A Month of Healthcare-AI Headlines, Read From the Floor

If you only read the headlines about healthcare AI from the past few weeks, you would be confused. AI is going to save medicine. AI is going to make your insurance more expensive. Nobody knows who regulates it. Even the best models still make a meaningful share of dangerous mistakes.

All of that happened in June. None of it contradicts. From the floor, it points at one uncomfortable question: when AI gets it wrong, or drives costs up, who absorbs it? Right now the answer is still the clinical staff.

Insurers are now blaming AI for rising costs

Start with the money, because it is the most concrete. Healthcare Dive reported on June 11 that health plans expect commercial healthcare costs to rise about 9% next year. The figure comes from a PwC report based on a survey of actuaries at 27 US health plans covering more than 103 million employer-sponsored members.

What is new is that AI made the list. About 70% of surveyed plans put provider-side AI documentation and coding tools among the top three trends pushing commercial costs up, and about 20% called it the single biggest factor. This is not science fiction. It is billing. PwC's Glenn Hunzinger told Healthcare Dive the trend is providers using technology and AI to code more appropriately, or to code things that could not be coded before. He was careful to add that AI is not the biggest driver overall, since labor and supply costs still lead, but that it does move that 9%.

Sit with that. The first place AI has made a measurable difference in healthcare is not diagnosis. It is the size of the bill.

And nobody fully agrees on who regulates it

On June 12, Becker's Hospital Review summarized a Bipartisan Policy Center report on the gaps in health AI oversight. The center's Maya Sandalow put it simply: most of the AI actually deployed in healthcare right now is administrative, not clinical. Ambient scribes that listen to a visit and write the note. Tools insurers use to decide prior authorization. Tools that affect patients indirectly all day, and that mostly sit outside medical regulation.

Why the gap? Because authority is scattered. The same AI can be regulated differently depending on where it is used and what data it was trained on, the report notes, listing the FDA, CMS, the HHS Office for Civil Rights, the FTC, and ONC, on top of a patchwork of state rules. That fragmentation creates an incentive the report names directly: developers have reason to label products as not a medical device to stay out of FDA review. In the same week, the administration proposed rescinding ONC transparency requirements for AI built into electronic health records, which would make it harder for the people using those tools to understand how they work.

Meanwhile the building is speeding up

While regulators sort out who is responsible, companies are not waiting. On June 11, AI documentation company Abridge announced it is working with Nvidia on a foundation model for clinical conversation and has taken a strategic investment from Eli Lilly. Abridge's documentation system is already in use at more than 300 health systems including Kaiser Permanente, Johns Hopkins, and Yale New Haven. CEO Shiv Rao framed the Lilly partnership as supporting evidence-based care and improving access to research at every point of care.

This part is both encouraging and slightly unsettling. A model trained specifically on how clinicians and patients actually talk will likely outperform a general chatbot in a white coat. It also means more clinical conversation is being captured, modeled, and commercialized faster than the rules covering it are being written.

How good is the AI, really?

Here is a number worth getting right, because it gets misquoted often. Stanford and Harvard researchers published a preprint in January 2026 called First, Do NOHARM, testing 28 large language models against a benchmark of 1,100 real cases across 10 specialties, drawn from primary care referrals to specialists. Across the models, up to 22.6% of cases produced recommendations that could cause severe, life-threatening harm. Worth noting on the other side: the best models scored safer than generalist physicians, and combining several models in a multi-agent setup reduced harm further. The risk is not evenly distributed.

The detail that matters most for nursing: more than 80% of the severe errors were omissions. The model did not raise a dangerous flag. It left something out, like failing to recommend a follow-up that should have happened. (This is a preprint and has not been peer reviewed, so treat the specific figures as preliminary.)

Omission errors are exactly the kind an exhausted, overloaded nurse is least likely to catch, because catching one means noticing what is missing rather than noticing what is wrong.

What nurses actually said about it

The people doing the work have a view, and it is not anti-AI. A Black Book Research report released during Nurses Week in May surveyed 118 nurse leaders. Most were open to AI, with one significant condition: it has to be designed for nursing. 74% said the physician-style ambient documentation tools getting all the attention do not solve nurses' documentation burden unless they are rebuilt around how nurses actually chart. 77% wanted AI to start with low-risk, high-volume tasks before it goes anywhere near clinical judgment.

That is not resistance. That is people who have been handed the wrong tool before and want a say this time.

Where we land

Eipnare makes scrubs, not software. We are not going to lecture anyone about machine learning. But our job is to pay attention to what makes a shift harder, and that gets lost easily under the hype.

• Every one of these stories pushes the cost of imperfect AI onto the clinician. An error rate like that does not reduce a nurse's workload. It adds a checking task to a load that is already full.
• The nurses in that survey have it right: give AI the boring, low-stakes work first, like charting, not judgment. Earn trust where being wrong is cheap.
• What we contribute to a better shift is unglamorous, and we are fine with that. On a day that now includes double-checking a model nobody fully regulates, what you are wearing should be one less thing to think about.

The technology is moving faster than the rules and faster than staffing. We are watching who ends up holding that gap. For now it looks like the floor.

Key takeaways

• Insurers expect commercial healthcare costs to rise about 9% next year, and a PwC survey names provider AI coding and documentation tools as one of the main drivers.
• Health AI oversight is split across multiple agencies, and most AI currently deployed is administrative and largely outside medical regulation.
• An early Stanford-Harvard preprint found leading clinical models still make severe errors in a meaningful share of cases, most of them omissions.
• Surveyed nurse leaders want AI built for nursing and started on low-risk tasks first. In each of these stories, the cost of imperfect AI lands on clinicians.

Frequently asked questions

Why are insurers blaming AI for rising costs?

Health plans expect commercial costs to rise about 9% next year, per a PwC report cited by Healthcare Dive, and about 70% of surveyed plans ranked provider AI documentation and coding tools among the top three trends driving costs up. The mechanism is billing: AI helps providers code more, and more completely.

Who regulates AI in healthcare?

Oversight is split across the FDA, CMS, the HHS Office for Civil Rights, the FTC, and ONC, plus state rules, according to a Bipartisan Policy Center report. Most deployed AI is administrative rather than clinical and sits largely outside medical regulation.

How accurate is clinical AI?

A Stanford and Harvard preprint from early 2026, First, Do NOHARM, found that even leading models gave severely harmful recommendations in a meaningful share of cases, most of them omission errors where the model failed to recommend a necessary follow-up. It is a preprint, so treat the specific percentages as preliminary.

What do nurses want from healthcare AI?

In a Black Book Research survey of 118 nurse leaders, 74% said physician-style ambient documentation tools do not solve nurses' charting burden unless redesigned around how nurses work, and 77% wanted AI to start with low-risk, high-volume tasks before clinical judgment.

Sources

• Healthcare Dive on the PwC medical cost trend report (June 11, 2026)
• Becker's Hospital Review on the Bipartisan Policy Center report (June 12, 2026)
• Healthcare Dive and Fortune on Abridge, Nvidia, and Eli Lilly (June 11, 2026)
• Stanford and Harvard preprint, First, Do NOHARM (January 2026)
• Black Book Research nursing AI report (May 2026)

Related reading

• Two Nurse Unions, the Same Two Asks: Staffing Ratios and AI Limits in the Contract
• The Joint Commission Now Has a Rulebook for Hospital AI
• Technology Can Cut Nurses' Paperwork. It Can't Replace the Basics.
• What Nurses Actually Need From Scrubs in 2026
• Shop best-selling scrub sets

Edited by Hedy Nie, COO of Eipnare. Connect with her on LinkedIn.