The future of agentic AI in real-world evidence generation
As Datavant builds the agentic AI platform to accelerate the path from data to trusted evidence for Life Sciences organizations, it is critical not only to consider what can be built today, but also to consider where the field of real-world evidence (RWE) is headed as agentic AI evolves. Today, task-oriented AI agents support discrete parts of the RWE workflow, such as code list development, feasibility assessment, literature review, and protocol drafting. As agentic RWE systems develop, they will enable organizations to generate RWE more efficiently, continuously, and at greater scale. As they mature, they will take on more advanced work: designing studies, identifying sources of bias, recommending alternative analytic methods, and synthesizing evidence.
These agentic capabilities have the potential to support more seamless and continuous evidence generation—the ongoing cycle of 1) identifying evidence gaps, 2) articulating a research question, 3) assessing the feasibility of answering the research question in available data source(s), 4) implementing the study with a valid design & analytic approach, 5) interpreting and synthesizing the resulting evidence, 6) understanding its implications, and 7) articulating additional research questions that reflect newly identified evidence gaps. Future RWE agents theoretically could drive dynamic design, analysis, and data assessment, alongside ongoing evidence synthesis. Instead of treating evidence generation as coming from a discrete study to be optimized in terms of design and feasibility, continuous evidence generation will be accelerated and more coordinated through the establishment of agentic systems throughout.
For Life Sciences organizations, this shift will require changes to evidence-generation strategies to incorporate agentic reasoning and evidence synthesis. Adopting agents for individual tasks is the starting point; organizations will need integrated agentic approaches while continuing to apply frameworks and workflows that support rigorous, transparent, and principled generation of RWE. Below, we describe several RWE agentic systems which may exist in the future and the implications for the field.
The potential for RWE agents
Agentic systems will make discovering and sourcing real-world data (RWD) faster and more scalable. Life Sciences organizations will be able to more seamlessly build fit-for-purpose datasets through privacy-preserving data linkage and stacking, with the ability to update them over time as data accrue. These agentic data discovery and data sourcing systems will support more iterative study design, especially around cohort eligibility, available outcomes and follow-up times to ensure robustness and feasibility. Furthermore, as data accrue, agentic systems may support the updating of study implementation at a regular frequency.
Causal inference agents may support epidemiologists by helping identify potential biases and methodological limitations within a given study. Agents may flag design choices and data limitations that could introduce immortal time bias, selection bias, measurement error, or residual confounding. They may also conceive the hypothetical target trial and define the estimand based on the literature and the study objectives as a benchmark for robust study design. In addition, causal inference agents can search the literature for potential confounders, minimize inappropriate variable control, augment researchers’ existing causal knowledge, and recommend sensitivity analyses to address known and unknown limitations. As these tools mature, they may also support quantitative bias analysis, including estimating the potential direction and magnitude of bias in the observed effect estimate.
Evidence synthesis agents will expand what is possible before, during, and after a RWE study. Prior to study conception, they will be able to identify relevant evidence gaps within a specific indication or product area and from there, articulate potential research questions. During study design and implementation, agents will reveal additional relevant evidence as it is published in the peer-reviewed literature, creating opportunities to refine study objectives, update populations and subgroups of interest, revisit design assumptions, and articulate new analyses that address evolving evidence gaps. Following generation of primary study results, agents will synthesize study outputs – including tables, figures, and primary results – and suggest relevant post hoc analyses that may help explain or contextualize findings. They will also aid in the interpretation of study results incorporating the latest and most relevant published literature, with suggestions for implications for the next evidence gap to fill. Over time, these systems may also help connect RWE results to clinical development, regulatory, market access, or commercial strategy by clarifying what the findings mean for a product team and what evidence generation activities should come next.
Implications
The introduction of data discovery, causal inference, and evidence synthesis agents into real-world evidence workflows has wide-ranging implications for Life Sciences organizations. These systems may help reduce bias and can support more informed evidence generation across the product lifecycle. However, agents must be deployed intentionally within existing RWE workflows, with clear alignment with established scientific best practices and stakeholder needs, such as those stipulated by regulators.
For example, an evidence synthesis agent could search and summarize the latest literature relevant to a study’s research objectives at initial design and throughout the study lifecycle. Based on that synthesis, the agent may suggest a change to eligibility criteria, an additional subgroup analysis, an alternative approach to balancing baseline characteristics between exposed and comparator groups, or a method to mitigate informative censoring. These recommendations may be valuable, but for certain downstream uses or stakeholders (e.g., regulators or for peer-reviewed publication), they must be incorporated into the study workflow in a way that preserves pre-specification, transparency, auditability, and objectivity. Given the ability to more seamlessly incorporate additional patients and events over time as data accrue, organizations will need clear decision rules for when to conduct primary analyses, particularly when sample size thresholds are reached for a prespecified effect size. Altogether, the roles of agents in continuous data discovery, data sourcing, and study execution raise important methodological considerations that need to be addressed in agent development and deployment, including the risks of type I error and cherry-picking in inferential studies.
As agentic systems become more knowledgeable and capable, RWE workflows will continue to evolve. Agents may become increasingly central to evidence orchestration, supporting both individual studies and more dynamic evidence-generation activities. This shift will create more space for subject matter experts (SMEs) to provide needed oversight, methodological governance, and value demonstration. RWE SMEs will be essential to determining when agent-generated recommendations are appropriate, when they require further scrutiny, and how resulting evidence should inform decisions in clinical development, portfolio strategy, market access, and commercialization.
Agentic AI will make rigorous evidence generation continuous, traceable, and scalable while preserving human scientific judgment. The organizations that win will have the best scientific operating system around their agents.
As Datavant transforms its platform to support agentic RWE study design, data discovery (via Explore Assistant), data utility, protocol development, variable operationalization, analytic cohort generation, and the implementation of RWE studies (e.g., natural history), we are building with an eye toward the future of agentic systems, particularly where and how human and agent SMEs play a role. Datavant will be releasing a series of subsequent blogs to explore this transformation in greater depth, both across our platform and within the broader field of RWE.
To learn more about how Datavant is building agentic AI capabilities to accelerate trusted real-world evidence generation, connect with our team here.
