The AI Act deadline moved a year. Three obligations did not

The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the AI Act high-risk deadline for AI embedded in regulated products, which includes AI-enabled medical devices, from 2 August 2027 to 2 August 2028. Three obligation sets kept their original dates and apply now: Article 50 transparency duties from 2 August 2026, general-purpose AI provider obligations since August 2025, and the Article 5 prohibitions since February 2025. The deferral also lapses if a system undergoes a significant design change, and no Medical Device Regulation obligation was affected.

This edition covers 27 July to 10 August 2026 in six movements: the regulatory read, what just became possible, voices, money, adoption reality, and the argument. Meanwhile the strongest result published in the last month came from a hospital training a model on its own archive and beating a frontier system by twenty-one points.

The regulatory read

The Digital Omnibus is law, and it is narrower than the headlines. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July. It came into force on the third day after publication rather than the customary twentieth, because 2 August was too close to leave the text hanging.

What moved:

• High-risk AI embedded in regulated products, Annex I. This is where an AI-enabled medical device sits: 2 August 2027 becomes 2 August 2028.
• Stand-alone high-risk systems, Annex III: 2 August 2026 becomes 2 December 2027.
• Pre-existing high-risk systems used by public authorities: not previously set, now 2 August 2030.
• National regulatory sandboxes: 2 August 2026 becomes 2 August 2027.

What did not move, and this is the part being lost in the coverage:

• Article 50 transparency and AI-content marking duties applied from 2 August 2026. They are in force now.
• General-purpose AI provider obligations have applied since August 2025.
• Article 5 prohibited practices have applied since February 2025.
• Two new prohibitions land on 2 December 2026, covering non-consensual intimate imagery and AI-generated child sexual abuse material.
• Everything in the Medical Device Regulation. The AI Act deferral does not touch MDR at all.

Two details in the adopted text deserve more attention than they have had.

The dates are now unconditional. The Commission’s November 2025 draft tied the high-risk deadlines to an assessment that harmonised standards and support tools were ready. The final text removed that trigger, so these are fixed calendar dates that cannot slip again without a fresh legislative procedure. That cuts both ways: no further delay is coming, and nobody has to wait for a standards readiness finding to start.

The transitional relief lapses on significant design change. A high-risk system already on the market before the new application date sits outside the obligations only until it undergoes a significant modification. For a static device that is a genuine reprieve. For an adaptive model, or any product on a normal release cadence, it is a clock that a routine engineering decision can stop.

Article 4 on AI literacy was also rewritten, from a duty to ensure a sufficient level of literacy into a duty to take measures supporting its development. It is now an obligation of effort rather than of result, still binding on every deployer, with national supervision beginning 3 August 2026.

What just became possible

A hospital’s own archive beat a frontier model, and the margin was not close. On 10 July, a University of Michigan team published NeuroVFM in Nature Medicine: a three-dimensional visual foundation model trained on 5.24 million clinical MRI and CT volumes drawn from 566,915 studies acquired over twenty years at Michigan Medicine.

The premise is the interesting part. Frontier models trained on internet-scale public data are structurally blind to neuroimaging, because the identifiable facial features embedded in head MRI and CT keep those scans out of public datasets. The paper shows frontier models underperforming on neuroimaging tasks as a direct consequence, then shows that training on uncurated data generated during routine clinical care closes the gap. The authors call the paradigm health system learning.

NeuroVFM was trained with a self-supervised, vision-only method, without hand labelling, without curated datasets, and without radiology report supervision. In a prospective feasibility study run across the whole health system for one week in January 2026, it outperformed GPT-5 on critical-findings triage by 21.4 percentage points, at 92.6% against 71.2%, with roughly half the hallucination rate.

The benchmark gap nobody has closed. A related result is worth holding alongside it. Large language models routinely score around 92% on standardised medical licensing examinations. The same models evaluated against real-world clinical tasks on the BRIDGE benchmark, developed at Mass General Brigham and published in Nature Biomedical Engineering, scored 44.8%.

Separately, a head-to-head benchmarking study published in Nature Medicine on 23 June 2026 compared general-purpose models against two cleared clinical AI tools on questions submitted by practising physicians. The general-purpose models outperformed both cleared tools across every benchmark tested.

Voices

Three published positions inside four days, arguing with each other. This is the most useful disagreement in the field right now and it is worth reading in order.

