Workflow in the Age of AI

“Workflow” consists of the network of business processes that organizations use to get their work done. For each clinical study, clinical research professionals work together in a complex network of interconnected business processes. Study workflow is part of a broader clinical development workflow that interacts with workflows across the organization and in other organizations, including study sponsors, CROs, clinical sites, solution providers, etc. It even extends to the study patients. In addition, these workflows interact with multiple forms of communication, which are, themselves, special types of business processes.

Clinical research is complex, inconsistent, fragmented, transactional and resistant to change. With experience, clinical research professionals learn the study protocols, standard operating procedures (SOPs), regulations, guidances, employee handbooks, and other documents that outline their business processes. These documents are supplemented by know-how, the knowledge that clinical research professionals develop over time to interpret all this documentation and put it into practice.

Artificial intelligence (AI) agents have started taking over various bits of workflow, often in the form of “AI teammates” or with a “human in the loop.” For example, an AI agent could analyze a study protocol and draft an informed consent form that a human could then polish into its final form.¹

In certain cases, AI agents can handle sequential steps in a business process, e.g., chatting with a potential study participant and then scheduling a visit with a study coordinator. In this example, it may be necessary to divert the chat to a human, e.g., to answer questions the AI agent cannot answer. This stage of the patient recruitment process thus becomes a small workflow involving the patient, an AI app, and a study coordinator. The key point here is that the AI app has not replaced the workflow; it has just made it more efficient for the study coordinator.

Workflow can involve moving documents, e.g., essential documents from a site’s eISF to the sponsor’s eTMF. An AI agent can download documents (“artifacts”) from the eISF and upload them to the eTMF nightly. Other AI agents can extract data from one database and insert it into another. These agents would be scheduled and managed as part of a workflow.

AI apps do not function in isolation. They are useful only when they advance a workflow, which is probably connected to many other workflows. In some cases, they can be strung together sequentially in their own mini-workflow. Over time, these sequences will become longer and require less human intervention, but they will still constitute workflow.

Today, we are in an AI era of “let a thousand flowers bloom.” (One large pharmaceutical company has over 200 employees vetting employee-developed AI apps.) Many employees developing home-grown AI apps are not thinking about workflow. Instead, they are thinking about how an AI app can solve a specific problem. Nevertheless, that specific problem exists in a workflow.

Workflow management

The emergence of AI agents begs the question: How do organizations, e.g., those involved in clinical research, manage their workflows? How do they ensure that their workflows are efficient, timely, compliant and reliable? How do they ensure that their personnel (and AI agents) are working on the highest-priority tasks, completing them in a timely manner, and properly supporting any connected workflows? How do they identify structural problems in their business processes and find ways to improve them?

Some organizations have standardized on workflow apps like Monday.com, accepting their mix of strengths and weaknesses. Others employ a variety of technologies, including workflow apps, the workflow built into other apps, checklists, tracking sheets, and know-how. In this scenario, there is no unified workflow management system that connects all these disparate tools and ensures that business processes flow smoothly across the entire organization and with business partners, who may have their own workflow management tools.

To complicate matters further, these workflows incorporate multiple forms of communication, typically a mix of face-to-face conversations, email, text messages, telephone calls, virtual meetings, and discussion groups, which are often ephemeral, usually hard to monitor, and never fully integrated.

As a result, management’s ability to see and control an organization’s workflows is very limited. Imagine a bunch of engineers designing the parts of a car with only a vague understanding of the car as a whole or even what a car actually is. That’s where workflow management stands today. Adding a swarm of AI agents may improve individual workflow steps, but it will exacerbate this broader problem.

A new perspective

AI agents are changing the way we look at information technology. We are moving from an app (e.g., CTMS)-centric perspective to an agent-centric perspective. We will increasingly interact with SaaS (Software as a Service) programs through agents—not necessarily AI-based—rather than as monolithic environments. In other words, SaaS products will become more of a resource than the world in which we live.

But what we really need is a workflow-centric perspective. We need a four-level IT architecture: Workflow - Agents - SaaS (optional) - Data.

A workflow-centric perspective encourages a systems approach to business operations. With a systems approach, each element is important only to the extent it affects the system. A workflow-centric perspective thus broadens the perspectives of personnel who develop their own AI agents; they must consider how a specific AI agent fits within their workflow and, preferably, within the entire system of workflows. Is this particular AI agent the one they should be creating (or acquiring)? Secondly, it encourages personnel to consider how their AI agent supports standards of quality and compliance that are apparent only at the process level. Thirdly, it provides a framework for categorizing AI agents so they can be discovered by colleagues and need not be reinvented.

A workflow-centric perspective comes naturally at the management level. An important part of many managers’ jobs is to find, prioritize and fix bottlenecks and other problems in business processes.

Example: Source document template builder

Helios Clinical Research has tested an early version of CRIO’s AI-based Source Template Builder. According to Rustin Rolen, director of study and site integration at Helios, “It has the potential to greatly reduce the time and resources needed to get eSource built and ultimately decrease study startup timelines.”

The process map below describes the life cycle of source document templates in a clinical study. (This process map is not the only possible version, and does not include study sponsor or CRO roles.) As can be seen, a source document template builder, while performing the most important and time-consuming sequence of steps in the process, is only part of the business process for creating source document templates, which, itself, is only part of the complete source document template life cycle, which is only a small part of clinical study business processes, etc. Short-change even a minor step in the process, and the entire workflow may break down.

Most steps in this process map are performed by humans, but one can imagine possible roles for AI agents. For example, an AI agent could collect feedback from stakeholders on draft versions. Another AI agent could periodically analyze data queries to identify problems and perhaps fix them.

Digital data flow

CDISC and TransCelerate are developing Digital Data Flow (DDF) to “modernize clinical trials by enabling a digital workflow with protocol digitization.” The initiative enables machine-readable versions of protocols in Study Data Tabulation Model (SDTM) format that can be used to create and drive clinical study workflows from study protocols to NDAs.^2,3.4

Conclusion

In the beginning, there was workflow. By definition, business-oriented AI agents—and business software in general—exist only to support business processes. The DDF initiative demonstrates broad support for a workflow-centric perspective.

Clinical research is complex, inconsistent, fragmented, transactional and resistant to change. A systems-oriented, workflow-centric perspective addresses all these challenges. It facilitates the identification, prioritization, design, creation and deployment of AI agents in the context of a business’s objectives. Over time, clinical research processes will increasingly incorporate AI agents. Some may, eventually, consist entirely of agents. But, in the end, there will always be workflow.

About the author

Norman M. Goldfarb is executive director of the Site Council and founder & CEO of Portolo. Previously, he was chief collaboration officer of WCG Clinical, founded and led the MAGI conferences, and published the Journal of Clinical Research Best Practices.

References

1. “Is Artificial Intelligence Coming for Clinical Research?” Norman M. Goldfarb, November 25, 2025, Applied Clinical Trials, https://www.appliedclinicaltrialsonline.com/view/artificial-intelligence-coming-for-clinical-research.
2. “Digital Data Flow,” CDISC, https://www.cdisc.org/ddf.
3. “Digital Data Flow,” TransCelerate, https://www.transceleratebiopharmainc.com/initiatives/digital-data-flow/
4. “Digitizing The Clinical Protocol: Small Steps For Seismic Change,” Kelsey Jakee and Rob DiCicco, Clinical Leader, January 2024, https://www.clinicalleader.com/doc/digitizing-the-clinical-protocol-small-steps-for-seismic-change-0001