AI Drug Candidates Clear Phase I at 80%, Then Stall in Phase II

The Phase I Illusion

AI-designed drug candidates are clearing Phase I trials at rates between 80 and 90 percent, nearly double the historical industry benchmark of 40 to 65 percent. Yet of 117 AI-enabled assets currently tracked in clinical trials, only eight have successfully completed Phase II. The disconnect between early promise and clinical validation has prompted senior pharma executives to question whether the industry has been measuring the wrong success metric.

Vikram Singh, Head of Enterprise AI at Gilead Sciences, captured the tension in a recent post: "Speed is bought. Trust is earned in Phase III." His observation follows years of celebration around AI's ability to accelerate molecule generation and reduce early-stage costs. Insilico Medicine, for example, advanced a novel target to clinic in under 18 months, a process that typically costs more than $430 million. But Phase I only answers whether humans can tolerate a molecule. Phase II determines whether it actually changes disease progression, and on that measure, AI has produced no improvement over the historical 40 percent success rate.

Why it matters

The pharmaceutical industry has invested heavily in AI tools that optimize the front end of drug discovery while the fundamental constraint remains unchanged: validating biological mechanisms in human disease. If AI-enabled assets continue to fail at Phase II at historical rates, the technology may be accelerating the production of molecules against already-understood targets rather than identifying genuinely novel therapeutic mechanisms. That distinction will determine whether AI represents a paradigm shift or an efficiency gain.

The Bottleneck Nobody Moved

Greg Meyers, Chief Digital and Technology Officer at Bristol Myers Squibb, argues that AI is "moving rather than removing bottlenecks in drug discovery." The technology excels at generating plausible candidates, but plausibility was never the constraint. The constraint is understanding which biological mechanisms actually alter disease course in living humans, knowledge that remains gated by expensive, multi-year clinical trials.

Meyers points to two technologies the industry has over-valued: generative molecular design and self-driving labs. Both perform well when there is a fast, cheap scorecard. Drug efficacy in humans provides neither. The signal arrives years later, in Phase II, at costs that dwarf upstream savings. What the industry under-invests in, according to Meyers, is using compute to choose the next physical experiment rather than generating more screening candidates, and systematically mining failure data. Information about which targets failed, in which patient subgroups, and why, typically gets discarded after unsuccessful trials instead of being fed back into models.

Data Quality, Not Model Capability

A 2026 survey of 115 pharma and biotech digital leaders found that 68 percent cite poor data quality and governance, not model capability, as the primary reason AI initiatives stall. The training data for drug efficacy prediction remains compromised by publication bias, inconsistent patient stratification, incomplete biomarker annotation, and the discarded failure data Meyers identified as the field's most underutilized asset.

Simon Istolainen, founder of CURE51, argues the venture capital herd has funded identical platforms pursuing identical strategies, producing "the same predictable failures." If most AI-enabled assets in clinic pursue mechanisms that were well-characterized before AI arrived, the high Phase I clearance rate may simply measure tolerability for molecules generated more efficiently from already-understood biology.

The first wave of Phase III readouts will arrive in 2026, but even a wave of approvals would not prove AI identified the right targets. It would prove AI-generated molecules against validated targets perform as well as conventional molecules against those same targets. Real proof would require an AI-identified, previously unknown mechanism producing Phase III success in a disease where prior attempts failed. That trial does not yet exist in the public literature.

These details were first reported by Clinical Trial Vanguard.