When Eli Lilly signed a deal worth up to $2.75 billion with AI drug developer Insilico Medicine in March 2026, it was the latest in a cascade of nine-figure and ten-figure commitments from the world’s largest pharmaceutical companies to a single emerging technology. Takeda had just struck a potential $1.7 billion collaboration with Iambic Therapeutics. AstraZeneca had committed over $200 million to a raft of AI partnerships. Nvidia and Lilly had already announced a $1 billion joint laboratory. The pattern is unmistakable: every major name in global pharma is moving at speed into agentic AI in drug discovery, and the sums involved are growing.
For those outside the sector, the term may still be unfamiliar. What is agentic AI in drug discovery, how does it work inside a pharmaceutical pipeline, and what is it about this technology specifically that has prompted the largest coordinated capital commitment the industry has made to a single scientific approach in a generation?
Beyond the Chatbot: What Agentic AI in Drug Discovery Actually Means
Most people who have encountered artificial intelligence have experienced its assistive form: a system that responds to a question, generates a document, or predicts a protein structure when prompted. These tools are genuinely powerful, but they are fundamentally reactive. They execute a single task and stop.
Agentic AI is architecturally different. According to BioPharm International, it refers to autonomous systems that can reason, plan and execute multi-step workflows with minimal direct human intervention. Rather than waiting for instruction at every step, an agentic system sets its own sub-goals, selects tools, runs experiments, analyses results, and revises its approach within a continuous loop. It does not pause between tasks; it iterates.
The underlying architecture, as described in a 2026 review published in Drug Discovery Today, combines large language models with specialised tools for perception, computation, action and memory. In practical terms, this means a single agentic system can simultaneously mine scientific literature, design candidate molecules, predict their toxicity, and route promising compounds to robotic laboratory platforms for physical synthesis, all without a researcher managing each handover. Early deployments have compressed literature analysis from weeks to minutes, and assay development from months to hours.
Shweta Maniar, Global Director for Strategic Industries, Healthcare and Life Sciences at Google Cloud, has drawn the distinction between the two paradigms directly. Writing in PharmaTimes, she noted that assistive AI gave researchers better tools to predict specific data points, but did not solve the underlying methodological bottleneck. Agentic AI, she argued, provides autonomous systems that can reason, plan and execute entire complex workflows. The shift, in her framing, is from a smarter instrument to an independent colleague.
Inside the Pipeline: Where Agents Are Being Deployed
Traditional pharmaceutical R&D follows a largely sequential path: identify a biological target implicated in a disease, screen chemical libraries for compounds that interact with it, optimise a lead molecule for potency and safety, then run through preclinical and clinical validation. Each stage takes years. Costs escalate sharply the further a candidate advances before failing, which happens around 90 per cent of the time.
Agentic systems do not simply accelerate individual steps; they attack the architecture of the process itself. In target identification, knowledge graph platforms mine scientific literature, patents, clinical data and genomic datasets simultaneously, surfacing non-obvious connections between biological mechanisms and diseases. BenevolentAI built its early reputation on precisely this approach: its system identified baricitinib, an existing rheumatoid arthritis drug, as a potential treatment for COVID-19 in 2020, a finding validated in subsequent clinical trials.
In molecular design, generative models propose novel compounds from scratch rather than screening existing chemical libraries. These systems explore chemical spaces that human medicinal chemists could never navigate manually, designing molecules optimised across multiple properties at once: potency, selectivity, metabolic stability and synthetic accessibility. Schrödinger’s platform combines quantum mechanical simulations with machine learning to predict molecular behaviour at the atomic level, a physics-based approach that produces unusually precise binding predictions for difficult targets.
The most significant frontier is the connection between computational and physical laboratory work. Johnson and Johnson is deploying agentic AI to manage autonomous drug discovery workflows, including determining optimal timing for critical steps in chemical synthesis. AstraZeneca has integrated a multi-agent system called ChatInvent, documented in a 2026 ScienceDirect paper, into its discovery pipeline to handle molecular design and synthesis planning. In a separate development this month, AstraZeneca announced a three-year licensing agreement with AI firm Owkin, deploying agentic tools to monitor competitive intelligence across clinical trials and disease areas. Moderna has taken a broader approach, deploying AI agents across operational functions including regulatory document drafting and patient communications.
Elsewhere, AWS launched Amazon Bio Discovery in April 2026, an agentic platform giving researchers access to a library of biological foundation models, an AI agent for experimental design, and integrated wet laboratory partners that return results for rapid iteration.
“AI agents make powerful scientific capabilities accessible to all drug researchers, not just those with computational expertise,” said Rajiv Chopra, Vice President of AWS Healthcare AI and Life Sciences.
Eli Lilly’s $1 billion laboratory with Nvidia is designed to build a continuous learning system connecting agentic computational and wet laboratory environments around the clock. The competitive implications of these deployments extend well beyond the laboratory, reshaping how pharma companies monitor rival pipelines and identify licensing opportunities before they become public knowledge.
The Clinical Evidence That Changed the Investment Calculus
For most of the past decade, investment in agentic AI in drug discovery ran ahead of clinical proof. That changed in June 2025, and the effect on deal-making was immediate.
On 3 June 2025, Insilico Medicine published Phase IIa results for rentosertib in Nature Medicine. The compound is notable because both its biological target, a kinase called TNIK, and the molecule itself were identified and designed using generative AI, making it the first drug of its kind to demonstrate clinical proof-of-concept in a controlled human trial. The 71-patient study in patients with idiopathic pulmonary fibrosis showed a mean improvement in forced vital capacity of 98.4 mL at the 60 mg dose, compared with a mean decline of 20.3 mL in the placebo group over 12 weeks. Insilico had reached preclinical candidate nomination for rentosertib in approximately 18 months from project initiation, against an industry average of two and a half to four years.
