Electronic Health Records, Administrative Bloat, and the AI Reset
Who is coming to save health systems?
THESIS: Electronic health records turned healthcare into a digital industry, but at the cost of fueling an administrative arms race that now consumes a quarter of U.S. healthcare spending. Health systems are financially upside down in part because of this complexity, and while AI tools like ambient scribes show how fast the industry can adopt solutions when under pressure, the real test is whether AI can collapse—not add to—administrative workstreams. The winners won’t be those who bolt AI on for compliance, but those who redesign workflows so doctors, nurses, and patients spend less time in the bureaucracy and more time in care.
Everything is SaaS. Or at least it felt that way 4 years ago. Even if a startup was building a hardware product, they would say that they had a SaaS platform that would drive their growth or profit, or something else good that investors care about. And even now with LLMs and AI agents on the rise, SaaS still feels like it’s all part of one big acronym soup. Does SaaS mean anything anymore? I guess Software really did eat the world.
Except there are some parts of the technology world that don’t get to use the term “SaaS”. Like marketplaces, which only get the term SaaS applied when they have a specific digital product that is used outside of the marketplace dynamics. Nobody calls Uber a SaaS product. It is a software-enabled service, like most marketplaces these days. And it has some SaaS features. But is Uber SaaS? Not in any popular parlance.
And then there is the healthcare industry. A lot of healthcare technology (referencing software, not biotech or medical devices) actually does have most of the SaaS-like qualities required to earn that moniker. But for whatever reason, Healthcare SaaS is not a term that you frequently see getting thrown around. Instead you see stuff like “Digital Health” “Health Tech” or “Healthcare IT”. In any other industry, companies in those verticals would likely be considered sector-specific SaaS – they fit the SaaS definition pretty solidly.
Part of the reason for this phenomenon is that Healthcare, for a variety of reasons, was one of the last industries to adopt the cloud. Most healthcare digital tools really were not SaaS in a technical sense – the product was not a cloud product. Due to HIPAA concerns, storing patient data in the cloud was seen as too risky. This was a mostly bogus fear, of course, but health systems didn’t want to take that risk, no matter how foolish it was. So Healthcare IT or Health Tech was a more palatable term.
But it’s not just about the cloud and perception, “Healthcare IT” was the term used by policymakers and regulators. Electronic Health Records (EHR) mandates, which happened as a precursor to ACA and as a response to the financial crisis in the late aughts, came about as a result of Health Information Technology for Economic and Clinical Health (HITECH) Act. The term Health IT was literally legislated and voted upon by elected officials. And unlike other industries where technology adoption was catching on like wildfire in the 2000’s, healthcare technology was adopted largely because industry practitioners were forced to through HITECH. The federal government rewarded early adopters of EHRs and even penalized folks who took too long to adopt EHRs into their systems quickly enough. In 2008, 10% of US Hospitals had an HER, but by 2015, 80% had an EHR.
(There are other healthcare technology tools outside of EHRs, but none that challenge the size and scope even remotely. EHRs are the tool through which all healthcare technology flows through today – they are the de facto operating system of healthcare.)
HITECH also set the standard for what an EHR needed to do – the feature roadmap was heavily influenced by the government. Of course, the big players like Epic and Cerner had a seat at the table when these standards were set, but the point is that the market did not play as big of a role as they normally do regarding how these products would be built. But again, the product wasn’t driven to maximize user experience / user ROI. It was developed in a legislative office and might be the ultimate example of the principal-agent problem.
Despite the issues with having politicians determine the feature roadmap for a product, EHRs were still heavily adopted (see above). So EHRs are now ubiquitous and have been for a decade. What are the results?
The workload for doctors and care providers spiked – they now had a government mandated part of their job that was largely focused on data entry. And it was not built for them, the primary users, but rather the federal government. We now have a serious burnout problem in health care systems of the United States and this is a big reason why.
EHRs helped launch digital health as a category. Suddenly, we had real data about care and patients that helped people figure out how to solve problems with a data-first approach. Florence Nightengale used data to revolutionize sanitation in hospitals – so it’s not like this is the first time we had data to drive decision making. But we had a TON of data, in new vectors we probably didn’t fully understand at the time.
Because of the way that the EHR process is used to determine costs and payments and reimbursements in the healthcare industry, they in turn drove the number of administrators through the roof at health systems. EHRs aren’t just about imputing notes into a glorified CRM. They also drive how regulators think about billing and costs for a health system. Health insurance changes and federal mandates made the medical coding process way more complicated, which required a ton more people to try and solve.
