We've updated our Privacy Policy to make it clearer how we use your personal data. We use cookies to provide you with a better experience. You can read our Cookie Policy here.

Advertisement

Building Regulatory-Ready AI Workflows for Drug Development

Two female researchers looking at a computer screen, reviewing data in a clinical setting.
Credit: iStock.
Read time: 5 minutes

The use of AI in early drug development is becoming commonplace, with four out of five pharmaceutical researchers now using AI tools to support their work in some capacity.


But there has been greater hesitation to adopt AI in clinical trials for several reasons: incomplete data can limit the reliability of tools; hallucinations and flawed outputs raise concerns about liability and overreliance; and regulatory uncertainty leaves investigators unsure how to implement the technology safely.


At the same time, when used well, AI has the potential to answer several key questions that can improve trial efficiency and inform future drug development: Where should the trial be held? How should it be conducted? What are the early indicators of success or failure? And how can that data be fed back to the drug discovery stage to inform future efforts?


As the use of AI in clinical trials increases, Technology Networks spoke with Shekhar Thumake, head of AI at Ephicacy, to explore how the technology can be applied while keeping regulatory compliance and patient safety front of mind.

Izzy Hirst (IH):

How is AI being applied in clinical trials and what is important from a regulatory perspective?


Shekhar Thumake (ST):

In the clinical data management and biometrics space, at a granular level, the most valuable applications of AI in clinical trials are not isolated chatbots. They are targeted interventions across workflows where people repeatedly interpret documents, transfer information, apply rules, and verify outputs. For example: building electronic case report forms from protocols; automation, from query management to data reviews; and regulatory document authoring, such as writing clinical study report documents


In real-world evidence, AI is used to mine unstructured clinical data for evidence generation, support synthetic or external control arms, and to convert unstructured data into traceable, queryable, AI/machine learning-ready assets using agents.


The differentiator, however, is not the number of agents or the complexity of orchestration. Many organizations are trying to solve multiple use cases through large multi-agent frameworks. Ephicacy’s focus is workflow agility, predictable performance, and biometrics-grade control. An AI system authoring SDTM [Study Data Tabulation Model] specifications, for example, should not work in isolation and then leave a reviewer to find every error. It should request expert input at defined decision points, record those decisions, and allocate clear task boundaries and decision rights between people and AI.


SDTM

The SDTM is a standardized format for organizing clinical trial data prior to regulatory submission. It is required by major regulatory agencies, including the US Food and Drug Administration. Historically, converting raw data into SDTM formats has been a manual, labor-intensive, long process.


This is more rigorous than simply placing a human “in the loop” or “in the lead.” People must actively steer the process where domain judgement changes the outcome. That design reduces reviewer fatigue, improves trust, and prevents human oversight from becoming a bottleneck to adoption.


For regulated use, we must also engineer agent action tracing, context management, reproducibility, GxP-compliant validation, and consistent handling of unseen cases.


Unseen AI cases

AI tools produce an output based on input, using pattern recognition and following predetermined instructions. Unseen AI cases are those in which the AI is unfamiliar with the input provided and faces a novel scenario. This can result in the tool breaking down or “hallucinating” and producing fabricated outputs. Unseen AI cases can be handled proactively with defensive engineering.



IH:
What scientific, technical, and governance challenges arise when using AI to extract insights from clinical documents? 

ST:

The first challenge is scientific; clinical documents were created to support care, not machine interpretation. Clinicians use inconsistent abbreviations, acronyms, symbols, and writing styles across institutions, departments, and individual practices. Notes may contain typographical errors, outdated information, and copied-and-pasted text that propagates irrelevant history.


“An AI system must also interpret nuanced medical context, including negation, temporality, uncertainty, and attribution. For example, it must distinguish “no fever” from a confirmed symptom, and a patient’s condition from family history. Converting this narrative into structured fields can remove precisely the context that gives the data meaning, such as when an event occurred, why therapy changed, or whether two events were causally related.” — Dr. Shekhar Thumake.


The technical challenge is therefore not simply extracting entities; it is preserving clinical meaning, provenance, and confidence while producing outputs that remain traceable to the source.


Governance must be designed into the workflow. Sensitive clinical text requires strong privacy controls, management of re-identification risk, appropriate consent, secure deployment, and compliance with data protection requirements.


Equally important, organizations need defined decision rights: what the AI may extract or infer, when a qualified expert must intervene, how corrections are recorded, and who is accountable for an error. Responsibility remains legally and operationally ambiguous across clinicians, healthcare organizations, sponsors, and technology vendors.


Trustworthy AI extraction means more than deploying a capable model. The objective is not to remove human judgment, but to focus it on consequential decisions while AI reduces repetitive effort without weakening scientific integrity or regulatory accountability.



