2MM: AI Roundup: Food and Drug Administration reviews liver injury prediction tool, Joint Commission launches healthcare artificial intelligence certification, governance playbooks aim to standardize adoption, and pediatric hospitals bring generative tools to the frontline

NHS England is rolling out Microsoft Copilot to 505,000 clinicians after a 30,000-person trial saved nearly an hour of admin time per day

Recently, NHS England announced it will deploy Microsoft 365 Copilot to 505,000 clinicians and support staff, following a nine-month pilot across 30,000 workers at 90 NHS organizations. The headline result from that pilot was 43 minutes of administrative time saved per staff member per day, which amounts to roughly five weeks per person annually. Full rollout is expected by October 2026. The tool runs inside the Microsoft 365 apps that NHS staff already use, handling things like drafting clinical correspondence, summarizing patient records, generating discharge documentation, and building meeting notes. The rationale for the scale of the deployment is real: a 2026 study published in the National Library of Medicine found NHS resident doctors spend approximately four hours on administrative tasks for every hour of direct patient contact. Whether the 43-minute savings from a motivated pilot cohort will hold across half a million users with varying digital literacy and institutional culture is the empirical question that matters most here. The pilot produced some inconsistency: departments with heavy documentation loads saw large gains, while others saw little. There were also instances of the tool generating clinical text that required correction before it went into a patient record, which is why NHS England is positioning Copilot strictly as an assistive tool requiring human review of all clinical outputs. A Royal College of Physicians survey from earlier this year found 68% of doctors worry AI tools will erode clinical skills over time. The NHS has allocated £45 million for training and has built a network of 5,000 super-users from the pilot to support colleagues during the rollout. An independent evaluation by the Nuffield Trust is scheduled for 2028. For health systems watching from outside the UK, this is the closest thing the field has to a controlled experiment in healthcare AI at national scale. The results will be worth paying attention to.

ARPA-H has picked its teams for ADVOCATE, a program to build the first FDA-authorized AI agent that manages heart disease patients around the clock

ARPA-H selected innovation teams in June 2026 for ADVOCATE, the Agentic AI-Enabled Cardiovascular Care Transformation program, a roughly three-year federally funded competition to develop an autonomous AI agent capable of providing specialty cardiovascular care 24 hours a day. The program has two parallel tracks. The first builds a patient-facing agent that can adjust medications, schedule appointments, give dietary and exercise guidance, and interpret data from wearable devices without requiring a physician to be in the loop for each decision. The second builds a supervisory AI that monitors the clinical agent after deployment, watching for safety drift in a system that learns over time. That second track addresses something the healthcare AI field has largely avoided dealing with: most cleared AI tools are validated at the time of approval and then deployed with minimal structured post-market surveillance. The supervisory agent is designed to do that monitoring continuously. As STAT News pointed out when the program launched, an agent that actually adjusts a patient’s medications carries a categorically different liability profile than a tool that shows a clinician information. ADVOCATE is in development, not in patient care. For cardiologists and primary care physicians managing chronic heart disease, nothing changes at the bedside while the competition runs. What ADVOCATE signals is that the federal government is now actively investing in the regulatory pathway for clinical AI that takes real actions, rather than waiting for industry to define it. The FDA is involved in building the authorization framework alongside the technology. How that framework develops will shape what any future autonomous clinical AI system is allowed to do, in cardiology and well beyond it.

The FDA’s vaccine advisory committee voted 9-0 to recommend Moderna’s mRNA flu shot, putting an August approval on the table

On June 18, 2026, the FDA’s Vaccines and Related Biological Products Advisory Committee voted unanimously to recommend Moderna’s mRNA-based influenza vaccine, mFlusiva, for adults 50 and older. The VRBPAC meeting was the committee’s first review of a new vaccine application since May 2023, and the 9-0 vote came after an unusual regulatory episode earlier this year in which the FDA initially refused to file the application before reversing course after public backlash. The pivotal phase 3 trial data, published in the New England Journal of Medicine, showed a relative vaccine efficacy 26.6% higher than a licensed standard-dose comparator in adults 50 and older over one flu season. FDA staff noted gaps in the evidence: single-season data, limited information in frail older adults and immunocompromised patients, incomplete efficacy data against influenza B strains. The panel found those limitations acceptable given the overall benefit-risk picture. The clinical case for an mRNA flu vaccine comes down to speed. Current flu vaccines require strain selection roughly six months before the season, and the egg-based production process can introduce mutations that reduce match accuracy. mRNA manufacturing could compress that timeline to two to three months, which could meaningfully improve how well the vaccine matches whatever strain is actually circulating. The FDA’s final decision is expected by August 5, 2026. But as BioPharma Dive noted in its coverage, a second hurdle remains: the CDC’s Advisory Committee on Immunization Practices must recommend the vaccine before insurers are required to cover it without cost-sharing. That committee is currently in a legal dispute over its composition following changes made by HHS Secretary Robert F. Kennedy Jr. Whether it will have the legal standing to convene and act on mFlusiva before the 2026 to 2027 flu season is genuinely uncertain.

UpDoc just got the first FDA clearance for a patient-facing large language model, and it raises real questions about what that distinction is worth

UpDoc received FDA clearance for the first Software as a Medical Device that uses a patient-facing large language model. The FDA has cleared more than 1,000 AI-enabled medical devices to date, but nearly all of them are predictive tools operating in physician workflows: imaging algorithms, early warning systems, clinical decision support tools. A patient-facing LLM is different. Its outputs go directly to patients who may act on them without a clinician reviewing them first. The clearance establishes that such a tool can satisfy the FDA’s safety and effectiveness standards when designed with the right guardrails. Initial deployments are announced at four health systems. The practical question this raises for hospital systems is whether FDA clearance creates a real accountability and governance advantage over the uncleared AI chatbots many institutions have already deployed for appointment scheduling, symptom triage, and post-visit follow-up. Those tools exist today without regulatory review, operating under administrative billing codes that do not require clearance. The UpDoc approval arrives the same week ARPA-H advanced the ADVOCATE program, which is working toward FDA authorization for an autonomous cardiovascular care agent. Both point in the same direction: the regulatory boundary around clinical AI is being extended to include patient-facing interactions. Whether CMS reimbursement policy eventually distinguishes cleared from uncleared patient-facing tools, and whether that distinction translates into economic incentives, will determine whether FDA clearance in this category matters beyond press releases.