You have probably seen a tweet that references that last bullet point:
But the reality is that this has been going on well before HITECH and ACA. This started in the 1970’s. The number of doctors has 2-3x’d since that period, which is reasonable considering advances in medicine and general population growth. The number of administrators has grown by 30x over that same period. There are now roughly ten administrators for every doctor.
Now, I am DEFINITELY not saying that administrators don’t do anything. Even without federally required healthcare digitization, the healthcare industry is extremely complex (and only getting more complex). Healthcare administrators do very important work to solve very complex problems. Insurance has gotten more complicated, more regulations and standards required more oversight, and the conglomeration of the industry leads to larger organizations with more middle managers.
Plus, admin is 25% of the costs of healthcare – not 50% or 100%. There are still a lot of other things that have made the cost structure in healthcare skyrocket. An analysis estimated total waste in U.S. healthcare at roughly $760–935 billion annually, spread across categories: administrative inefficiency ($266 billion), pricing failures ($231–$240 billion), overtreatment/low-value care ($76–$101 billion), fraud and abuse ($59–$84B), and failures in care delivery or coordination (~$130–$244B). So, while admin complexity is one of the largest single waste categories, drug prices and device prices (pricing failure) and unnecessary services (overtreatment) together account for at least as much waste
Even still, you can’t just cut out the administrative costs by cutting out administrators. A lot of these folks don’t just have rote digital workflows that are mindless. There are endless complications with the work they do. On top of that, they are frequently auditors as much as they are doers of tasks. A healthcare company isn’t like a normal company, where if you screw up a task (or a robot screws up an automated task) the result is increased costs or lost revenue. If you screw up a task in a health system, it can result in much worse than lost revenue (for instance, if you screw up someone’s prior authorization, they may not receive necessary care to help alleviate their condition, resulting in an adverse outcome). I have written about this in the past, but a big part of an administrator’s job is to be a pain sponge – protect the organization from liability and mistakes. Of course, mistakes still get made, but you want to minimize them as much as possible.
And it’s not just about mistakes in current care, but the risk of future care mistakes as well. One of the biggest issues within health systems is early readmissions. This occurs when a patient comes back to the health system sooner than they should because the care they received did not adequately solve their problem. Aside from the fact that this fractures patient-doctor trust, it is also a big no-no according to insurers, who frequently cause health systems to eat the bill on the readmit’s cost of care. So you aren’t just auditing for the here & now, but also for the near future.
So healthcare administration is endlessly complicated, and for decent reasons. These are high stakes and a gordian knot of regulations, risks, payment conditions, etc. It’s not like you can just slap an LLM together and start working through the problems overnight. It requires the EXTREMELY rare combination of a deep understanding of the status quo PLUS an ability to want to upset the apple cart (you know what they say about apples and doctors).
The good news is that a lot of companies are trying to solve this administrative problem – health systems are currently upside down financially largely because these costs are running rampant (among other reasons). It is impeding a health system’s ability to generate profit as well as focus on improving revenue optimization. Which might sound icky when thinking about doctors, but those two things allow for there to be more doctors, covering fewer patients per doctor, and achieving better outcomes.
We are seeing health systems adopt new technologies to try and solve this problem. And while I want to give them credit and say they are way ahead of the curve, the reality is that the problem is pretty dire, so they need to act quickly. It’s funny how even a historically slow moving sector is starting to reduce sales cycles when their neck is on the line.
A cross-system taskforce report from the Peterson Health Technology Institute calls healthcare “an industry with notoriously long sales cycles” and then notes there is “no technology in recent memory” that’s been adopted “more enthusiastically” or “scaled so uncharacteristically fast” than ambient AI scribes (i.e., systems are moving from pilots to scale much faster than usual).
Bain & KLAS’ 2024 joint research says providers accelerated tech projects “at a pace previously thought unfeasible” and that gen-AI adoption has “accelerated,” with buyers willing to experiment if ROI is clear—again signaling a shift toward faster decisions.
A 2025 Bain/AWS/Bessemer survey of 400+ healthcare buyers shows providers are furthest along on AI: 30% report systemwide deployment of ambient scribes and another 60% are piloting/implementing—evidence of unusually rapid movement through the funnel.