IH:
To what extent can AI overcome challenges with unstructured data, or does structure need to be built in from the start for the best results?

ST:

AI can substantially improve the use of unstructured healthcare and clinical trial data, but it does not remove the need for structure. Modern language models can identify clinical concepts, interpret narrative context, summarize longitudinal histories, and transform information from sources into usable data. This can reduce manual abstraction and make previously inaccessible information available for clinical development and real-world evidence generation. That’s the reason why use cases like building electronic data capture screens from unstructured protocols are possible today.


However, the best results come from combining AI with structure by design. In clinical trials, standardized protocols, controlled terminology, consistent metadata, and clearly defined data models give AI a reliable framework for interpretation. Without this foundation, the model must infer more from inconsistent language and incomplete context, increasing the risk of ambiguity, silent errors, and outputs that are difficult to validate.


The practical goal should not be to force every source into a rigid format or expect AI to solve poor data quality after the fact. A hybrid approach is more effective: introduce structure at data capture wherever feasible (e.g., mining unstructured electronic medical record data to build structured patient profiles for queryable insights), then use AI to interpret the narrative content that remains necessary for clinical care and research.



IH:

With this in mind, how can protocol digitization and standards support successful AI adoption in clinical trials?


ST:

Protocol digitization gives AI a reliable, machine-readable foundation for clinical trial automation.


Today, much of the protocol remains locked in narrative documents, forcing teams to interpret and re-enter the same information across study design, data management, statistical programming, operations, and regulatory workflows. Converting legacy and new protocols into structured data based on the Unified Study Definitions Model (USDM) creates a consistent representation of study objectives, endpoints, eligibility criteria, activities, schedules, and design decisions.


USDM

The USDM is a data standard that includes all parts of a clinical trial, representing components in a single machine-readable model. Previously, clinical trials were designed and conducted on paper, and data had to be re-entered across different stages. USDM was developed around a “write once, read many” model, aiming to reduce the time required to re-enter the same data repeatedly.


Once the protocol becomes a governed dataset, pharmaceutical companies can apply protocol intelligence across the study lifecycle. They can compare designs across programs, optimize inclusion and exclusion criteria, assess protocol complexity and patient burden, identify operational risks earlier, and evaluate how design choices affect recruitment and execution. Structured protocol data also supports digital data flow by enabling information to move into downstream systems without repeated manual interpretation and transcription.


AI strengthens this model by extracting and mapping content from legacy protocols, identifying missing or conflicting information, and supporting evidence-based design decisions. However, the value does not come from digitization alone; the output must preserve semantic meaning, provenance, and other attributes, so that experts can verify decisions.


Standards such as USDM, therefore, provide more than interoperability. They establish the common language and traceable data foundation that AI needs to operate predictably across clinical development. This combination
can make trials more efficient, intelligent, and responsive, helping sponsors shorten development timelines and bring innovative, affordable therapies to patients sooner.



IH:
What are the next steps required to build trust in AI in clinical development?

ST:

Trust in AI for clinical development will not come from better model accuracy alone. It will come from demonstrating that AI-enabled workflows remain scientifically sound, operationally predictable, and accountable under regulated conditions.


The next step is to move from broad AI principles to use-case-specific assurance.

 

“An AI system supporting medical monitoring should not be governed in the same way as one that drafts SDTM specifications. For each use case, sponsors must define the intended purpose, risk level, acceptable performance, prohibited actions, and decisions reserved for qualified experts.” — Dr. Shekhar Thumake.


Validation must also evolve. A one-time benchmark is insufficient for probabilistic systems that may encounter new protocols, therapeutic areas, data patterns, and operating environments. Organizations need representative test sets, scenario-based testing, predefined acceptance thresholds, monitoring for drift and bias, and controlled revalidation when models, prompts, agents, source data, or connected systems change. Output and material decisions should remain reproducible and traceable to their source, model version, instructions, intermediate actions, and human inputs.


Human participation must be engineered into the value chain, not added as a final approval step. Experts should guide AI at decision points where clinical interpretation or regulatory judgment changes the outcome. Clear task boundaries allow AI to handle repetitive work while people focus on consequential decisions. This approach is more sustainable than asking reviewers to continuously inspect large volumes of AI-generated work.


“Trust requires transparency between all parties. Shared evidence on limitations, failure modes, validation methods, and post-deployment performance will be more valuable than unqualified claims of automation.” — Dr. Shekhar Thumake.


Ephicacy’s perspective is that dependable AI in clinical development must combine domain-led workflow design, GxP-aligned validation, action traceability, strong data foundations, and accountable human-AI collaboration. Trust will be earned workflow by workflow through consistent evidence that AI improves efficiency without compromising patient safety, data integrity, or regulatory confidence.



Google News Preferred Source Add Technology Networks as a preferred Google source to see more of our trusted coverage.