[Adobe Stock]
FDA targets agency‑wide genAI by June
As the FDA transforms its internal processes, a natural question arises: What level of AI-driven efficiencies can the pharmaceutical industry achieve with AI the extensive clinical documentation it must prepare? For AI vendors like Yseop, who partner with major pharma companies, there is already significant precedent. Timothy Martin, VP Product at the company, points to sizable gains to date. Even heavyweight dossiers, including clinical study reports, clinical summaries and clinical overviews and the like, can see time cuts “in the 50% to 70% range” once AI handles first-draft generation and humans stay on the hook for accuracy. More typically, the clinical study report (CSR)-drafting gains are in the 30% range.
Timothy Martin
Beyond accelerating the drafting of comprehensive reports, the efficiency gains can appear even more dramatic for highly structured, repetitive tasks. For instance, a scientific writing team member at Lilly, a Yseop client, reported that traditionally, it took a minimum of four hours to write a single patient narrative. “In a large trial you might face tens of thousands of patient narratives,” Martin said. “Using a structured data pipeline, we can now generate those narratives in seconds instead of the four or five hours writers used to spend on each one.”
Accuracy first
Of course, the raw speed gains don’t matter if they come at the cost of accuracy. “One of the most important things that we can do for our customers, and what they expect is truth and accuracy,” Martin said. When delivering the results sections from clinical data, the goal is 100% accuracy. Achieving precision, Martin explained, involves a mix of humans and machines. In other words, a robust technological foundation combined with human oversight. A series of elements that work in tandem to enhance accuracy as well, including retrieval augmented generation (RAG), knowledge graphs, and metadata, but humans need to be in the loop. “At the end of the day, the medical writers and scientists are still responsible for the output of that study,” Martin said.
While creating standardized workflows that keep humans firmly in the loop is important, best practices need to extend across the board. If you manage data processes well, writing processes, too, AI tools are much more effective, Martin said. Dossiers must be built from machine‑processable data. It is also helpful to standardize the input going into the AI: “We manage the prompts so there’s a level of consistency and accuracy,” Martin said.
It takes a village
Martin noted the gains can plateau quickly unless teams plug AI straight into the systems they already trust. “One of the things we’re doing is working with Veeva,” he said, referring to the cloud-based content‑management platform that many large pharmas use. “Source documents live there, we author new ones, and the audit trail stays intact: GxP compliance, version control, everything.” By piping structured data out of Veeva and back in again, Yseop can apply the same governed prompts across studies. As a result, writers get a consistent template while still flagging any anomalies for human review.
Now that the FDA is also chasing genAI-driven efficiency gains, perhaps the next phase of work for the wider industry will shift from head-spinning productivity gains to a deeper focus on disciplined data hygiene, prompt governance and accountability. Those maturity gains could lead to continued progress over time. “If you manage data processes, well, if you manage writing processes, well, AI tools are much, much more effective,” Martin stated. There is a need for “pushing really hard to have very solid processes — data consistency — [and] machine processable data for all studies, moving forward, consistently across all studies.”
In other words, standardizing the human element and the raw data is as vital as refining the machine learning models themselves. If those pieces fall into place, the industry’s busiest writers could soon spend more of time interpreting data rather than just formatting it.
Filed Under: Data science
