Bristol Myers Squibb makes Claude its "shared intelligence platform" — and the AI-scientist papers land in Nature
BMS deploys Claude Enterprise across 30,000+ employees as the unified agent layer. DeepMind's Co-Scientist and FutureHouse's Robin clear Nature peer review the same week Edison Scientific lands at Incyte. Plus: the EU AI Act high-risk draft, Helio × Syneos, and a Benchling adoption snapshot.
This week pharma's enterprise-AI race tipped decisively toward the agent layer. Bristol Myers Squibb named Claude its "shared intelligence platform" for 30,000+ employees, capping a six-week run of nine- and ten-figure commitments from Merck, Novo Nordisk, Lilly, and AstraZeneca. Underneath the megadeals, the AI-scientist category earned its first real validation — two Nature papers and a live Incyte deployment in the same week — while regulators in Brussels and White Oak opened the consultation windows that will shape how any of this actually ships. The throughline: the competitive edge in pharma AI is shifting from which model you pick to the data you feed it and the workflows you wrap around it.
Lead — BMS makes Claude its "shared intelligence platform"
DealsOn May 20, Bristol Myers Squibb and Anthropic announced a strategic agreement deploying Claude Enterprise across research, clinical development, manufacturing, commercial, and corporate functions — more than 30,000 employees. The near-term focus is agentic: Claude Code for engineering, embedded agents in commercial and medical affairs, and secure connections to BMS's institutional knowledge, all under governance and audit controls. BMS explicitly names Claude its "shared intelligence platform" while framing the choice as part of a "deliberate multi-vendor strategy" — standardizing one frontier model at the agent layer while still evaluating others below it. That's a different posture than Merck's all-Google bet or Novo Nordisk's all-OpenAI bet, and it's the first of the recent megadeals where a frontier vendor's flagship is the unified surface employees actually touch. No productivity numbers were disclosed, and with three-plus years of prior internal AI investment, Claude is replacing a working baseline, not arriving green-field. The real test is whether agentic read-access finally unifies the regulatory, clinical, and commercial knowledge that traditional knowledge management never could.
Two "AI scientist" papers land in Nature — one lands at Incyte
PapersDeepMind's Co-Scientist and FutureHouse's Robin moved from bioRxiv preprints to peer-reviewed Nature publications the same week Edison Scientific — FutureHouse's commercial spinout — announced its first major pharma deployment with Incyte. Both systems run the same pattern: an LLM coordinator orchestrating sub-agents for literature review, hypothesis generation, experimental design, and interpretation; the Robin paper reports a generated hypothesis validated in patient-derived cells, not an immortalized line. Edison spun out in 2025 with $70M from Spark Capital, Triatomic Capital, and others, and its May 19 agreement positions Kosmos as a continuously learning system inside Incyte's discovery and development data. Incyte R&D head Pablo Cagnoni frames the deal as turning company data into a "compounding asset." Peer review plus a real mid-cap deployment is the strongest validation this hyped, under-tested category has received — and contracting terms remain unusually favorable while the field is still uncrowded.
Helio Genomics × Syneos pair AI liquid biopsy with CRO commercial muscle
DealsAI TechBio Helio Genomics is partnering with CRO giant Syneos Health to drive adoption of its blood-based test for early detection of hepatocellular carcinoma. Helio's assay reads methylation patterns in cell-free DNA through an AI classifier trained to flag early-stage HCC in at-risk populations — people with hepatitis B, hepatitis C, or NASH-related cirrhosis. What makes the deal unusual is the role of the CRO: not trial logistics but go-to-market commercialization, adding market access, medical affairs, and physician engagement that Helio lacks in-house. Liquid-biopsy MRD assays have been technologically credible for years; the binding constraint is now adoption pathways through community oncology and hepatology. Pairing an AI developer with a CRO's commercial-services infrastructure is a post-clearance distribution model worth watching, though primary-source coverage of the deal terms is still thin.
EU AI Act high-risk classification guidelines drop; consultation closes June 23
RegulatoryOn May 19 the European Commission published draft guidelines for classifying high-risk AI systems, opening a consultation ahead of the AI Act becoming fully applicable on August 2, 2026. For any biopharma running clinical decision support, biomarker classifiers, trial-eligibility AI, or AI-driven pharmacovigilance in EU markets, these guidelines set the conformity-assessment and post-market monitoring obligations that go live in roughly ten weeks. The treatment of general-purpose LLMs sitting on top of clinical data is one of the contested zones — directly relevant to anyone deploying agentic AI of the BMS-Anthropic variety in EU operations. Expect industry feedback to push for narrower scope on general-purpose use cases, broader transparency carveouts for proprietary models, and clearer "human-in-the-loop" qualifications. That last line determines whether a tool faces the full conformity regime or a lighter touch.
