The New R&D Playbook in the AI Era

For decades, pharmaceutical R&D has operated against an uncomfortable equation: enormous scientific progress, rising investment, and stubbornly difficult economics.

The industry even has a name for the paradox: Eroom’s Law — Moore’s Law spelled backwards. Since 1950, the number of new drugs produced per billion dollars of inflation-adjusted R&D spending has fallen roughly by half every nine years. Better computing, genomics, automation and molecular biology did not translate automatically into better R&D productivity.

The financial consequences remain visible today. Deloitte estimates that the average cost of developing an asset from discovery to launch reached approximately $2.67 billion in 2025, up from $2.23 billion a year earlier. Forecast returns on late-stage pharmaceutical R&D have recovered to around 7.0%, but that improvement remains concentrated in a relatively small number of high-value programs.

Artificial intelligence is therefore arriving at a particularly important point in the industry's evolution.

Its significance is not simply that researchers can perform existing tasks faster. The deeper change is architectural. AI allows companies to move away from a sequential R&D model — discovery, optimization, testing, trials, approval — toward a system in which data generated at every stage continuously improves the decisions made at every other stage.

The new R&D playbook is increasingly a learning loop rather than a pipeline.

The problem is not experimentation. It is experiment selection.

Traditional R&D is constrained by physical experimentation.

Compounds have to be synthesized. Biological systems have to be tested. Patients have to be enrolled. Clinical outcomes often require months or years to materialize. Those constraints do not disappear because computing becomes cheaper.

What AI changes is what happens before scarce experimental capacity is consumed.

The old discovery model was heavily empirical: identify a promising biological target, screen enormous libraries of compounds, find a hit and improve it through repeated experimentation.

The emerging model is increasingly generative.

Protein-structure models can predict biological structures at enormous scale. Generative chemistry systems can propose molecules optimized simultaneously for potency, selectivity, toxicity, synthesizability and other characteristics. Multimodal models can search relationships across genomic, imaging, pathology and clinical datasets that would be impossible for individual research teams to process manually.

The objective is not to eliminate experimentation. It is to increase the expected information value of the next experiment.

That distinction matters.

If models can eliminate low-probability candidates before they enter a laboratory, R&D organizations can direct expensive wet-lab capacity toward better hypotheses. The productivity metric consequently changes from how many experiments can we run? to how much uncertainty can each experiment eliminate?

That is a much more consequential use of AI than simply automating research workflows.

From pipeline to closed-loop discovery

The model becomes even more powerful when AI is connected directly to laboratory automation.

A conventional discovery workflow follows a familiar cycle: design a molecule, synthesize it, test it, analyze the result and use those findings to design another molecule.

Every handoff introduces time.

The AI-native laboratory compresses those handoffs. Models generate candidate designs. Robotic laboratories synthesize and test them. Experimental results return to the model. The model updates and recommends the next experiment.

The result is a design–make–test–learn loop.

Companies including Recursion, Isomorphic Labs and other AI-native biotechs are effectively attempting to build this capability at platform scale. The valuable asset is no longer simply one molecule. It is the infrastructure that can repeatedly produce and improve molecules.

This is the first major shift in the new playbook:

from pipeline to platform.

A conventional biotech may ultimately live or die on a small number of assets. An AI-native R&D platform attempts to create a repeatable discovery engine in which every experiment improves the probability distribution for the experiments that follow.

The early clinical evidence is encouraging, although it needs to be treated carefully.

Research published in Drug Discovery Today found that AI-discovered molecules in the sample achieved approximately 80–90% Phase I success, versus historical industry benchmarks around 50%. Phase II success was closer to 40%, where the advantage over historical benchmarks became considerably smaller.

That distinction is critical. AI appears increasingly capable of designing molecules with attractive drug-like characteristics. Whether it can consistently identify the right biology remains a harder question.

But even modest improvements matter because the most destructive R&D failures are not early failures. They are late failures, after hundreds of millions of dollars and years of development have already been committed.

Clinical development may be the bigger prize

Drug discovery attracts most of the headlines around AI. But clinical development may ultimately offer the larger economic opportunity.

Trials consume enormous amounts of both time and capital, and many of their bottlenecks are fundamentally information problems.

Which population should be enrolled? Which sites will recruit successfully? Which eligibility criteria will exclude too many patients? Which endpoints provide the clearest signal? Which subgroups are most likely to respond?

AI increasingly allows sponsors to model these decisions before committing capital to execution.

Models can analyze thousands of historical protocols to identify criteria associated with slow enrollment or amendments. Electronic health records and claims data can improve site and patient identification. Digital biomarkers can expand measurement beyond conventional endpoints. Real-world evidence can provide additional context around safety and effectiveness. Generative systems can accelerate documentation, medical writing and regulatory preparation.

In other words, the clinical trial itself becomes more computational.

And that is already visible in regulatory activity. FDA has accumulated experience with more than 500 drug and biological-product submissions containing AI components since 2016. Earlier FDA analysis shows that AI-related regulatory applications are overwhelmingly concentrated in clinical development, far more than in discovery, nonclinical development or post-market use.

The market is therefore moving beyond the idea of AI as a discovery technology.

It is becoming R&D infrastructure.

Healthcare data becomes part of the R&D stack

This also changes the boundary between pharmaceutical R&D and healthcare delivery.

