In drug R&D, excitement for generative AI has increased the adoption of technology at an unprecedented rate. In fact, over the past 18 months, budgets for early adopters have expanded, and most life sciences leaders plan additional increases this year — nearly half intend to raise AI spend by more than 10 percent.¹ At the same time, infrastructure players, frontier model companies, and smaller application-focused companies are extending their footprints across the R&D value chain. The focus has also moved from chats to multi-orchestrated agentic solutions to harnesses around agentic frameworks, and the industry has shifted its expectations for AI from efficiency alone to how solutions deliver ROI in complex use cases with humans-in-the-loop.

In a high-velocity environment, accelerating adoption requires three things: proving the technology and delivering value; mobilizing the ecosystem; and getting comfortable with human-in-the-AI-loop in place of AI-in-the-human-loop.

Proof of Technology and Delivering Value: A senior executive from a top 10 pharma put it plainly at a recent QuantHealth industry summit: "It's not worth getting out of bed for us unless we're talking about an impact worth at least $100 million." That level of value lives in the hardest problems, specifically drug development, where investments run high (~ billion dollars), approval timelines are long (~ decade), and the probability of success is low (~10 percent).

The industry has long known that changing the trajectory of drug development — from its high costs, timelines, and patient impact — starts with changing which clinical trials are ultimately pursued. Being able to quantify the potential of a drug for smarter decision-making and confidence, though, was largely a theoretical exercise.

But this change is finally upon us. Deep, model-driven outcome predictions and decision intelligence available months or years before trials begin, along with the ability to pinpoint areas of improvement — such as selecting endpoints and subgroups — while iteratively collaborating with experts from various disciplines, can meaningfully help improve the odds of a drug succeeding and have a real impact for patients invest time and effort into trials.

An important note: the economic value is clear if the drug succeeds, but the "fast-fail" approach reduces the opportunity cost, cutting spend on an asset more likely to fail and making space in the same infrastructure for the ones that might have a better probability to succeed (economic value compounds fast within the lifespan of an asset, and even faster on a portfolio level). This computation, resulting from fit-for-purpose, deep computational models augmented by LLMs, can help guide a decision to retire a drug asset earlier and clear a path to multi-million-dollar savings without strain.

While the examples in drug development of deep frontier technology and models integration are rare, in research recent examples provide guidance: CHAI Discovery’s CHAI-2 genuinely hard problem to designing antibodies for 52 targets with no known binders and confirmed hits in < 2 weeks³ and Ismorphic’s collaboration with Lilly and Novartis⁴, structured as relatively modest upfront milestones (compared to the overall potential payout) that pay only as frontier tech proves itself are good benchmarks: Earn conviction on a hard problem first by proving the value of your technology, then let the commercial terms follow the evidence.

Mobilizing the Ecosystem: No company alone can prove a new way of working; the surrounding system has to move with it. This is exactly what’s happening in drug development now.

• • • Government as the legitimizer: Regulators have moved from tolerating computational evidence to inviting it, and the binding constraint has shifted downstream — from building a model to establishing that its output can be trusted, through the validation, provenance, and context that can turn a prediction into evidence. Several initiatives now formalize that credibility layer: MIDD Paired Meeting Program,⁵ ICH M15 harmonized framework and terminology for how model-derived evidence is submitted and judged, ⁶ AI credibility guidance is the FDA’s risk-based framework for AI used to support decisions on drug safety, efficacy, and quality,⁷ and Operation Trialblazer is an HHS-wide, expedited-IND effort with QSP-based first-in-human dose guidance and updated master-protocol guidance.⁸
    • Infrastructure as the foundation: The compute to train and run models over genomic, clinical, and real-world data sits with a few providers, now partnering downward into a stack of raw compute, foundation models, and applications. NVIDIA’s BioNeMo, running on AWS, puts GPU-scale training within reach of companies that could never build it alone.⁹
    • Intelligence as the distribution: The frontier model companies are moving from horizontal tools towards specialization. Anthropic’s Claude Science, a workbench that connects models to the tools researchers already run — PubMed, Jupyter, and open biomolecular models such as Boltz-2 — launched alongside an in-house program to pursue drugs for neglected diseases.¹⁰ Specialist connectors extend that reach. This intelligence-layer-as-distribution is the ultimate mobilizer of the ecosystem.

Human-in-the-AI Loop vs. AI-in-the-Human-Loop: The third source of trust takes the longest to build because it is about us. Drug development is multidisciplinary, and therefore a useful technology system has to speak medicine, science, regulation, and engineering, and interface with multiple stakeholders who each hold a piece of the decision.

Only a handful of companies working alongside pharma have assembled solutions that reflect the needs of these diverse groups, and even fewer have the same range of talent as their core DNA. That presence is what moves an organization from AI-in-the-human-loop, where the expert does the work and the model assists at the margins, toward human-in-the-AI-loop, where a frontier model carries the first pass and the expert becomes its reviewer and approver. Professionals still prefer a human to AI for complex, high-stakes work by roughly nine to one¹¹ and where a wrong call costs hundreds of millions of dollars, that caution is well placed. Closing the gap is slow work: staying beside the teams as the technology advances, keeping AI in the human's line of sight, and letting comfort accrue one decision at a time.

These necessities — proof, ecosystem, and people — are also changing what is bought and sold. When a system can carry a workflow to a validated result, value moves from accessing a software towards outcome — and commercial terms move toward the milestone-based, risk-sharing structures now forming across AI-driven R&D.¹² This is why QuantHealth has built its product portfolio with a combination of deep foundational models trained by scientific, medical, and biology experts, with recommendations provided with further expertise – or, human-in-the-AI loop.

The companies that earn trust on all three fronts — proving value where the stakes are highest — will be the ones that win and inspire us to get out of bed in service of bringing life-changing medications to patients – ones participating in clinical trials and the ones awaiting a successful outcome.