Recursion's Najat Khan keeps setting the "tangible proof points" bar
OpinionRecursion CEO Najat Khan continues to lean into a "tangible proof points" framing — an explicit contrast against the AI biotech sector's overreliance on platform narratives — as the company heads toward Phase I readouts. Her rhetoric is becoming the industry's de facto positioning template; competitors are already echoing "we're past the model era, into the proof era." The asset to watch is REC-1245, targeting RBM39, with a Phase I safety and PK monotherapy readout guided for biomarker-enriched solid tumors and lymphoma in the first half of 2026. For investors, the question is whether the MMAI Gym infrastructure and pharma partnerships with Roche/Genentech and Sanofi translate into the asset-level data that makes the stock work. The goalposts for the entire AI-biotech subsector have moved from model claims to clinical proof.
Karpathy joins Anthropic's pre-training team
Industry dataAndrej Karpathy — OpenAI founding member and former Tesla AI lead — joined Anthropic's pre-training team on May 19. Talent gravity matters for pharma because the underlying model capabilities those teams ship determine whether agentic deployments like BMS-Anthropic actually deliver on the data-unification promise. When a company bets its entire enterprise AI stack on one frontier model, the strength of that model's core research bench becomes a form of vendor risk.
Insilico × Eli Lilly $2.75B deal keeps rippling
DealsInsilico Medicine's back-loaded AI drug discovery deal with Eli Lilly continues to move the sector, but the story is the structure, not the headline number. $115M upfront against $2.75B in development, regulatory, and commercial milestones has become the dominant pricing pattern for pharma-AI deals. That structure means AI biotech valuations now live and die on milestone-trigger probabilities — how likely each staged payment is to actually fire. For anyone modeling these companies, the diligence has shifted from platform narrative to the credibility of the milestone schedule.
FDA early-phase AI trial pilot — comment window closes May 29
RegulatoryThe FDA's AI-Enabled Optimization of Early-Phase Clinical Trials Pilot has a comment window closing May 29, and it's worth submitting even a short response. Phase 1 efficiency is where AI has the most measurable near-term impact, and the agency is signaling unusual willingness to shape the program around industry input. Leaving the feedback to trade associations alone would cede the framing on a program that could set early-phase AI norms. This is a low-cost, high-leverage moment to influence how the FDA structures the effort.
Benchling's 2026 report puts hard numbers on AI adoption
Industry dataBenchling's 2026 Biotech AI Report quantifies where AI is actually being used: 76% literature review, 71% protein structure prediction, 66% scientific reporting, 58% target ID, 42% generative design, and just 29% ADME prediction. The gap between literature review and ADME tells the whole story about where data infrastructure is mature versus where it isn't. The next 18 months of biotech AI investment will be about closing that ADME-adjacent gap, and the companies that solve it first will compound. Adoption clusters where clean, abundant data already exists — and stalls where it doesn't.
NVIDIA × Lilly "Lillypod" lab staffs up in South San Francisco
InfrastructureThe NVIDIA-Lilly "Lillypod" co-innovation lab is now staffing in South San Francisco: $1B over five years, marketed as "the most powerful supercomputer in pharma." Compute is increasingly the differentiator, and Lilly has now stacked NVIDIA infrastructure, Insilico models, and Chai biologics access. That's a vertically integrated AI R&D stack no other pharma can replicate without comparable capital commitments. The move signals that at the frontier, infrastructure ownership — not just model licensing — is becoming a competitive moat.
Owkin spins out Waiv with $33M
FundingOwkin has spun out Waiv — formerly Owkin Dx — as a standalone company with $33M, led by OTB Ventures and Alpha Intelligence Capital. The notable pattern is the partition: AI diagnostics is being split from AI drug discovery as a separate fundable business with its own investor base. Expect more carve-outs from companies that built diagnostics and discovery in parallel, since the two now attract distinct capital and distinct comparables. The structural separation is a tell about how the market is learning to value each half.
The Onion Desk
This week's satire: AstraZeneca licenses Owkin's "AI Scientist" platform, reportedly to help locate AstraZeneca's previous AI scientist, last seen submitting a competitive-intelligence dashboard in Q3 2025. Verily raises $300M in a round investors describe as "definitely not just Alphabet putting money in a different pocket." The FDA's Elsa 4.0 now asks reviewers if they've tried turning the drug off and on again, and big-pharma earnings calls are "legally required" to include a slide quantifying AI's "measurable impact" — Pfizer's reportedly reads only "AI: Yes." Rounding out the desk: Anthropic acquires eight-month-old Coefficient Bio (whose team it "already has here, but with different LinkedIn URLs"), BMS plans to run all 12 licensed Hengrui programs through Claude "just to see what happens," Medidata and CRIO promise to solve data friction by generating higher-quality friction, and Iambic updates its deck to promise IAM1363 Phase 1 data "definitely this year, definitely."