Vikram Singh, Head of Enterprise AI at Gilead Sciences, put the industry’s central anxiety into a single sentence this week: “Speed is bought. Trust is earned in Phase III.” He posted it after laying out numbers that every pharma board already knows but hasn’t fully processed: AI-designed candidates are clearing Phase I at 80 to 90 percent, compared to the historical industry rate of 40 to 65 percent. And yet, of 117 tracked AI-enabled assets in clinical trials, only 8 have cleared Phase II. The launchpad is crowded. The orbit remains empty.
That gap deserves more attention than it is getting. The industry has spent three years celebrating Phase I clearance rates as proof of concept for AI-driven drug discovery, and the celebration is not entirely wrong. Insilico Medicine took a novel target to clinic in under 18 months, a process Singh notes typically costs more than $430 million. That is genuinely remarkable. But Phase I answers a narrow question: can humans tolerate this molecule? Phase II answers the question that actually matters: does this molecule change a disease’s course? On that question, the AI era has produced no new answer. The success rate sits at roughly 40 percent, the same figure the industry has lived with for decades.
The problem is not that AI failed. The problem is that the industry celebrated a victory in the wrong race.
The Bottleneck Nobody Moved
Greg Meyers, Chief Digital and Technology Officer at Bristol Myers Squibb, framed the structural issue with unusual precision: “AI is moving rather than removing bottlenecks in drug discovery.” He was reacting to a blog post from Daphne Koller at Insitro, and his argument cuts directly against the dominant narrative. AI excels at generating plausible candidates, he wrote, but “plausibility was never the hard part.” The hard part, the part that has not changed, is knowing which biological mechanism actually alters a disease’s course in a living human. That knowledge is still gated by the slowest and most expensive experiment in science: a randomized clinical trial in human subjects.
This is the counterintuitive truth the AI-in-drug-discovery complex has been reluctant to state plainly. Conventional wisdom holds that better molecule generation should improve downstream outcomes because better inputs produce better outputs. But that logic assumes the constraint is candidate quality. Meyers and Singh are both pointing at a different constraint: biological mechanism validation. You can generate ten thousand structurally elegant molecules, but if you do not understand why a target behaves differently in a diseased patient population than in a cell line, you are optimizing the front end of a pipeline whose back end remains ungoverned by the same tools.
Meyers named two specific technologies the industry has been over-bullish on: generative molecular design and the self-driving lab. Both perform well, he noted, when there is a fast and cheap scorecard. Drug efficacy in humans does not provide a fast or cheap scorecard. The signal arrives years later, in Phase II, at a cost that dwarfs the savings captured upstream.
What he argued the industry is under-investing in is more instructive: compute pointed at choosing the next physical experiment to run rather than generating more candidates to screen, and systematic mining of failure data. Which target, in which patient subgroup, with which biomarker profile, failed, and why. That information currently gets discarded after every unsuccessful trial instead of being fed back into the model. The industry is running an algorithm on incomplete training data, and the missing data is the most valuable data it has ever produced.
The Herd Is Chasing the Same Cliff
Simon Istolainen, founder of CURE51, offered the sharpest structural critique of the current AI investment thesis. The Insitro and a16z conversation about “uncovering new biology,” he wrote, has prompted a belated industry recognition that drug design and clinical trial segmentation may not be the optimal use of AI. His argument: “We don’t know enough about diseases and cancer specifically. There is a whole new world to explore.” The herd of VCs, he noted, has funded the same platforms pursuing the same strategies, producing “the same predictable failures.”
That framing matters beyond venture capital dynamics. If the majority of AI-enabled assets in the clinic are pursuing mechanisms that were already well-characterized before AI arrived, then the 80 to 90 percent Phase I clearance rate is measuring something misleadingly narrow: tolerability in humans for molecules generated more efficiently from already-understood biology. The question Singh’s data raises, but cannot answer, is whether any of those 117 tracked assets represent genuinely novel mechanism-of-action hypotheses that AI identified and that human researchers would not have reached through conventional methods. Per Singh’s sources, only 8 have cleared Phase II. That number will grow in 2026, as the first wave of Phase III readouts arrives, but the baseline is not encouraging.
Terrel Marks of IBM approached the same problem from the commercial side, arguing that most life sciences companies have their AI investment priorities inverted, pouring resources into back-office efficiency while underweighting the functions that directly shape patient outcomes: access, trial recruitment, and supply reliability. His critique maps onto the discovery debate in a precise way. The industry optimized for the part of drug development where AI produces fast, measurable, reportable wins. Phase I clearance rates are fast and measurable. Phase II biological validation is slow, expensive, and frequently humbling. Investment, narrative, and media attention have all followed the faster signal.
Per Singh’s ZS 2026 CDIO Research surveying 115 pharma and biotech digital leaders, 68 percent of technology leaders cite poor data quality and governance, not model capability, as the reason AI initiatives stall. That figure is doing a lot of work. It tells you that the industry’s core AI constraint is not algorithmic: it is the trustworthiness of the biological and clinical data the algorithms are trained on. A model is only as good as its training set, and the training set for drug efficacy prediction is riddled with publication bias, inconsistent patient stratification, incomplete biomarker annotation, and exactly the kind of discarded failure data Meyers identified as the field’s most underutilized asset.
The 2026 Phase III readouts will not resolve this debate cleanly. Even a wave of approvals would not prove that AI identified the right target; it would prove that AI-generated molecules against already-validated targets perform as well as conventionally generated molecules against those same targets. What would constitute real proof is an AI-identified, previously unknown mechanism of action producing a Phase III success in a disease where prior attempts failed. That trial does not yet exist in the public literature. Until it does, the industry is measuring the speed of a tool against a finish line the tool has not yet crossed.
References
- Vikram Singh, LinkedIn post on AI Phase I clearance rates and Phase II stagnation, 2026
- Greg Meyers, LinkedIn post on AI moving rather than removing drug discovery bottlenecks, 2026
- Simon Istolainen, LinkedIn post on AI, new biology, and VC herd behavior in drug discovery, 2026
- Terrel Marks, LinkedIn post on AI investment priorities in life sciences, 2026
Moe Alsumidaie, MBA, MSF, is founder and Chief Editor of Vanguard Publications, which publishes Clinical Trial Vanguard, Pharma Vanguard and BullScope, and Head of Research at CliniBiz. He has two decades in clinical trial operations and data science, with earlier roles at Genentech, Abbott Vascular and Stanford University Medical Center, and is a guest lecturer in clinical trial sciences at Rutgers University.
