AI drugs clear Phase I at record rates, but Phase II success stalls at decades-old baseline

The central tension emerged this week when Vikram Singh, Head of Enterprise AI at Gilead Sciences, posted a striking observation: "Speed is bought. Trust is earned in Phase III." He had just laid out data showing that AI-designed drug candidates clear Phase I clinical trials at 80 to 90 percent—a dramatic improvement over the historical industry rate of 40 to 65 percent. Yet of 117 tracked AI-enabled assets currently in clinical trials, only 8 have advanced past Phase II. Phase I success is crowded; Phase II remains sparse.

That gap exposes a celebration the industry got wrong. For three years, pharma has cited Phase I clearance as proof that AI works in drug discovery. The success is real in one narrow sense: Insilico Medicine brought a novel target to human trials in under 18 months, a process Singh notes typically costs more than $430 million. But Phase I answers only whether a molecule is safe to test in humans—not whether it treats disease. Phase II answers that harder question, and the answer has not changed: the success rate sits at roughly 40 percent, the same baseline the industry has lived with for decades.

Greg Meyers, Chief Digital and Technology Officer at Bristol Myers Squibb, framed the structural problem with precision: "AI is moving rather than removing bottlenecks in drug discovery." The distinction is crucial. AI excels at generating plausible drug candidates, but "plausibility was never the hard part," Meyers wrote. The hard part—knowing which biological mechanism actually alters disease in a living human patient—is still gated by the slowest and most expensive experiment in science: a randomized clinical trial. 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 identified two specific technologies the industry has over-bullish on: generative molecular design and the self-driving lab. Both perform well when there is a fast and cheap scorecard, but drug efficacy in humans provides neither. The signal arrives years later, in Phase II, at a cost that dwarfs the savings captured upstream. What the industry is under-investing in, he argued, is compute pointed at choosing the next physical experiment to run rather than generating more candidates to screen, and systematic mining of failure data. The industry is running an algorithm on incomplete training data, and the missing data is the most valuable data it has ever produced.

Simon Istolainen, founder of CURE51, offered a sharper critique: the venture capital herd has funded the same platforms pursuing the same strategies, producing the same predictable failures. If most AI-enabled assets are pursuing mechanisms already well-characterized before AI arrived, then the Phase I clearance rate is measuring something narrower than advertised: tolerability in humans for molecules generated more efficiently from already-understood biology. Whether any of the 117 tracked assets represent genuinely novel mechanism-of-action hypotheses that AI identified—and that human researchers would not have reached conventionally—remains unanswered. Only 8 have cleared Phase II.

A ZS 2026 survey of 115 pharma and biotech digital leaders adds concrete weight to the diagnosis. Sixty-eight percent cite poor data quality and governance, not model capability, as the reason AI initiatives stall. 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 the kind of discarded failure data Meyers identified as the field's most underutilized asset. The core AI constraint is not algorithmic; it is the trustworthiness of the biological and clinical data the algorithms are trained on.