I started Cloudticity in 2011 because three things arrived at once. The federal government had put real money behind electronic health records through a program called Meaningful Use, which was heading into its second stage with penalties scheduled to follow the incentives. Medical devices and digital records were producing more health data than most organizations knew how to store. The cloud had become a practical, affordable place to keep and process it. I thought that combination would let a technology company help people in a way that hadn’t been possible before.
Meaningful Use is easy to criticize. I think it mostly worked. In 2008, fewer than one in 10 hospitals had an electronic health record. By 2024, more than 99% of non-federal acute care hospitals and 91% of office-based physicians had adopted a certified one, according to federal data. The Congressional Budget Office estimated the incentive programs would cost $30 billion through 2019. For that money the country got a digital record of care. Everything since has been built on it.
The money explains only part of the result. The program told the industry what to adopt, what counted as using it, and when. The first stage gave hospitals 14 core functions to demonstrate, and Medicare payment adjustments began in 2015 for those that hadn’t. A hospital executive in 2012 could say exactly what the organization was working toward and how it would be judged.
The strongest case against the program is about who paid for it. In a 2016 study published in Annals of Internal Medicine, researchers observed 57 physicians in four specialties and found that for every hour of direct face time with patients, they spent nearly two more on the electronic record and desk work during the clinic day. That is a small study of office practice, and it can’t separate what Meaningful Use caused from what billing and documentation rules already required. The complaint behind it is still fair. Some of the measures were blunt, and a physician who spent evenings finishing charts had reason to think the program counted clicks while the work of care went unmeasured.
I still think the program mostly worked, for one reason. Its job was to get the record into digital form. A badly designed screen can be redesigned, and much of today’s investment in ambient documentation is an attempt to do exactly that. A paper chart can’t be searched, exchanged, or learned from at all. The cost was real and it fell on clinicians. I think the country was right to accept it, on the condition that the industry then spent years making the record easier to live with. That work is still under way.
AI has the adoption and none of the definition
AI is arriving in healthcare with comparable force. A survey of executives at 120 health systems published in March found that 75% were using at least one AI application, up from 59% a year earlier. Ambient note-taking alone had reached 68%.
Nobody has told those health systems what meaningful use of AI is. No program defines which outcomes count, what evidence shows them, or what happens when a tool falls short.
The federal government has now made clear it doesn’t intend to. On October 26, HHS finalized a rule known as HTI-5 that eliminates 14 health IT certification criteria and revises seven others. Among the requirements removed is the one that made developers of certified health IT publish model cards, the documents describing how an AI decision-support tool was built and evaluated. Much of the commentary since has said the rule shifts oversight to health systems.
I read it differently. The accountability was always the hospital’s. A model card describes how a tool performed somewhere else, on someone else’s patients. It could never tell a health system whether the tool works on its own population, inside its own workflow, with its own data. The rule changed the paperwork. The responsibility stayed where it was.
Patients already see it that way. In a West Health-Gallup survey of more than 5,000 adults released this week, 56% said they were comfortable with AI making diagnosis or treatment decisions when their provider could change or override them. When AI made those decisions without provider review, the figure fell to 9%. People will accept the tool if a person they trust answers for it.
What your own definition should contain
If no one is going to write Meaningful Use for AI, each health system should write its own. It can borrow the structure that made the original work, and it can fit on a page:
- The outcome. State it in terms the board already tracks, before a tool is chosen: time to a diagnosis, denied claims overturned, hours clinicians spend on notes.
- The measure and the baseline. Record where that number stands today. A pilot without a baseline can report enthusiasm and little else.
- The stopping rule. Name the evidence that would end the deployment, and the date it will be reviewed.
- The owner. One person who can pause the tool, with the authority written down.
- The data. What the tool reads, where that data goes, and who else holds a copy.
Take ambient note-taking, the AI tool most health systems have already bought. Here is what the page might say for a hypothetical primary care group.
The outcome is time: clinicians finish their notes during the clinic day. The measure is the share of notes closed before the clinician leaves and the hours of charting done after hours, both taken from the record system for the month before go-live. The stopping rule has two parts. If after-hours charting hasn’t fallen by an agreed amount in 90 days, the pilot ends. If the weekly review of a sample of notes turns up errors that would change care, the tool pauses the same day. The owner is the group’s chief medical information officer, who can switch the tool off without convening a committee. The data line says where the recordings go, how long the vendor keeps them, and whether they’re used to train anything.
None of that is sophisticated, and it fits in a paragraph. It changes the 90-day review from a conversation about whether people like the tool to a decision about whether it did what the organization bought it to do.
The data line deserves more attention than it gets. Earlier this year a phishing attack on a single healthcare AI vendor reached nine health systems and about 1.4 million patients. An AI tool that needs patient data becomes another place that data lives.
I served on the HITRUST AI Working Group in 2024. The lesson I took from that work is simple to state: safety, security, and compliance are far cheaper to design in than to add after a pilot has succeeded and everyone wants it in production next month.
The objection, and the dependency
The reasonable objection is speed. Competitors are deploying, pilots teach things a planning document can’t, and a health system that waits for a perfect definition will wait a long time. I agree with all of it. The page I’m describing takes an afternoon. It exists so the pilot can teach something, because a team that hasn’t written down what it expected can’t tell whether it was wrong.
There is also a dependency underneath the definition. Meaningful Use produced records first and exchange later for a plain reason: you can’t share a record that doesn’t exist in digital form. By 2023, 70% of hospitals sent, found, received, and integrated outside records at least some of the time, more than a decade after the incentives began.
AI has the same dependency one layer down. An AI system inherits the environment it runs in. If nobody can say which data sits where, who can reach it, and which controls apply, the AI system has the same problem at higher speed.
During the pandemic, Cloudticity built New York State’s first COVID-19 data hub in 11 days. That speed came from two things that were in place before anyone wrote code: a clear answer to what the system was for, and a compliant foundation to build it on.
Before your organization approves its next AI purchase, ask to see the page. If the outcome, the baseline, the stopping rule, the owner, and the data can’t be written down, the purchase isn’t ready, however good the demonstration was.