The current most prominent example is AI notetaking scribes for doctors. There are several large players who are gobbling up market share. If you have been to an annual physical this year, there is a good chance that you have heard the phrase “are you alright if my AI scribe listens in on our appointment?”.
The reality is that these tools are having a big impact and growing pretty rapidly from the outside looking in:
How much capital have these folks raised
Companies like Abridge, Nuance (Microsoft), Suki, and DeepScribe have collectively raised hundreds of millions. Abridge alone has raised over $715M in the last ~18 months, including a $315M Series E in June at a $4.5 billion valuation. Investors are betting heavily that this is the first real at-scale AI application in healthcare.How quickly are they growing
Adoption has been unusually fast for a historically slow-moving industry. Abridge went from pilots to systemwide rollouts in large health systems (UPMC, Emory, Yale New Haven) within a couple of years. Surveys show nearly one-third of health systems already have ambient scribes deployed systemwide, with most of the rest in pilots or implementation.What are some signs that the product is working
Early case studies report physicians saving 1–2 hours per day on documentation, a meaningful reduction in “pajama time.” Providers report higher satisfaction scores, and patient feedback has generally been positive (patients often like that doctors are less focused on typing during visits). Integration with major EHRs like Epic and Cerner is another signal of traction.What are common themes in the product
Most tools are “ambient,” listening passively in the background and generating structured notes that flow directly into the EHR. They emphasize linked evidence (tying notes back to the conversation transcript), human-in-the-loop verification, and seamless Epic integration. The common value prop is less clinician burnout, improved accuracy, and faster documentation.
If these tools can reduce the admin burden for doctors taking notes, there is a good chance they can allow docs to be more effective, see more patients, and focus on what really matters. What’s maybe most notable about these tools is that EPIC has released their own version of the tool this Summer. Who knows what that is going to do to the rest of the market.
But these scribes don’t necessarily impact administrative costs. That is really administrative burden, but specifically for doctors and nurses. While it can down the road start impacting more prevalent admin bloat like billing and claims management, it’s not there yet.
There are other examples of room for AI improvement that might be less scaled to-date
· Prior Authorization Automation: Humata Health (founded 2023 by a former Olive AI executive) raised $25 million in 2024 to build “touchless prior authorization” solutions.
· AI for Medical Coding and Billing: CodaMetrix, an MIT-founded startup, received $55 million (Series A) in early 2023 and an additional $40M Series B in 2024 to scale its AI-based autonomous medical coding platform. CodaMetrix’s algorithms read clinical documentation and assign billing codes (ICD diagnoses, CPT procedure codes) with minimal human intervention, aiming to improve coding accuracy and reduce the need for large human coding teams.
· Scheduling, Triage and Front-Office Automation: Medallion, which isn’t patient-facing but rather addresses credentialing, raised $43M in 2025 to build an “automated credentialing clearinghouse”. Credentialing (enrolling providers with health plans and obtaining licenses in states) is highly administrative; Medallion’s platform automates verifying provider info and submitting it to many payers at once, aiming to save healthcare orgs huge time and labor.
On top of solving the problems listed above, health systems have built in issues of their own. Literally, physical space is a constraint for a lot of health systems, and digital workflows have always suffered from trying to be jammed into the physical world. Healthcare is about the human body, but not all care requires seeing someone in person. Which is obvious now, but almost impossible to fathom in 1990 (or even 2008). As a result – there is a lot of legacy infrastructure that limits our ability to maximize efficiency in the healthcare space.
This is likely why we have seen so many telehealth startups launch (and many take off like rockets) over the last two decades. These are inherently more efficient (even if you simply just think about the driving required to get to a hospital).
General purpose healthcare automation is likely a market that is larger than we even realize or could fathom today. I expect a lot of people don’t even get healthcare services appropriately because of all of the constraints foisted upon the industry today. What if we could strip them away? Could you make going to the doctor feel like a seamless, painless (metaphorical) experience. What if it felt like the doctor could actually focus on you, instead of all of the other bullshit?
The most important point of all, however, is that AI can’t add another layer of administration – that just makes the problem worse. If LLMs are just another software tool that are jammed into health systems for the sake of regulatory compliance, this is all a failure. The lesson here is that we need to have an economic driver behind all of these tools to really make sense. We need LLMs and AI to be a magnitude better than our current processes, otherwise they will just make the problem worse.
So that’s the question rattling around in my head: What are some areas where administration would be ripe for further automation? The winners won’t be those who add AI for compliance, but those who use it to actually collapse administrative workstreams.