It costs $2.67 billion, on average, to shepherd a new therapy from laboratory concept to pharmacy shelf. That figure, pulled from Deloitte's 2026 analysis of pharmaceutical R&D spending, has climbed steadily for years despite industry efforts to rein it in. Returns improved slightly—from 5.9% to 7.0%—thanks mostly to the windfall from GLP-1 weight-loss drugs. But the underlying math remains punishing. Clinical trials still fail more often than they succeed, and the successful ones frequently demand enrolling hundreds or thousands of patients just to prove a drug works.
Enter a new cohort of startups with a pitch that sounds almost too good: What if artificial intelligence could identify, before a trial even begins, which patients are most likely to respond? Train a model on millions of biological samples, they argue, and you can predict individual drug response with enough accuracy to slash trial sizes, boost success rates, and save billions in development costs.
The promise is seductive. The track record is still being written.
A Very Expensive Gauntlet
That $2.67 billion encompasses the full journey—discovery, preclinical studies, three phases of human trials, regulatory filings. Clinical development claims the lion's share, especially Phase 3 trials that can sprawl across dozens of sites and enroll thousands of participants. When those trials stumble—and roughly half of Phase 3 programs do—the money evaporates.
Development timelines, counterintuitively, have actually lengthened in 2025, with overall development durations slowing due to longer enrollment and inter-trial intervals. According to IQVIA's Global R&D Trends report issued last May, enrollment periods are stretching out, and the gaps between trial phases have widened. But buried in the same report is a more tantalizing observation: AI-enabled programs at emerging biotech firms appear to be posting stronger success rates than non-AI counterparts. The dataset is small, the signal nascent, but it marks the first industry-wide hint that these tools might deliver more than hype.
Investors seem convinced. Market research from Fortune Business Insights pegs the AI-in-clinical-trials sector at $3.8 billion in 2025, with projections showing a compound annual growth rate of approximately 39%, reaching $77.3 billion by 2034. Whether that trajectory holds depends on whether the technology actually works at scale.
Fishing for Responders
The core idea is disarmingly simple: use AI to identify patients most likely to benefit from a given drug, then enroll those responders in trials. Even modest predictive accuracy can shrink the number of participants needed to hit statistical significance—and smaller trials mean lower costs, faster timelines, and less exposure to potential side effects for patients who won't benefit anyway.
Atlas Discovery, a Y Combinator-backed startup operating out of San Francisco, offers one version of this approach. The company built what it calls a foundation model trained on millions of transcriptomic and biological samples. According to the company, for each drug and disease pairing, they attach a small supervised learning module designed to classify likely responders.
The team recently circulated a retrospective analysis of UNIFI, a 2019 Phase 3 trial of ustekinumab for ulcerative colitis that was published in the New England Journal of Medicine. Using baseline colon biopsy transcriptomes from 358 treated patients, Atlas reported an AUROC—a measure of predictive accuracy—of 0.76. According to the company's modeling, that level of precision could theoretically reduce enrollment requirements by roughly 3.5 times while maintaining statistical power.
Atlas is careful to frame this as retrospective number-crunching, not a claim about how the original trial should have been designed. Prospective deployment, they acknowledge, introduces a thicket of complications: regulatory guardrails around safety exposure, shifts in placebo response rates, the messy reality of coordinating biopsy collection across multiple trial sites.
Still, the concept has traction elsewhere. In January 2025, Nature Medicine published research on SCORPIO, a National Cancer Institute-backed tool that predicts immunotherapy response in cancer patients using routine lab values and clinical data. Validated across real-world datasets and 10 trial cohorts, it hit similar predictive benchmarks. The NCI's explainer emphasized potential uses in treatment decisions and trial enrollment strategies.
Worth noting: this is different from the digital twin approach championed by companies like Unlearn.AI or Medidata, which construct virtual control groups from historical data to shrink placebo arm sizes. Atlas and similar firms are aiming to identify responders upfront using transcriptome-based embeddings—a related but distinct bet.
Regulatory Reckoning
The tools exist. The harder question is whether drug sponsors and regulators will trust them enough to redesign trials around their predictions.
Unlearn.AI has made perhaps the most regulatory progress. The company secured qualification from the European Medicines Agency for its PROCOVA framework, which deploys digital twin generators to enable smaller sample sizes in Phase 2 and 3 trials with continuous endpoints. In February of last year, Unlearn announced a collaboration with VectorY on the PIONEER-ALS trial and launched a design tool to help sponsors model potential sample size reductions. The company had raised a $50 million Series C in early 2024, led by Altimeter Capital.
