Lessons from Hundreds of Clinical Trials: AI’s Successes, Gaps, and Opportunities

The industry now has models capable of drafting study documents, reviewing data, generating code and summarizing operational signals. Yet adding these capabilities does not automatically make a trial faster, less expensive or more reliable. Each task can move more quickly while the trial itself moves at the same speed.

This challenge can be seen from all three sides of clinical development: the global biopharmaceutical company, the clinical technology provider and the contract research organization. Each vantage point reveals the same underlying problem. Clinical development has accumulated disconnected systems, study-specific processes, multiple vendors and incentives that do not always reward the same outcome. AI is entering that environment one tool at a time.

This article will examine four lessons that emerge across these settings: The model is no longer the main constraint for many useful applications; customization creates downstream costs that study teams often underestimate; automating individual tasks does not shorten a trial unless decisions and workflows are connected; and AI should change the work of clinical experts without removing their accountability.

The model is no longer the main constraint

The first wave of clinical AI focused on whether a model could automate a task that is part of a workflow: draft a protocol section, propose data queries, write statistical code or summarize a monitoring report. For many bounded uses, the harder question now is: What must change for the model to produce an operational result?

Today’s large language models can interpret metadata and reduce manual data-harmonization work, but an agent still needs to know what a field means, which definition is authoritative and whether the data is fit for its intended use. AI does not eliminate the need for a governed data layer, shared definitions and traceability.

AI and automation depend on high-quality, structured protocol data because unreliable inputs limit predictions about study design, patient burden, site selection and operational risk.¹ The question is no longer simply whether an organization has data. It is whether the data carries enough consistent meaning for people and machines to act on it and generalizes across studies and programs.

Customization is not free

Novel endpoints and unusual patient populations may require study-specific approaches, but customization has a compounding cost.

Consider a sponsor that requests a bespoke case report form field where an established standard could answer the same question. The change may appear small, but it can require new specifications, mappings, edit checks, validation and downstream transformations. A different definition also makes historical data and agents built around standard structures harder to reuse.

What looks like flexibility in one study can become AI debt across a portfolio. Not every trial belongs in an identical template, but exceptions should be intentional. Teams should weigh the value of departing from a standard against the loss of interoperability, automation and reuse.

This becomes more important as the protocol moves from being a document to becoming a structured source for study execution. Drug development experts have described the protocol as the defining document at the center of a clinical trial and outlined the industry’s movement toward digital protocol approaches.² The more consistently protocol concepts are represented, the more reliably they can drive downstream study build, analysis and reporting.

This example clearly illustrates an unrealized use case for AI: Can a model’s data harmonization capabilities be leveraged during the study’s operations phase to simplify downstream activities like data review, reporting, and transformation into analysis-ready formats? A truly connected AI solution would do this, allowing the sponsor to introduce the bespoke case report form field from our example without paying the high price of customization.

Faster tasks do not necessarily produce a faster trial

The most common AI business cases measure local productivity: hours saved drafting a document, generating code or reviewing data. Those measures are useful, but they can be misleading. Saving 500 hours does not matter to a sponsor if database lock is not any sooner.

A protocol amendment illustrates the difference. AI may help draft the revised language in hours, but the change may also affect case report forms, edit checks, randomization logic, site instructions, training materials, the statistical analysis plan and regulatory documents. If different functions and vendors interpret and implement the amendment separately, the trial can continue operating from inconsistent assumptions even though the first task was completed faster.

The larger opportunity is to shorten the distance between a decision and coordinated action. An AI-enabled workflow should identify the downstream implications of a change, propose aligned updates, preserve traceability, surface exceptions and direct them to accountable people. The value is not another document; it is fewer days of latency, fewer back-and-forth emails and less rework across the study. Unfortunately, many sponsors did not deliberately design their current clinical architecture with the end-to-end in mind. It is accumulated through purchasing decisions, mergers and study-level exceptions. The result can be platforms and data structures that work within individual functions but require substantial effort to connect.

There is also an economic concern. When AI enters the budget as an additional product while platform fees, service contracts and review processes remain unchanged, local productivity does not automatically become systemwide savings. Sponsors and partners should define who owns the end-to-end outcome and how faster work will shorten the critical path.

Kenneth Getz, MBA, and Kenneth Kaitin, PhD, of Tufts University School of Medicine recently described an “execution translation gap:” Organizations have become better at detecting trial problems, but fragmented accountability, misaligned incentives and rigid processes can prevent timely, coordinated action.³ AI can make signals arrive faster. Without clearer ownership, it may simply help the industry identify delays sooner.

Clinical trials do not need to be fully autonomous

The answer is not to automate the expert out of the trial. Clinical research needs controlled, semi-autonomous workflows in which AI performs defined work, shows its sources and escalates uncertainty to qualified and empowered decision makers.

The appropriate boundary depends on risk. An agent may draft routine site correspondence, compare documents for inconsistencies or prepare code for review. It should not independently determine the clinical significance of a safety signal or make an unexplained decision to exclude a patient’s data. Human review must be substantive, not ceremonial.

This changes the role of clinical experts. A biostatistician may spend less time manually coding repeatable transformations and more time testing assumptions, reviewing agent-generated code and interpreting results. A data manager may move from finding every inconsistency manually to supervising automated checks and resolving the exceptions that require context. AI contributes speed and pattern recognition while people provide strategic oversight, judgment and relationship management.⁴

The most useful measure of AI, then, is not how much content it produces but whether it improves the performance of the trial: fewer avoidable amendments, faster study build, more consistent execution, quicker exception resolution, better endpoint quality and earlier trustworthy decisions.

The lesson across hundreds of trials is that AI’s greatest opportunity is not to automate more fragments of clinical development. It is to connect them. That requires standardized and meaningful data, workflows built around decisions rather than functions, experts prepared for new roles and commercial and operating models that assign responsibility for the whole outcome. The technology is ready for many of these small incremental deliverables to be done more efficiently and effectively. The main challenge is adapting and transforming the clinical development ecosystem with redefined accountable decision makers equipped with AI rather than a stream of handoffs between small deliverables and task masters.

About the author

Rama Kondru, PhD, is chief executive officer of Emmes Group. He previously served as co-CEO of Medidata Solutions and as senior vice president and chief information officer for Janssen Pharmaceuticals, Americas, a Johnson & Johnson company. He also served as the global head of data sciences at Johnson & Johnson.

References

1. Mooney M, Whitney N, Woodruff C. “Beyond the Buzzwords: What SCOPE 2026 Revealed About Clinical Trial Planning and Operations.” Applied Clinical Trials. Feb. 24, 2026. https://www.appliedclinicaltrialsonline.com/view/scope-2026-clinical-trial-planning-operations
2. Georgieff T. “Navigating Toward a Digital Clinical Trial Protocol.” Applied Clinical Trials. Oct. 10, 2023. https://www.appliedclinicaltrialsonline.com/view/navigating-toward-a-digital-clinical-trial-protocol
3. Getz K, Kaitin K. “Recognizing and Addressing the Execution Translation Gap in Clinical Trials.” Applied Clinical Trials. April 17, 2026. https://www.appliedclinicaltrialsonline.com/view/recognizing-addressing-execution-translation-clinical-trials
4. Agrawal G, Studna A. “Human-AI Collaboration in Clinical Operations.” Applied Clinical Trials. Dec. 11, 2025. https://www.appliedclinicaltrialsonline.com/view/human-ai-collaboration-clinical-operations