Tufts model estimates 82x ROI and up to $21 million in value from Medable AI agent. Tufts Center for the Study of Drug Development (CSDD) and clinical trial platform company Medable recently announced an analysis estimating that Medable's Clinical Monitoring Agent could deliver an 82x return on investment for phase 3 trials and 64x for phase 2. Based on expected net present value (eNPV), that equated to roughly $21 million in added financial value for phase 3 and $7.5 million for phase 2. The analysis focused on oncology. The analysis comes as enterprise use of agents is surging, and many organizations are using agents for workflows ranging from coding to tackling routine business workflows. McKinsey reported in July that 93% of its survey respondents reported exceeding AI budgets, while, 41% of software developers reported agents saved them one to two hours per day, on average, according to IBM in late 2025. Meanwhile, in the clinical trial domain, a 2025 Tufts CSDD analysis of 36 company-reported AI/ML use cases found an average 18% reduction in cycle time, reporting the largest trial-execution gains in patient monitoring and enrollment assessments. Putting numbers on the pilots. The latest Tufts-Medable collaboration sought to translate those reported operational gains into the language of investment decisions. "We've known for some time that companies have been piloting AI enablement tools, but what they've often been lacking is any kind of quantification of the actual ROI, the value proposition," said Ken Getz, executive director of Tufts CSDD. "We were really interested in doing a study where we could put some hard numbers around a specific AI enablement solution." Dr. Pamela Tenaerts, Chief Medical Officer at Medable, sees the potential savings expanding for sponsors with sizable oncology portfolios. She projected an incremental portfolio eNPV of up to $226 million across phase 2 and 3 studies for sponsors with 20 active indications and up to $565 million for those with 50 indications based on modeled portfolio scenarios that apply the same estimated value across each indication. Prebuilt agents, then about 20% tweaking. While agents can tackle a range of workflows, their clinical-trial configuration varies from sponsor to sponsor. Medable starts with prebuilt monitoring and trial master file (TMF) agents with connectors to clinical, data and enterprise systems. Medable's connector page lists roughly 40 integrations, spanning EDC (Medidata Rave, Oracle InForm, Veeva EDC), eTMF and CTMS (Veeva), safety (Oracle Argus), randomization (Oracle IRT), warehouses (Snowflake, Databricks, BigQuery) and general enterprise tools (Slack, Teams, SharePoint-adjacent storage). Sponsors configure the agents depending on their needs. "A lot of those agents still need about 20% tweaking, because systems are different and it has to be configured," Tenaerts said. In a Medable demonstration reviewed by the author, the agent pulled together information from multiple clinical systems and flagged a site with unusually low adverse-event reporting alongside concomitant medications recorded without corresponding adverse events. It then recommended opening EDC queries and sending correspondence to the site. "It can also see across the sites that person has, whether that's a recurring issue or not," Tenaerts said. Wrangling millions of data points per protocol. The ability to find patterns in mountains of data efficiently remains a priority as phase 3 protocols now involve an average of 5.9 million data points, up from 929,000 in 2012, per the Tufts CSDD/TransCelerate analysis of 105 protocols across 15 sponsor companies. "As the number of items collected multiplies, imagine you're a CRA [clinical research associate] and you have a study," Tenaerts said. "Your job is to wrangle that data together and make sure it all makes sense... You have an EDC system that collects data around the actual protocol data. You have CTMS [Clinical Trial Management System], you have TMF, and the CRA has to put all that data together in context and decide what to focus on." Clinical trials can be difficult to optimize through iteration. Randomization compares groups under prespecified conditions, leaving unobserved how the same trial would have performed with a different endpoint, eligibility rule, dose or monitoring plan. Testing those alternatives generally requires another trial with new participants, more time and substantial capital. Meanwhile, in clinical trials, avoiding outright failures is, of course, a priority. According to a study published in Journal of Medicinal Chemistry, the cumulative failure rate of small-molecule cancer drug candidates has hovered around 95% during phase 1 through phase 3 for roughly three decades. The odds improve in phase 3, but, still, failures outnumber successes. For instance, a 2026 analysis of 824 phase 3 oncology drug trials registered from 2007 through 2023 found a success rate of 43.4% (358 out of 824 trials). Can agents help with clinical trial failures? A host of problems are behind those long odds. Asked how an agent might catch a problem a human would miss, such as a fuzzy endpoint, a screening error, or a design flaw surfacing mid-study, Tenaerts described the agent's advantage in spotting errors a human might initially miss. "The agent has the ability to see things slightly earlier than a human would," she said. "That likely happens faster than a human can make that connection, so the person can intervene on that issue." She pointed to examples where an agent could point out sites that are falling behind on enrollment, protocol deviations traced to a lab value that repeatedly runs above the range specified in the protocol, and an adverse event emerging sooner on a trend basis. Getting a handle on the complexity can free up humans to focus less on data wrangling and more on interpreting the nuances of a given trial, Getz said. "There are very tactical decisions, like at what point do we need to alert the sites if we've detected a pattern in the deviations," he said. "And then there are major decisions of importance. Should we change the strategy of the study? Do we modify the design in some way? Do we need to address an amendment? What an agentic tool can help do is give the human the opportunity to focus on those higher priority decisions, and a lot of the more mundane ones can be streamlined or managed a lot more effectively." To some extent, the adoption of agents in clinical trials is at