The pharma industry is investing aggressively in AI across drug R&D, but there’s still limited data on whether it’ll translate into treatments that work in the clinic.
But a recent analysis by the Tufts Center for the Study of Drug Development suggested the technology could pay off in a number of other ways during the R&D process, delivering expected net financial value of up to $21 million for a single phase 3 drug development program, and up to 82 times the return on investment.
While the study offered new insights about AI’s potential, it was also limited to oncology trials, leaving questions about what the findings may mean for broader AI adoption.
Still, interest in the findings has been high as industry leaders look for answers about whether AI is worth the investment, said Ken Getz, executive director of Tufts CSDD. This is especially true as federal regulators develop frameworks for broader AI use in drug development.
“There are … regulatory tailwinds,” he said. “But what we were missing was quantification of the value proposition, and that's really where we are now.”
To gauge AI’s impact, Tufts zeroed in on net financial value to show the full implications of speed and cost improvements over the lifecycle of a product, which is a “really meaningful measure for senior management,” Getz said.
The analysis combined Tufts’ benchmarked oncology data with operational and cost data from Medable’s AI clinical monitoring agent. In the model, efficiencies including faster patient enrollment and earlier database lock shaved development timelines by approximately 18 weeks, paving a path toward faster revenue generation while reducing development costs.
Here, Getz and Dr. Pamela Tenaerts, chief medical officer at Medable, offer additional details about the analysis and what it might mean for an industry that is still sometimes embracing AI tools with caution.
This interview has been edited for brevity and style.
PHARMAVOICE: Why did the analysis focus on oncology, and are results generalizable to trials in other therapeutic areas?
PAMELA TENAERTS: We chose oncology because there was good data that we could benchmark off. Oncology is also a large indication. It’s a complex therapeutic area with high unmet need, so we thought modeling it there made sense. Can you extrapolate from that? That’s the question. To a degree, I feel like we can, but we don’t have data. It makes logical sense that something you do in an oncology program we could probably replicate in other programs. I feel comfortable saying that we can, but we haven’t done the model.
KEN GETZ: One of the most powerful parts of the study is that all of our financial data has been adjusted to reflect the current market environment. Not only are we looking at valuing time using the most prevalent rates, but even the market potential of drugs in recent years that have targeted oncology-related illnesses. It can be generalized. We're showing a way companies should do their own custom modeling using their own target market and therapeutic areas. They can use the same basic approach we've used to derive the net financial valuation.
Did your modeling factor in the need for human oversight?
TENAERTS: We modeled it after what the Medable CRA Agent does. So, if another agent does something different, the model might come out slightly differently. In the Medable scenario, we are reducing the time it takes to review data, because the agent is helping with that. But the CRA is still involved in checking to ensure the data makes sense. The agent is involved with drafting emails, but the CRA still makes sure that the email sounds like something she or he would have written. The CRA agent is linked to the CRA, and it logs in as the CRA. We have audit trails showing what the agent does and what the CRA does. But it is modeled with a lot of human-in-the-loop oversight. There are things that don’t use human-in-the-loop oversight such as a missing value. The agent knows that we need to have that value, so it might go ahead and send that query automatically. All of those settings are set up with the sponsors and what they're comfortable with.
GETZ: In the press release, we tried to convey that there are efficiencies gained in so many different domains. There's the whole on-site monitoring area, which has a level of human oversight with people actually going out to the site and reviewing source material, meeting with the investigative site personnel. There's a place there where there's some efficiency. But there's also all of the offsite monitoring, which are more administrative or task-based efficiencies. Each has different places where there's human review. We thought comprehensively about areas where impact could be measured. It wasn't just a simple, straightforward area. It's really going to touch on so many different aspects of the monitoring function.
What does this analysis say about the use of AI in trials overall?
GETZ: In an academic environment, doing these big macro-level studies, we generally see remarkable amounts of inefficiency. Everything new that's incorporated only adds more time, and can in many cases create even more inefficiencies because we're building on a foundation of inefficiency. A lot of these new AI tools, agentic AI in particular, seem to be fundamentally addressing some of those inefficiencies and freeing up staff and teams to focus on the higher value areas. That's particularly exciting, and I think you're seeing more organizations target areas that were historically labor intensive and known to be inefficient. You're going to see similar value creation from a lot of these new AI-enabled tools.
