally organized, is the organizational mechanism by which a biostatistics function develops, evaluates and operationalizes statistical methodology. They provide a collective structure for translating emerging methods into practice, ensuring consistency, scientific rigor and regulatory alignment. In the AI era, methods groups face new challenges and have new capabilities.

5.1 The four functions of a methods group

• Scientific advancement: Developing and evaluating novel methods

• Standardization: Establishing organizational position on acceptable approaches

• Education: Transferring methodological knowledge across the function

• Regulatory intelligence: Monitoring and interpreting evolving guidance

Many methods groups conflate standardization with scientific advancement, producing conservative guidance documents that protect the organization but suppress innovation. The best groups separate these functions explicitly.

5.2 Structural recommendations for methods groups

Dual-track operating model: Maintain a standards track for validated, guideline-ready methods and an innovation track for methods in evaluation. Members of the innovation track should be protected from the pressure to produce deliverables — methodological exploration requires slack.

External collaboration mandate: Methods groups that operate in isolation calcify. Every methods group should have formal mechanisms for external engagement: academic collaborations, consortium memberships, secondments to regulatory agencies and active publication programs. The goal is to remain a net importer of ideas.

AI as a methods tool: Methods groups should lead the internal adoption of AI-assisted simulation, literature synthesis and code validation — not waiting for the broader organization to set the agenda. They should also critically evaluate AI methodological tools, asking whether a machine learning approach performs better or only differently than established methods.

5.3 Addressing the talent challenge

Methods groups are facing a growing talent shortage. A 2025 analysis of the pharma AI skills gap found that organizations recruiting statisticians with both deep statistical methods expertise and AI fluency face a market that competes directly with technology firms and academia, with recruitment costs running up to 50% above standard roles. Three interrelated challenges require attention:

• Recruitment: Statisticians with both deep methods expertise and AI fluency are rare and in high demand. Organizations using AI extensively achieve up to 14× lower quality costs than peers, creating strong incentives to close this gap.

• Retention: Methods work can feel distant from trial delivery; visible impact and career recognition are essential.

• Succession: Senior methodologists carry institutional knowledge requiring deliberate investment in apprenticeships programs to transfer organizational insights.

6. Leadership imperatives for transformation

Successful AI-enabled transformation in biostatistics requires more than technology adoption. Leaders must link AI to scientific rigor and patient impact while advancing governance, validation and change management alongside technical capability. Targeted investment in real-world use cases is critical. Roles will shift from execution to orchestration, interpretation, and oversight, requiring leaders who balance technological ambition with protection of human statistical judgment.

Defining the biostatistician’s non-delegable role and communicating it to regulators, partners and leadership is the central governance challenge of the next five years [2]. As AI evolves, the boundary between automation and human judgment will require continuous negotiation led by biostatistics leaders.

7. Training and development implications

Current training models have not kept pace with the evolving competency requirements of biostatistics. Graduate education continues to underemphasize regulatory science, communication and AI literacy, creating gaps that must be addressed through organizational development and on‑the‑job learning. Building the next generation of biostatisticians therefore requires structured investment beyond formal academic preparation.

Organizational development should combine cross-functional rotations, project leadership opportunities, external engagement and structured mentorship to build statistical judgment, influence and strategic thinking. Embedding statisticians across regulatory, clinical, and data functions deepens context, while conference participation and publication prevent methodological insularity. Mentorship that explicitly surfaces reasoning and trade-offs accelerates development. Together, these investments ensure that growing technical capabilities are matched by the judgment, context and regulatory awareness required for effective statistical leadership.

7.1 Core competencies for the next-generation biostatistician

The competency profile for biostatisticians is expanding faster than formal training pathways are adapting. While the foundations of the discipline remain essential, modern practice now requires deeper engagement with regulation, computation and AI‑enabled workflows. The core competencies below represent the minimum skill set required for effective practice in contemporary clinical development.

• Statistical theory: Core foundations remain assessment. Human judgement is required to assess validity and limitations.
• Regulatory science: Estimands, guidance interpretation and submission strategy are now central competencies.
• Programming and computation: Proficiency in R, Python and reproducible workflows is essential. AI‑generated code must be reviewed and validated.
• AI and machine learning literacy: Understanding model evaluation, bias and interpretability is critical for effectively leveraging AI.
• Communication: Clear writing, data visualization and scientific communication is essential. Analytical judgment only has value if it can be communicated persuasively and accurately.
• Clinical domain knowledge: Therapeutic area expertise and endpoint science guide sound statistical decisions.
• Data engineering: Skills in data structures, database querying and data quality assessment is increasingly important for complex data environments [2].

Conclusion: Biostatistics’ strategic moment

Biostatistics is not at risk of being replaced by artificial intelligence. Rather, it is at risk of being marginalized if the profession fails to adapt, either by clinging to methods and workflows that AI has rendered obsolete, or by abdicating scientific judgment to algorithmic systems that cannot bear it.

Simultaneously, AI elevates the value of statisticians who understand why analyses are correct and can defend them. The priority is to ensure the next generation develops this depth of judgment, not just technical proficiency.

This requires intentional investment in building statistical judgment, engaging regulators to enable innovation, repositioning methods groups as scientific co-designers, and expanding skill sets to include AI literacy, regulatory science and communication. Organizations that balance digital acceleration with human capability development will define the future standard for rigorous, trustworthy evidence generation.

Read more insights from ZS.

This article reflects our personal views. They do not necessarily represent any official position of ZS.