Utah AI Prescribing Expands as AstraZeneca Opens AI-Driven Lab

A week that showed both sides of AI in biopharma

One week in early October brought two kinds of AI news to the life sciences. In Utah, regulators let software write a patient's first prescription for a medicine. In Cambridge, Massachusetts, AstraZeneca opened a $1 billion research building designed around robotics and agentic AI. On the same day, Genentech lost the executive who led its AI drug discovery work. Dealmaking continued alongside all of this: Shionogi bought a rare disease company and Genentech licensed a brain-delivery asset.

Taken together, these stories suggest the industry's AI ambitions have moved past the pitch-deck stage. Clinical deployment is now outpacing medicine's governance structures, and the AI talent inside large companies is turning out to be less stable than the buildings meant to house it.

Utah moves from AI refills to new prescriptions

The most significant development came from Utah's Office of Artificial Intelligence Policy. It authorized Nolla Health, a New York startup, to use AI to issue initial acne prescriptions. The company describes this as the first U.S. clearance for AI to write new prescriptions rather than renew existing ones.20 Under the pilot, the AI combines a score it generates with the patient's medical information and then recommends one of eight approved treatments.13 Only topical products are allowed, including tretinoin, adapalene, azelaic acid, and clindamycin with benzoyl peroxide. Isotretinoin, oral drugs, systemic hormonal therapy, and treatment for severe acne are all excluded.17

Nolla's case rests on access. The company says only about 10% of Utahns with acne ever see a dermatologist, and the average wait for an appointment in the state is 61 days.20 It also reports that clinicians agreed with the AI's treatment choices in more than 96% of real-world cases.16

Coverage splits sharply on how much human oversight is actually involved. The Los Angeles Times reported that the startup would offer prescriptions without direct human oversight, and it quoted the CEO saying prescriptions will carry a clinician's name even though no human writes them.13 HealthExec's headline also stressed the absence of human oversight.15 Other outlets described a staged structure in which doctors review cases at the start. Under that structure, the AI can submit prescriptions on its own only after at least four weeks, at least 100 patients, physician agreement of at least 95%, no missed safety stops, and no serious adverse events. State approval is still required to move to the next stage.17

The most reasonable reading is that both accounts are accurate but describe different points in time. Physicians review cases now, and the pilot is designed to reduce that review over time. The headline about a world-first reflects where the pilot is designed to end up, not how it operates today.

Doctronic's track record shaped Nolla's pilot

Nolla's careful design reflects what happened with Utah's first healthcare pilot. Doctronic's AI prescription renewal program was approved in October 2025 and publicly announced on January 6.14 It ran into resistance quickly. In April, 11 of the 14 physicians on the Utah Medical Licensing Board signed a letter demanding that the pilot be suspended immediately, and they objected that the board had not been consulted.18 The state kept the program running but removed the sedative butalbital and the arrhythmia drug flecainide from the formulary. It also ordered the company to fix the liability disclaimers in its terms of service.14

The early results were mixed. Over the first several months, the AI recommended renewal in 72% of cases and sent the rest to a human. Physicians agreed with 91% of the renewals, but they judged nearly a third of the escalations to be overly cautious.16 Those figures came from Doctronic's own physicians, and the state has started a separate independent review.18 As of the latest public update, the pilot was still in its first phase, with every renewal reviewed by a physician, and no serious safety incidents had been reported.14 Even the size of the formulary varies across reports, with figures of 190, 191, and 192 medications.111814 That gap is small, but it shows how loosely documented some of these programs remain.

Nolla's approach was noticeably different. It worked with the licensing board before launch, and its pilot added a third-party physician appointed by the board to oversee prescription reviews.13 Utah also announced five new healthcare sandbox pilots and added independent third-party evaluators.12 These include August AI, which will renew chronic medications for $4 per refill, with a clinician reviewing every decision at launch.16 The state is also moving to put physicians back into the oversight structure.12

The main open question in this program is a legal one, not a technical one. A legal analysis noted that sandbox agreements can relax state rules but cannot waive federal law. Whether these systems count as FDA-regulated medical devices therefore remains unresolved.18 Doctronic is already in talks with Arizona, Missouri, Texas, and Wyoming.14 That means Utah's experiment may become a model for other states before its central legal questions are settled.

AstraZeneca builds AI into its labs

On the research side, AstraZeneca opened its new global strategic R&D center at 290 Binney Street in Kendall Square. The building is 18 stories and 570,000 square feet, with ten floors of interconnected labs equipped with robotics, continuous automation, and agentic AI.47 Combined with its genomic medicine site nearby, the campus will hold nearly 2,000 researchers working on oncology, cell therapy, chronic disease, and rare disease. The company says the investment will grow its Massachusetts workforce by more than 50%.41 The site belongs to a $50 billion U.S. investment plan that AstraZeneca first announced in July 2025.48

Most coverage repeated the company's own framing. STAT added useful context by describing Kendall Square as an unsettled center of U.S. biotechnology after a wave of dealmaking with Chinese companies.21 From that angle, the opening is partly a statement of confidence in a U.S. hub that has lost some licensing activity to China. AstraZeneca itself has been an active buyer there, paying $100 million upfront for Jacobio's pan-KRAS inhibitor earlier this year.27 The company is pursuing both strategies at once: in-house AI-equipped labs in the U.S. and external assets sourced from China.

The agentic AI branding deserves some skepticism. An AstraZeneca bioinformatics director wrote this year that AI has not revolutionized drug development. He argued that its clearest gains have come in less visible process work, such as statistical programming and trial operations.26 He also identified the lack of high-quality biological data as the main constraint, and he pointed to robotic wet-lab startups as one possible solution.26 Read that way, a building full of automated labs is less an AI breakthrough than a bet on producing better data at scale, and that may be the more durable strategy.

Genentech loses its AI drug discovery chief

Genentech had a very different week on the AI front. The company parted ways with Richard Bonneau, its global head of AI for drug discovery. Sources told Endpoints the exit followed tension with research chief Aviv Regev, and John Marioni has taken over on an interim basis.33 Endpoints described this as the latest of several high-profile departures at the Roche unit.33 It adds to earlier concerns that Genentech is moving toward licensing outside assets instead of relying on in-house innovation.22

The same day brought evidence for that view. Genentech paid Alector $100 million upfront for AL050, a preclinical enzyme replacement therapy for Parkinson's disease that uses Alector's blood-brain barrier carrier. The deal includes up to $1.17 billion in milestones.32 The asset has no human data yet.34 For Alector, which rebuilt around this platform after GSK ended their partnership, the payment extends its cash runway into 2029.3139

Shionogi continues its rare disease buildout

Shionogi agreed to buy Austin-based IntraBio for $2 billion upfront.10 The deal gives it Aqneursa, which is approved for neurological symptoms of Niemann-Pick disease type C and received FDA approval on September 18 for ataxia-telangiectasia.6 Closing is expected in November or December.28 The purchase follows Shionogi's $2.5 billion deal for Tanabe's ALS franchise earlier in the year.27 The company is clearly building a U.S. rare disease commercial operation through acquisitions rather than discovery.

What it adds up to

The pattern this week is clear. Regulators and buyers are acting on AI faster than the evidence supports. Utah is approving autonomous prescribing in stages while its first pilot is still under independent review and still in its first phase. AstraZeneca is pouring money into agentic labs while its own scientists say AI's biggest wins so far have come from process work. Genentech, meanwhile, lost its AI leader on the same day it paid for a preclinical asset from outside the company. The near-term returns on AI in biopharma are likely to come from cautiously regulated, limited clinical uses and from data infrastructure. The broader claim that AI will transform discovery remains unproven.