Medidata, a clinical trial software giant, has leaned into its Synthetic Control Arm offering, built atop patient-level historical data from more than 38,000 trials and 12 million participants. The platform has been deployed in oncology programs; Medidata points to the Celsion GEN-1 ovarian cancer trial as a case study where external controls reduced enrollment needs.
Regulators, meanwhile, are sketching out frameworks. The FDA issued draft guidance in early 2025 outlining a seven-step credibility assessment for AI used in regulatory decision-making. Between 2016 and 2023, the agency received more than 500 drug and biologic submissions containing AI components. And this April, the FDA announced a "Real-Time Clinical Trials" initiative in partnership with Paradigm Health, exploring cloud and AI pipelines that could funnel validated safety and efficacy signals to regulators during ongoing trials rather than after they conclude.
Europe is moving on a parallel track. The EMA has published reflection papers and qualification pathways for novel AI methodologies. The EU AI Act's general provisions took effect in August, establishing transparency and risk management expectations for healthcare AI tools—though how those mandates interact with trial design remains an open question.
The Doubt Brigade
Not everyone is buying in. A 2025 academic preprint flagged error risks when virtual controls are spun up from internal trial data alone, without robust external validation. Another paper cautioned about the perils of non-concurrent controls and shaky causal inference in platform trial designs. The worry is that sponsors, desperate to cut costs, will deploy models that sparkle in retrospective analysis but crack under prospective pressure—victims of data drift, population mismatches, or plain overfitting.
Medidata surveyed trial executives last October and found that 93% are either using or investigating AI, with most reporting the technology met or exceeded expectations. But adoption of fully decentralized trial elements—often bundled with AI-enabled remote monitoring—remains sluggish despite patient acceptance, according to a PLOS Digital Health study published in January. Operational headaches persist, especially around coordinating sites across different regulatory jurisdictions.
And there's a nagging question of marginal impact. ICH E6(R3), the international clinical trial standard adopted globally through 2025 and into this year, codifies risk-based and technology-enabled approaches. But it also doubles down on quality-by-design and data integrity—requirements that apply whether you're deploying cutting-edge AI or pencil-and-paper protocols. The tools may help. Rigorous science and meticulous execution still determine outcomes.
The Calculus Shifts
For founders developing therapeutics, the equation is changing—perhaps. Enriched trial designs using biomarker-based patient selection have been around for years, but AI models expand the playbook considerably. A sponsor working on a therapy with a modest effect size—too small to power a traditional Phase 3 economically—might now consider a responder-prediction model to make the trial financially viable.
The catch is precedent. Atlas Discovery is a three-person team that emerged from Y Combinator not long ago. Unlearn has EMA qualification and live partnerships, but hasn't yet published Phase 3 readouts demonstrating realized sample size reductions in practice. Medidata and other CRO-affiliated platforms have deployment experience, though public outcome data remains limited. IQVIA's report hints at better success rates in AI-enabled programs, but the data is aggregated and early-stage.
Medicare drug price negotiations, which began influencing pricing earlier this year, add another wrinkle. CMS published negotiated prices for the first 10 Part D drugs, effective in January, with additional rounds rolling out through 2027 and 2028. Lower prices squeeze margins, making development cost reduction even more urgent. The Congressional Budget Office expects only modest long-run impacts on innovation volumes, but sponsors are already recalculating return-on-investment assumptions.
The FDA's real-time trials initiative, assuming it scales beyond pilot programs, could compress feedback cycles and enable adaptive designs that tweak enrollment or dosing based on interim signals. ICON's partnership with Microsoft, announced in June, and IQVIA's expanded AI planning suite signal infrastructure investments by major contract research organizations to support these workflows.
Waiting on the Data
The market is moving—that much is clear. Whether the technology justifies its $77 billion valuation projection depends on whether the next generation of trials, designed with AI woven into protocols from the start, actually succeed at lower cost and higher rates than their predecessors.
Atlas Discovery and its peers are placing that bet. The proof should start arriving within the next 18 to 24 months, as prospective trials incorporating these models begin reading out. Until then, it's a numbers game—just one with potentially smarter odds.