Hospitals, diagnostics companies, payers and health systems possess precisely the multimodal datasets that increasingly sophisticated models require: longitudinal medical records, imaging, pathology, genomic information, prescribing behavior and real-world outcomes.

Historically, much of that information was downstream of pharmaceutical development.

In the AI-era R&D model, it becomes an input.

That makes the quality, accessibility and interoperability of healthcare data strategically important to pharmaceutical companies. Data partnerships, federated-learning models, real-world-evidence capabilities and common biological ontologies become part of the innovation infrastructure.

The implication is subtle but important:

the data generated after a drug enters the healthcare system can eventually improve the discovery of the next drug.

Discovery informs clinical development. Clinical development creates biomarkers and treatment-response data. Commercial use produces real-world evidence. Real-world evidence reveals new patient subgroups or biological signals. Those signals return to discovery.

The R&D organization becomes a network

The organizational architecture is changing at the same time.

Global pharmaceutical R&D spending increased from approximately $144 billion in 2014 to $251 billion in 2022 and is projected by McKinsey to approach $350 billion by 2029.

But external R&D spending has expanded even faster.

Spending through CROs and CDMOs increased from around $32 billion to $82 billion between 2014 and 2022 and could reach approximately $135 billion by 2029.

That means tomorrow’s R&D organization may resemble an orchestrated network more than a vertically integrated laboratory.

Pharma companies will combine internal biological expertise with AI platforms, specialist biotech capabilities, CRO infrastructure, automated laboratories, academic research, external data and technology partners.

The competitive question becomes less about owning every capability and more about integrating them without losing the learning loop.

That requires a very different data architecture.

If assay results sit in one system, clinical data in another, external evidence under incompatible licenses and scientists' experimental knowledge remains buried in PDFs and notebooks, increasingly powerful models cannot learn effectively across the organization.

Every experiment increasingly needs to produce not only a scientific answer, but also machine-readable institutional memory.

Validation becomes part of the product

Healthcare also places a hard boundary around the AI thesis.

A plausible answer is not enough.

Drug-development decisions need to be explainable, reproducible and ultimately defensible to regulators. Model quality therefore cannot be separated from model governance.

FDA's emerging framework emphasizes a defined context of use, appropriate data, risk-based credibility assessment and model validation. Regulators are not treating AI as inherently unacceptable. They are asking whether evidence generated with it is sufficiently trustworthy for the decision being made.

That changes the definition of a good AI system.

The winning model may not always be the model with the highest benchmark performance. It may be the one whose training data, assumptions, outputs, limitations and performance can be traced and validated.

For life sciences, trust becomes an engineering requirement.

This is why simply giving scientists access to a large language model will not transform R&D. The larger opportunity requires rebuilding processes, data architecture, validation, incentives and organizational boundaries around AI-enabled decision-making.

The real advantage is learning velocity

Five shifts therefore define the emerging playbook.

R&D moves from pipeline to platform, because value increasingly comes from repeatable discovery engines rather than individual assets.

It moves from data exhaust to data asset, because every experiment should create reusable information.

It moves from sequential to parallel, as biology, chemistry, safety and clinical design can increasingly be optimized together.

It moves from build everything internally to orchestrate an ecosystem, as external AI, data, CRO and biotech partners become embedded in R&D.

And it moves from scientists versus machines to scientists with machines, where biological judgment and computational capabilities increasingly operate together.

None of this eliminates the fundamental uncertainty of biology.

No AI system can guarantee that manipulating a target will improve human disease. Samples of AI-discovered clinical assets remain small. Phase III and commercialization will ultimately provide much stronger evidence than impressive discovery benchmarks. Bias, intellectual property, privacy, compute costs and regulatory complexity remain meaningful constraints.

But AI does not need to eliminate drug-development risk to reshape its economics.

It needs to change when risk is identified and how cheaply companies can learn from failure.

That may ultimately be the biggest difference between the old playbook and the new one.

The traditional pharmaceutical R&D organization attempted to move promising assets efficiently through a pipeline.

The AI-era organization attempts to make every asset, every experiment and every patient observation improve the system that chooses what to do next.

In the old R&D playbook, information moved down the pipeline.

In the new one, it moves around the loop.

And the companies that learn fastest may ultimately be the ones that discover fastest.

Sources & References

Deloitte. (2026). Navigating the GLP-1 boom. https://www.deloitte.com/uk/en/Industries/life-sciences-health-care/research/measuring-return-from-pharmaceutical-innovation.html

Madura KP Jayatunga, Margaret Ayers, Lotte Bruens, Dhruv Jayanth, Christoph Meier,

How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons,

Drug Discovery Today, Volume 29, Issue 6, 2024, 104009, ISSN 1359-6446, https://doi.org/10.1016/j.drudis.2024.104009 

Nature. (2016). New Tricks for Old Drugs. https://www.nature.com/news/polopoly_fs/1.20091%21/menu/main/topColumns/topLeftColumn/pdf/534314a.pdf

OECD. (2026). Artificial Intelligence in Science. https://www.oecd.org/en/publications/artificial-intelligence-in-science_a8d820bd-en/full-report/eroom-s-law-and-the-decline-in-the-productivity-of-biopharmaceutical-r-d_f42df75c.html 

Scannell, J., Blanckley, A., Boldon, H. et al. Diagnosing the decline in pharmaceutical R&D efficiency. Nat Rev Drug Discov 11, 191–200 (2012). https://doi.org/10.1038/nrd3681