Aviv Goldenberg and Jenna Wiens, in Nature Medicine, asked the question directly: “Is AI actually improving healthcare?” Their answer is a qualified yes with a substantial caveat, that in many cases we do not know, because a large share of deployed tools are too new for anyone to say whether they improve patient outcomes. Nat Med 32, 1182–1183 (2026). DOI 10.1038/s41591-026-04329-2

Priya Abani, chief executive of AliveCor, replied in STAT on 6 August under the headline “Stop asking if AI is good for medicine.” Her argument is that the question is malformed, because healthcare treats AI as a monolith when it is not. Her analogy: asking whether AI improves healthcare is like asking whether lasers improve surgery. In the hands of a skilled surgeon using a validated tool, they allow lifesaving precision. In an unproven setting the question is still open. Her concern is that overgeneralisation slows the adoption of tools that demonstrably work. Declared interest: AliveCor sells AI-powered cardiology products. She is arguing a position her company benefits from, which does not make it wrong and should be stated.

Frances Mei Hardin, a physician, took the opposite side in STAT the following day: “AI won’t enhance physician autonomy. It will further diminish it.” That is a direct rebuttal of the most common argument for clinical AI, which is that automating documentation returns time and agency to clinicians.

Money

Two rounds dominated, both consumer-facing, both European in origin, and both worth reading for what the paperwork says rather than for the number.

Neko Health, $700m Series C, 15 July. The Swedish full-body scanning company raised $700 million, co-led by Lightspeed Venture Partners and O.G. Venture Partners, with Atomico, General Catalyst and Lakestar returning and Liberty City Ventures, BDT & MSD and Positive Sum joining. Reported at a valuation near $7 billion, against $1.8 billion at the $260 million Series B in January 2025. Disclosed funding since 2023 now exceeds $1 billion. The money opens the first US clinic, in New York. Neko reports more than 100,000 scans delivered across the UK and Sweden against a waitlist of 350,000.

The detail that matters is in the FDA record. Two Neko devices were cleared through the 510(k) pathway in May 2026: Derma-2 as an adjunctive telethermographic system, and Spectrum-2 as a tissue-saturation oximeter for cardiovascular measurement. Those clearances cover the two named devices for their stated intended uses. They do not clear the Neko Health Scan as a screening service, and Neko describes its US clinics as preventive health and wellness providers rather than full-service medical practices, advising customers to keep seeing their own clinicians for diagnosis and treatment.

Bunkerhill Health, Series B, 16 July. Khosla Ventures led, with Sequoia Capital, Felicis, Optum Ventures and Y Combinator continuing. The platform, Carebricks, lets a health system build and deploy its own AI agents rather than buy a fixed product, and is running at Cleveland Clinic, Intermountain Health and the University of Texas Medical Branch, where more than twenty agents are live across clinical, operational and administrative work.

Note the figure carefully, because most of the coverage has it wrong. The company’s own announcement states that the Series B brings total funding to date, including its Seed and Series A rounds, to $55 million. At least four outlets reported a “$55M Series B”. The round size has not been disclosed.

Adoption reality

The largest number of the month, and what it actually says. Wolters Kluwer’s 2026 half-year report, published 5 August, states that as of 31 July, more than 90% of its US Enterprise Edition customers, representing approximately 2,500 hospitals, have signed up to adopt the UpToDate Expert AI version, against a mid-year goal of 70%. Internationally, more than 230 sites are activated across 36 countries. Drug dosing guidance was added after validation by clinical experts and pharmacists.

Three days later the company’s own newsroom carried the same fact under a different verb: adoption has expanded to approximately 2,500 hospitals and health systems.

Signed up to adopt is a procurement fact. Adopted is a deployment fact. Used is a clinical fact. They are three different numbers and only the first one has been published.

The argument

Three things happened in the last month and they point the same way.

A regulator deferred the obligations that are hardest to meet and kept the ones that are easiest to breach. A hospital demonstrated that its own unglamorous archive outperforms the frontier at the task that matters clinically. And the field’s most credible voices published, within four days of each other, three incompatible answers to whether any of this is working.

The common thread is that the proxies have stopped tracking the thing. A CE mark does not tell a buyer the tool performs. An exam score does not tell a developer the model is clinically useful. A deferred deadline does not tell a compliance lead they are safe. And a 2,500-hospital deployment figure does not tell anyone whether a single patient was better off.

For a company selling clinical AI in Europe over the next eighteen months, the practical consequence is that the evidence that will sell the product is not the evidence that gets it approved. Both have to be built, they are built from different work, and only one of them has a deadline attached.

Vital Signs: AI Healthcare is produced by HealthSeed AG, Weidmannstrasse 5, 8046 Zurich, CH-020.3.056.106-6. HealthSeed is a Swiss healthcare commercialisation firm. We take medtech, diagnostics, and digital health companies from an approved product to a paying European customer, running regulatory, reimbursement, and commercial as one sequence rather than as three handovers. Vital Signs is written out of that work, which is why it reads for the decision rather than for the announcement. Every claim here carries a source you can open. Where a figure comes from secondary coverage rather than the paper itself, the source line says so.