The arithmetic alone is enough to induce vertigo. Pharmaceutical companies now spend well into the billions to shepherd a single drug through to approval—a journey that typically devours 10 to 15 years. But here's the number that should truly alarm industry executives: up to 60% of newly opened clinical trial sites don't make it past their first year, according to data compiled last October.
For an industry that has watched R&D productivity flatline even as costs metastasize, the moment for a genuine reset has arrived. Whether that reset actually materializes is another question entirely.
The latest answer—or at least the latest wager—comes in the form of agentic artificial intelligence. These aren't the predictive models or image classifiers that have been kicking around research labs for years. Agentic systems purport to reason, plan, and execute across sprawling workflows, functioning less like tools and more like tireless colleagues who never sleep and rarely make transcription errors.
In recent months, the biopharma sector has backed this vision with genuinely staggering sums. Isomorphic Labs, the Alphabet-backed drug design venture, closed a $2.1 billion Series B this past May. Eli Lilly pushed its collaboration with Insilico Medicine to $2.75 billion in total potential value by March. Even second-tier players are writing substantial checks: Incyte handed Genesis Therapeutics $80 million in May to expand a partnership built around Genesis's 3D Pearl platform and other orchestrated models.
The pitch is seductive. Collapse timelines from years to months. Automate the soul-crushing documentation that regulatory submissions demand. Design clinical trials that actually hit enrollment targets. Yet beneath the deal announcements and the consultant projections—McKinsey suggests EBITDA lifts of 3.4 to 5.4 percentage points over the next few years—lies a far messier reality. Validation frameworks remain incomplete. Regulatory guidance is still evolving. And industry observers note that many companies report limited realized value from their AI investments so far, a pattern McKinsey has documented in its research on agentic AI adoption.
Where the Money Is Going
Pinning down the exact size of the AI-in-drug-discovery market requires a tolerance for ambiguity. Analyst estimates for 2026 range from around $2.9 billion to north of $3.25 billion, depending on methodology and what counts as "AI." But the directional trend is unmistakable: rapid expansion across three overlapping domains—drug discovery, clinical development, and regulatory operations.
On the discovery side, the players have matured well beyond academic proof-of-concepts. Isomorphic Labs rolled out its IsoDDE platform in February, positioning it as more than a binding affinity predictor—it's meant to handle candidate optimization end-to-end. Recursion and Exscientia merged late in 2024, creating what amounts to a vertically integrated AI discovery factory with shared datasets and infrastructure. These are production systems now, backed by pharma partnerships worth hundreds of millions of dollars.
Clinical development has spawned its own insurgent class. QuantHealth announced in March that its LRDM v1.0 foundation model now powers a portfolio north of 600 simulated trials, with vendor-reported accuracy claims hovering around 90% across roughly 30 disease areas—though independent verification of these performance metrics remains pending. Phesi's Trial Accelerator platform crossed the 200 million patient threshold in real-world data by this year, offering enrollment forecasts designed to sidestep the very site-failure crisis that's plaguing the industry. Unlearn.AI secured EMA qualification and announced new ALS trial partnerships in 2026, pushing digital twins and synthetic control arms into the mainstream.
Then there's the documentation layer. Unglamorous, perhaps, but potentially the most immediately monetizable wedge. Vendors like Skaldi, Asthra AI, Indago, and Zenopsys now market automated drafting of investigational new drug applications, protocols, statistical analysis plans, clinical study reports, even full eCTD dossiers. A March case study—attributed to a top-10 pharma company but never publicly identified—claimed $2.1 million in annual savings and reduced dependence on outsourced medical writers. Independent verification of that figure? Still pending.
What binds these segments is infrastructure ambition. Eli Lilly and NVIDIA announced last October they're building what Lilly describes as pharma's most powerful AI supercomputer, purpose-built for "scientific AI agents" that plan experiments and reason through datasets autonomously. NVIDIA's BioNeMo microservices, refreshed in May, aim to turn biology and chemistry workloads into composable agent tasks. The underlying bet is stark: the next generation of R&D won't just use AI. It will be orchestrated by it.
Why Now?