A second programme is at a more advanced clinical stage. Zasocitinib, known as TAK-279, was developed through a collaboration between Nimbus Therapeutics and Schrödinger using AI-assisted structure-based design, and subsequently acquired by Takeda. In December 2025, Takeda announced positive topline results from two pivotal Phase 3 studies in plaque psoriasis, with approximately 70 per cent of patients achieving clear or almost clear skin at week 16. Full Phase 3 data were presented at the American Academy of Dermatology annual meeting in March 2026. Takeda plans to file a new drug application with the FDA on the strength of these results, potentially making zasocitinib the first AI-assisted drug to receive regulatory approval.
The aggregate data picture across the broader sector is striking. Phase I success rates for AI-discovered candidates have reached 80 to 90 per cent, against a historical average of around 52 per cent, suggesting that AI platforms are selecting better candidates before they reach human testing. Hit rates from AI-guided molecular screening are running at 16.7 per cent, compared with 0.1 per cent from conventional high-throughput approaches.
“If 2025 was the year of breakthrough research, we believe 2026 will become the year of deployment,” said Jack Dent, Co-Founder of Chai Discovery, an AI-driven biologics company targeting previously undruggable disease mechanisms.
Why the Investment Wave Keeps Growing
The commercial logic driving pharmaceutical investment in agentic AI in drug discovery is not complicated. Patent cliffs are eroding revenues from established blockbusters across the industry simultaneously. The cost of replacing those revenues through conventional R&D, estimated at upwards of $2 billion per approved drug, is prohibitive. Agentic AI offers the prospect of compressing timelines and improving candidate quality at both ends of the pipeline: fewer failures in development, and faster progression for the candidates that survive.
The deal flow since mid-2025 reflects that pressure across every major name in the sector. AstraZeneca has committed over $200 million across partnerships with BenevolentAI, CSPC Pharmaceuticals, VantAI and Immunai. Bristol Myers Squibb collaborated with Owkin for more than $80 million to pursue novel immuno-oncology targets. Roche and Genentech signed a deal valued at over $150 million with Recursion Pharmaceuticals. In February 2026, Takeda committed a potential $1.7 billion to Iambic Therapeutics. In March 2026, Lilly’s $2.75 billion deal with Insilico Medicine granted it exclusive global rights to develop therapies discovered using Insilico’s platform.
Thomas Fuchs, Eli Lilly’s Chief AI Officer, has been public about the company’s strategic thesis.
“The greatest AI advancements will come from the combination of our proprietary data, compute investments to train large foundation models, and deploying that tech to thousands of chemists and biologists,” he said.
Takeda’s Research Chief Andy Plump has been equally direct about the competitive stakes. “The winners over the next five years” will be the companies that “fully integrate” artificial intelligence into drug development, he told BioPharma Dive. Cumulative deal value across pharma-AI partnerships exceeded $18 billion across more than 120 transactions from 2022 through early 2026. The AI drug discovery market was valued at approximately $1.94 billion in 2025 and is projected to reach $2.6 billion in 2026, on a trajectory towards an estimated $16.5 billion by 2034.
Caution Is Warranted
The investment momentum is real. The risks are also real, and the sector has already produced high-profile casualties that should temper the more expansive claims made on agentic AI’s behalf.
Recursion Pharmaceuticals lost most of its market value following multiple research setbacks. Exscientia, after struggling with a depressed stock price, merged with Recursion. BenevolentAI was delisted from public markets following a difficult run. Each of these companies was regarded, at some point in recent years, as a demonstration case for AI drug discovery’s transformative potential. They are cautionary evidence that platform credibility and clinical validation are not interchangeable.
There are also structural limitations that optimistic projections tend to minimise. Medicinal Chemist Derek Lowe, writing in his widely-read industry blog, reviewed a sample of AI-discovered drug candidates and found that in almost every case the biological targets were already known to be implicated in the disease in question. The AI contribution, in those instances, was to the chemistry rather than to the underlying scientific hypothesis. That is valuable, but it is not the same as autonomous scientific discovery.
Governance is a further unresolved challenge. A survey reported by Axios found that 53 per cent of life sciences firms that have adopted AI have formal policies, and 51 per cent conduct regular audits, but fewer than half have integrated these into a fully mature governance framework. In regulated pharmaceutical development, explainability and auditability are not optional. Agentic architectures are being designed to document their reasoning, but robust regulatory frameworks are still being established: the US FDA published draft guidance on AI in drug development in January 2025, with final guidance expected in the second quarter of 2026.
The Stakes of 2026
The year ahead will produce the clearest evidence yet on whether the claims being made for agentic AI in drug discovery are justified at scale. Between 15 and 20 AI-originated drugs are expected to enter pivotal trials in 2026. Multiple analysts estimate a 60 per cent probability that the first AI-designed drug will receive regulatory approval by 2027. Zasocitinib’s FDA filing, if it proceeds as Takeda has indicated, will be the first direct test of that estimate.
In February 2026, Insilico Medicine and Eli Lilly published a framework in ACS Central Science outlining how an advanced agentic system could integrate target discovery, generative chemistry, automated synthesis, biological validation and clinical planning into a single unbroken workflow. It is a vision, but one being pursued with considerable capital behind it. For a broader view of how artificial intelligence is expected to transform the sector, see Five Ways AI Will Reshape Life Sciences in 2026.
What the industry appears to have concluded, fairly or not, is that the cost of being too slow is greater than the cost of being too early. The deals of the past twelve months suggest that window for deliberation has closed. Whether agentic AI in drug discovery ultimately delivers on its promise will be determined not by the scale of the bets being placed, but by what emerges from clinical trials over the next two years.
