his past August, Tufts CSDD (Center for the Study of Drug Development) and Medable jointly announced the results of a study that asked questions about the economic value of implementing an AI monitoring agent in a clinical trial and answered these questions with empirical evidence. Based on a benchmarked oncology program, Medable contract and experience data, and Tufts CSDD analysis of key activities in oncology trials that directly impact study timelines, resource utilization, and efficiency, this study demonstrates that AI agents genuinely accelerate study timelines and improve productivity, which will ultimately deliver more cures to more patients.
Q: How did this study originate? Who first proposed the research question, and how did Medable and Tufts CSDD divide their respective roles in designing the analysis, supplying the data, choosing the assumptions, performing the calculations, and interpreting the results?
A: Tufts CSDD has a long history of conducting independent empirical studies quantifying the net financial value of various innovative solutions (e.g., patient engagement, Decentralized Clinical Trials [DCTs], digital endpoints) to inform industry-wide adoption. Medable approached Tufts CSDD to conduct a similar, independent study. The Tufts CSDD research team, led by Dr. Joseph DiMasi, directed and managed all aspects of the research effort. Medable provided data on actual use cases associated with its agentic AI study monitoring solution including clinical trial outcomes data from actual use experience and the investment required to deploy the solution.
Q: The announcement describes the analysis as using both “actual use” data and benchmark data, while the methodology says Medable supplied operational data from its Clinical Monitoring Agent “modeled in an oncology development program.” Can you walk through the underlying data set? Which improvements were actually observed in trials using the agent, which came from Tufts benchmarks, and which were assumptions or modeled extrapolations?
A: The underlying data populating the model are oncology clinical trial- and program-level benchmarks compiled and maintained by Tufts CSDD. We also gathered data on global travel costs and global salary levels of clinical research associates (CRAs). As mentioned earlier, Medable provided clinical trial outcome data (study timelines) from actual agentic AI use cases as well as the financial investment required to deploy the solution per individual CRA.
In modeling impact, Tufts CSDD used data from recent sponsor-reported operational changes to on-site and off-site monitoring in transitioning from traditional to risk-based/centralized monitoring. This data informed the change in CRA workload and number of visits under an agentic AI-supported approach. The primary assumption modeled (a 20% reduction in study enrollment) came from published reports (e.g., McKinsey).
Q: What clinical-monitoring activities did you actually evaluate? What tasks were performed or augmented by the AI agent, what would a CRA or other study personnel ordinarily have done instead, and what parts of clinical monitoring were outside the scope of this analysis?
A: The agentic AI solution largely addresses the very high time- and labor-intensive tasks handled by a study monitor, including 100% source data review and verification, query management, eTMF (electronic Trial Master File) and ISF (Investigator Site File) document review and filing, site communications and notifications, and action item and follow-up tracking.
Q: One of the most consequential findings is roughly 18 weeks of acceleration, including a 109–119-day reduction in enrollment time and about two weeks to database lock. How did you establish that the monitoring agent was responsible for those improvements? Were those differences directly observed against comparable trials or sites, or are they estimates based on changes in workflow? What other factors could account for some of that difference?
A: Enrollment time improvements were informed by published reports (e.g., McKinsey, MIT) and corroborated elsewhere. It largely reflects an agentic AI solution accelerating signal detection and alerting the site sooner so that it can implement remedial steps. Although we modeled time savings, we did not include cost reductions due to lower drop-out rates. This is an area that we will be pointing out in the forthcoming peer-reviewed manuscript.
Q: You estimate reductions in on-site monitoring costs of about $4.4 million for phase 2 and $5.6 million for phase 3, plus another $600,000 and $1.7 million respectively in off-site administrative efficiencies. What portion represents cash that a sponsor would actually stop spending, versus CRA time or organizational capacity that is freed up and could potentially be redeployed elsewhere?
A: The $4.4 million and $5.6 million ROI figures represent both cost savings (e.g., reduced CRA travel costs associated with on-site visits) plus time savings that can be reallocated. On average, CRAs typically oversee sites participating in three studies. For the CRA, the reduced work effort on the agentic AI-enabled study frees up their time to focus on stronger relationships with site personnel and to work on other studies.
Q: The numbers likely to attract the most attention are the estimated 64x ROI in phase 2 and 82x in phase 3. What exactly is the numerator and denominator in that calculation? Which costs of implementing and operating the agent are included—software, integration, validation, training, change management, and ongoing human oversight—and which costs are excluded? How sensitive are those ROI figures to different assumptions?
A: The eNPV values and multiples represent the difference between the eNPV of a traditional oncology development program compared to the eNPV of the agentic AI-enabled oncology program. The eNPV model is risk-adjusted and includes all cash-in and outflows associated with program development expenses, the full program investment required to deploy the innovation for the agentic AI-enabled program, development durations, and the commercial performance of approved oncology therapies.
The program-level eNPV value and ROI multiples represent expected outcome if an agentic AI solution were implemented effectively. We describe this as average expected net financial impact because we have not taken into account order-of-entry effects and individual program commercial potential. We have averaged the commercial performance of a large number of marketed cancer therapies.
Q: This analysis is based on an oncology development model. What characteristics of the modeled phase 2 and 3 trials—number of sites, patients, countries, monitoring intensity, enrollment duration, and so forth—drive these results? How confident are you that the findings would translate to a smaller oncology study or to therapeutic areas with different enrollment and monitoring requirements?
A: All the oncology development data is drawn from Tufts CSDD benchmarks published in the literature (most often in Therapeutic Innovation & Regulatory Science [TIRS]). The oncology area was deliberately selected for several reasons: Not only is it the most active area in drug development, but also because investigational oncology therapies are sponsored by a highly diverse community of emerging, small, and large sponsor companies. Also, these investigational therapies are highly personalized, targeting very narrowly defined patient communities.
We’re confident that the results translate well across oncology programs. We have conducted sensitivity analysis to directionally show how the eNPV delta changes with a much larger targeted patient community (e.g., chronic diseases). However, we encourage sponsor companies to adapt and customize the model to best reflect their unique portfolio experience.
Q: Is the $21 million/82x result best understood as a representative expectation, a plausible upper-end scenario, or something closer to a best-case situation? That last distinction matters because the press release says that AI agents “can deliver up to” $21 million and 82x ROI rather than presenting those figures as average expected performance.
A: The $21 million eNPV is for phase 3 trials, so when you look at phase 2 trials, the eNPV is $7.5M for the program, as the risk to the development success is greater.
Q: The detailed paper has not yet gone through peer review. If you were reviewing this work yourselves, what would you identify as its two or three most important limitations? And what additional evidence—prospective studies, controlled comparisons, independent data sets, or replication with other technologies—would you want to see before concluding that these economic benefits are broadly achievable in clinical development?
A: The detailed paper has been submitted for peer review, and it touches on two really important limitations. First is that our financial modeling is based on averages and benchmarks. Companies need to customize the model with their own data to get a more representative measure of the value proposition. The second limitation is that our model assumes that an organization will effectively deploy the innovation. But for many reasons, organizations often struggle to fully leverage an innovation’s impact.
Global Forum thanks Data/Digital Editor Lisa Barbadora for coordinating this Q&A.