Three forces have converged to push agentic AI from speculative to strategic.
First, the productivity crisis isn't improving—it's calcifying. Deloitte's annual pharma innovation index showed top-20 companies' late-stage internal rate of return ticking up to 7.0% last year, a modest gain driven largely by GLP-1 blockbusters. Strip out those outliers and the underlying picture remains bleak. A review of more than 19,000 cardiovascular trials, published recently in JACC Advances, documented persistent enrollment and protocol-design failures. Another study examining over 10,000 phase III trials found premature termination rates disturbingly high across therapeutic areas. The industry's established approach—sequential, siloed, heavily manual—faces mounting evidence that it cannot deliver the productivity gains stakeholders demand.
Second, the technology has crossed a threshold. FDA guidance issued early last year noted the agency has received more than 500 submissions with AI components since 2016, a quiet signal that AI has already embedded itself in regulatory processes. The agency's April announcement of "proof of concept" real-time trial monitoring with Amgen and AstraZeneca suggests regulators are willing to experiment, provided transparency and auditability are ironclad.
Standards, too, are finally coalescing. ICH M11, which introduces structured, computable protocols, gained FDA backing with guidance updates as recently as May. CDISC sessions earlier this year emphasized protocol-to-SDTM automation via the Unified Study Definitions Model. eCTD v4.0 migration is underway. These aren't minor technical updates. They're the plumbing that makes agent-generated, end-to-end submission packages conceivable in the first place.
Third—and perhaps most revealing—there's urgency emanating from the C-suite. Jensen Huang told an audience in January that Eli Lilly's investment in scientific AI agents reflects a belief that experiment planning, not just computation, can be delegated to machines. OpenAI's lobbying push in April claimed lab automation could compress timelines from months to days, though it conceded the proof is still forthcoming. Ben Liu at Formation Bio put it more bluntly in a February interview with Time: "The biggest problem in drug development hasn't been drug discovery for a long time." The real bottleneck sits downstream—in trials, operations, documentation. Exactly where agents promise the most leverage.
The Contenders
The competitive landscape sorts roughly into three tiers: platform giants, clinical specialists, and regulatory automation upstarts.
Platform giants like Isomorphic, Recursion, and Insilico Medicine are constructing vertically integrated stacks spanning target identification through lead optimization. Isomorphic's $2.1 billion raise in May signals investor conviction that AI-designed candidates can advance to the clinic at commercial scale. Recursion's post-Exscientia merger gives it petabyte-scale datasets and high-throughput screening infrastructure most academic labs can only dream about. Insilico's expanded Lilly partnership—now valued at $2.75 billion in total potential—covers multiple programs and reflects a shift from one-off tool licensing to strategic R&D alliances.
Clinical design and execution specialists are attacking trial productivity from multiple vectors. QuantHealth's foundation model approach attempts to simulate trial outcomes before enrolling a single patient, a hedge against that brutal 45-60% first-year site failure rate. Phesi layers real-world data—more than 200 million patients at last count—to surface enrollment risks and competitive dynamics. Unlearn's digital twin methodology, now EMA-qualified and aligned with FDA guidance, aims to shrink placebo arms and accelerate readouts in rare diseases. Its ALS partnerships with VectorY and SOLA Biosciences, announced this year, offer early tests of that thesis.
Then there's the regulatory automation tier, where startups like Enjamb Labs are making early-stage bets. Founded in 2025 and accepted into Y Combinator's Spring 2026 batch, Enjamb raised $650,000 in April with a pitch centered on eliminating "handoff losses"—the gaps created when data in Benchling, statistics in JMP, and regulatory documents drafted by consultants never quite align. The two-person team, led by ML researcher Rayan Mubarak and HPC specialist Maadhav Deekshitha, is building what it describes as a "full stack AI agent" workspace integrating 66 scientific databases, Python/R compute environments, and document editors for IND packages, protocols, statistical analysis plans, and NDA submissions. The company claims adoption at 11-plus institutions, though at this nascent stage that likely translates to academic pilots rather than multi-site pharma rollouts.
