From paper to cloud to agents: the digital transformation of R&D

If you work in biopharma, you’ve certainly heard some version of the AI pitch before. A new technology has come to transform your workflow, compress timelines, cut costs, transform R&D. Maybe you've looked on with skepticism: after all, the last generation of "transformative" software still left you doing most of the work yourself, just on a screen instead of on paper. But something about this time is different: the technology industry, an industry that famously characterized itself as the permanent ‘disrupter’ rather than the ‘disrupted’ has found itself restructured from within, a shift with direct implications for the biotech industry.

The conventional technology diffusion story tells us that consumer and tech-native industries move first, while regulated industries, like healthcare and legal, follow years later, burdened by compliance overhead, liability exposure, and institutional conservatism. That pattern held for cloud, for SaaS, for mobile. But peculiarly, the AI wave has penetrated the most regulated, complex industries early on. More than half of the AmLaw 100 now rely on Harvey for support; the AMA found that 66% of U.S. physicians reported using healthcare AI in 2024, up from 38% in 2023. In other words, the technology became too useful for organizational inertia to ignore: law, medicine, and drug development are all language-dense, judgment-heavy, and dependent on synthesizing large quantities of heterogeneous information.  

That inversion isn’t incidental. While earlier software could store information and move it around, it clearly couldn’t reason across information; frontier models (today’s most advanced, large-scale, generative AI systems) increasingly can, and to a meaningful degree. In this way, the pre-2020 software wave was evolutionary rather than transformative; as useful as software was to organize what scientists produced, software itself could not conduct science.

Beyond just increased efficiency of compute and inference, the ever-expanding context window (the amount of information a model can hold in working memory in a prompting session) has been instrumental in enabling improvement. In late 2023, the best AI models scored 39% on GPQA Diamond, a benchmark of 198 PhD-level multiple-choice questions in biology, chemistry, and physics, where human domain experts score roughly 65%. The amount of context they could work with was roughly 300 pages of text. As of early 2026, AI outperforms experts in expert-level scientific reasoning, and the context window has expanded to over one million tokens, capable of interpreting more and more data types. As any scientist knows, program decisions require triangulating across literature, assay outputs, study reports, protocol amendments, CRO deliverables, meeting notes, regulatory correspondence, and prior internal work, and AI excels at interrogating and synthesizing multimodal data.

To understand why the AI frontier represents a step-change in technological capability, it helps to understand what came before.

The first wave: digital foundations: The last two decades of software in biopharma were largely about moving from paper to the cloud. The focus was on data quality and compliance. This wave produced category-defining companies like Veeva, Medidata, Benchling, and Dotmatics. The alphabet soup of ELN, LIMS, LES, MES, QMS, EDC, CTMS, ETMF moved analog, paper-based data capture workflows to searchable, cloud-accessible systems of record, creating the digital layer on which most modern drug development now runs. But scientists were still responsible for the substantive work: literature review, data interpretation, writing conclusions, and making judgment calls.

The second wave: reasoning on demand. Large language models changed that. For the first time, software could read unstructured data, reason across it, and generate useful outputs: draft text, summarize literature, write code, interpret experimental results. The Benchling 2026 Biotech AI Report finds that 89% of scientists now use copilots or reasoning tools. But LLMs are reactive: they produce an output when prompted and stop. The human still drives every step, deciding what to ask, evaluating the response, and determining what to do next.

The third wave: delegating work. In this series, we define agents as systems that operate in closed loops, autonomously executing multi-step tasks and self-correcting toward a defined goal, rather than producing open-ended output with no capacity to act on or iterate from that output. Importantly, while conventional automation follows predefined rules (if condition X is met, execute action Y), agentic AI generates its own action plan based on the goal it has been given, defining its own conditions for action when confronted with ambiguity. The system breaks a complex goal into executable steps, determines what data it needs, connects to databases, APIs, and external services without user input to acquire said relevant data, synthesizes it against your criteria, critiques its own output, iterates to improve it, and only upon self-review presents a recommendation. Another advantage of the agentic approach is that it enables a multi-agent architecture that mirrors how organizations themselves work. Rather than routing all information through a single system, in a multi-agent architecture, each agent is tuned for different specialties and can connect to different data sources, running in parallel rather than sequentially, and a hallucination in one domain doesn’t corrupt reasoning in another. The human’s role, then, is to set the objective and review the outcome. What does that look like in practice? In this series, we'll dive into this new division of labor in biopharma R&D, and what it means for how drugs actually get developed and into patients.

Agentic AI across R&D Functions

Why segment by function rather than stage? The traditional taxonomy of target ID, lead optimization, preclinical, and clinical trials is less applicable to agentic systems, which are specialized for task type, not where a molecule sits in its pipeline. The prior generation of software was purpose-built for stage: ELN for research, EDC for clinical. Agentic AI is uniquely horizontal; the same cross-cutting capabilities (retrieval, synthesis, reasoning, drafting) deploy across entirely different contexts. The same system that analyzes the competitive landscape can help inform trial design. 

The life sciences industry has widely adopted chatbots for text-heavy, low-risk work: drafting documents, searching literature, summarizing information, and writing code. What comes next is more consequential: biopharma is beginning to recognize the limits of chatbot-style productivity gains, and is now searching for something much more consequential: systems that can structure, execute, and compound across an organization. Over the next seven pieces, we'll trace where agents are starting to do real work across biotech companies, function by function