an early phase. In an interview with Pharmavoice, Tenaerts cited the complexity of oncology and its high unmet need, and touched on the potential for the model to extend to other trial types. "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," she said. Where automation is landing in clinical trials today. Over the last couple of years, frontier AI labs have boasted about the ability of their models to do complex tasks, largely unattended. For instance, Anthropic said Mythos "conducted novel genomics research in over a week of largely autonomous work." Google DeepMind similarly said in July that AlphaEvolve, a Gemini-powered agent for designing and optimizing algorithms, "runs unattended, generating and improving candidates at a scale no human team could match." Medable said its agents tend to have a shorter leash. "Most of what the agent does right now is not autonomous," Tenaerts said. She described query resolution as close to the only action it takes on its own: when a field is missing, the agent writes the query, sends it to the site and files it in the TMF, and the CRA sees the record rather than approving it in advance. Everything else follows a recommend-and-review pattern in which the agent surfaces an issue, proposes next best actions and waits. "The agent doesn't suspend the CRA from thinking," she said. "The CRA still needs to be there." She also flagged where the work lands once monitoring speeds up. "If we are getting really efficient at sending out email to sites, we sure hope that the sites will get some help, because if you dump all that stuff on them, that's still a bottleneck, right?" Tenaerts said. "You push the balloon and it goes somewhere else. You need to figure out the whole system." Venu Mallarapu, chief transformation and AI officer at competing clinical data platform eClinical Solutions also put a ceiling on how far deployments have gone. Most eClinical customers are running assistive AI that drafts while a person handles the acting, he said, while the most advanced run scoped agents inside workflows that have been reengineered around them. "There's nobody that is running fully autonomous agents that not only process the data but also make decisions without human involvement," Mallarapu said. "At least I haven't seen that within our customer base."

Medable, Tufts CSDD value AI gain at $21m net per programme. New analysis from the Tufts Center for the Study of Drug Development (CSDD) and Palo Alto, CA-based Medable shows that AI agents accelerate clinical trials and improve staff productivity, delivering net financial gains to a sum as high as $21 million per drug development programme, as well as 82 times the return on investment (ROI). Medable partnered with Tufts CSDD to empirically answer the question: To what extent can clinical monitoring agents improve trial performance, and do those translate into measurable economic value for drug developers? Medable provided operational data - including operational time savings and agent cost - from its Clinical Monitoring Agent modelled in an oncology development programme. Tufts CSDD applied the expected Net Present Value (eNPV) framework, a widely accepted method for quantifying the financial impact of innovation investments, to assess the value created by the AI agent. On average, it takes 10-12 years to bring drugs to patients. That is expected to significantly shrink for the first time in the digital age. Indeed, for the first time since 1962, the average number of drugs approved each year in the US may finally jump past 50, breaking a long-held drug development glass ceiling. The Tufts CSDD analysis specifically assessed the impact of Medable's Clinical Monitoring Agent across three top-line metrics: expected net present value, overall return on investment, and direct operating cost savings. The agent showed eNPV gains of approximately $7.5 million (phase 2 trial), $11.3 million (combined phase 2 and phase 3) development, and $21 million (phase 3 trial). Meanwhile, the study found an estimated ROI of 64x for phase 2 and 82x for phase 3 clinical trials. From the data, Tufts CSDD estimated direct operating cost reductions in on-site monitoring per clinical trial of approximately $4.4 million per phase 2 and $5.6 million per phase 3 study. Additional analysis identified and valued administrative off-site monitoring task efficiencies of approximately $600,000 (phase 2) and $1.7 million (phase 3). These savings reflect clinical research associate time that could be reallocated to other studies and were not included in the eNPV calculations. Dr Pamela Tenaerts, chief medical officer at Medable, commented: "For a sponsor with 20 active indications, deploying a clinical monitoring agent across phase 2 and 3 studies could generate as much as $226 million in incremental portfolio eNPV. For a sponsor with 50 active indications, that figure could jump to as much as $565 million. Bottom line? We now have evidence demonstrating sizable value creation of agents in clinical research, helping break longstanding barriers." In the patent cliff race to 2030, where blockbuster drugs representing $200+ billion in annual revenue are expected to lose exclusivity, the benefits are clear. In addition to these high-level findings, Tufts and Medable will publish a more detailed peer-reviewed paper later this year. Dr Pamela Tenaerts previously spoke with pharmaphorum following ASCO 2025, for a conversation on the company's digital-first model to handle long-term follow-up care for cell and gene therapy cancer trials. 19 August, 2026

Medable has launched its Digital Data Flow Agent, which automatically converts static clinical trial protocols into machine-readable, structured data in CDISC USDM 4.0 format. The technology enables continuous alignment across the clinical development lifecycle, allowing trial sponsors to generate electronic clinical outcomes assessments in minutes whilst maintaining protocol consistency. The DDF Agent addresses the traditional challenge of manual protocol interpretation across multiple systems. When protocols are amended, changes automatically propagate across connected documents and workflows, reducing operational burden and enabling faster response to study insights. The technology supports the FDA's vision for real-time clinical trials, where ongoing data review becomes standard practice. Medable's platform has been deployed in nearly 400 trials across 70 countries, serving over one million patients globally.