Other entrants in this tier—Skaldi for clinical study reports and investigator brochures, Indago for IND submissions, Zenopsys for eCTD dossiers—are marketing similar value propositions: auditable provenance, ICH alignment, dramatic time compression. The Zelthy case study from March, citing that unnamed top-10 pharma and its $2.1 million in annual savings, didn't disclose validation methodology or the customer's identity. Which, in this industry, raises its own questions.
Sanofi offers a window into what internal adoption looks like at scale. The company's "plai" and "Concierge" agentic apps reportedly drew 20,000 daily users as of recent filings, spanning quality control to portfolio decision-making. Novartis has staffed AI-driven evidence generation roles and implemented board-level AI governance, according to job postings and corporate updates through mid-year. IQVIA, whose R&D Solutions segment pulled in nearly $9 billion in revenue last fiscal year, publicly describes AI as reshaping the R&D lifecycle—though executives cautioned in May that "credible signals" remain scarce.
What Happens Next
The coming 18 months will clarify whether agentic AI in biopharma represents a genuine inflection point or just another layer of vendor sprawl.
Regulatory scrutiny is intensifying in lockstep with adoption. The EU AI Act's phased enforcement began in August, with full high-risk system obligations arriving by August 2027. That intersects directly with the EMA's September 2024 reflection paper on AI in the medicinal product lifecycle. Providers building agent systems for European sponsors will need to map transparency, risk management, and data governance obligations across both frameworks simultaneously. In the U.S., FDA's draft guidance from early last year on AI in regulatory submissions emphasizes context-of-use documentation, performance drift monitoring, and AI's effect on overall evidence packages—requirements that favor vendors with auditable, source-level traceability over black-box architectures.
Standards convergence matters more than most executives seem to appreciate. ICH M11's structured protocols and eCTD v4.0's modernized submission format aren't mere compliance checkboxes. They're the rails that enable end-to-end agent workflows carrying context from literature review through SDTM mapping to final clinical study report. CDISC's sessions this year on protocol-to-SDTM automation and Analysis Results Standards reflect an industry beginning to think in machine-readable terms. Startups aligning early to these standards will face an easier adoption path than those constructing proprietary formats.
Validation, though, remains the elephant in every conference room. Multiple systematic reviews published over the past year—including a February scoping review in PLOS ONE and an AHRQ living surveillance report—document time reductions north of 50% in evidence synthesis tasks. But they also stress the imperative for quality control and human-in-the-loop governance. Cochrane launched a platform study in 2026 to evaluate AI tools across the evidence synthesis lifecycle. ISPOR proposed a transparent framework for generative AI in health economics aligned with NICE and Canadian guidelines. The message is remarkably consistent: speed without rigor is liability, not innovation.
McKinsey's projection of a 3.4 to 5.4 percentage-point EBITDA lift sounds compelling until you read the fine print: many companies still report limited realized value from their AI investments. BCG's series on agentic AI in healthcare, running from January through April, emphasizes that technology constitutes only half the equation. Operating model overhauls, data foundations, governance frameworks—those are the other half. And they're considerably harder to purchase off the shelf.
The Stakes
The opportunity is real enough. Drug development timelines and costs have remained stubbornly elevated despite decades of incremental tooling improvements. Trials fail at rates that would be unacceptable in nearly any other industry. Regulatory submissions are document-assembly nightmares consuming hundreds of person-hours per module, sometimes more. Agentic AI offers a plausible route to compression.
But only if vendors, sponsors, and regulators can align on transparency, auditability, and shared standards. Only if the validation studies catch up to the marketing claims. Only if the technology proves adaptable enough to handle the messy, exception-filled reality of actual drug development rather than idealized workflows.
For now, the smart money is watching for concrete proof points. Which AI-designed molecules actually advance through phase II? Which trial simulations accurately predict enrollment curves in the wild? Which automated clinical study reports sail through FDA review without triggering major information requests?
The billion-dollar deals signal belief. The validation gaps signal caution. Somewhere between those two poles, the next generation of biopharma R&D is taking shape—haltingly, expensively, and with far more questions than answers. That's often how these transformations begin. Whether this one delivers on its extraordinary promises remains very much an open question.
