Why you shouldn't use AI to fix things in healthcare

There is a version of AI in healthcare that I find genuinely exciting, and a version that makes me wince a little every time I see it. The exciting version changes what is possible. The wincing version changes what is tolerable. And the problem is that most of the investment, most of the press releases, and most of the LinkedIn thought leadership is flowing toward the second version while the first sits largely unexplored.

Let me explain what I mean, because this distinction matters enormously for anyone building, funding, or deploying AI in health systems right now.

The version that makes me wince is this: using a large language model to build a better electronic medical record. I have seen this pitch dozens of times. The AI makes documentation faster. The AI surfaces relevant information more intuitively. The AI reduces the number of clicks between a clinician and the data they need. The AI, in short, makes the EMR better. And every time I see it, I have the same thought. Why are we trying to make the cage more comfortable when we could be unlocking the door?

The EMR Was Never the Solution. It Was a Workaround.

To understand why the "better EMR" framing bothers me, you have to understand why EMRs exist in the first place. They exist because structured data, data that fits neatly into fields, rows, and relational tables, was the only way computers of the 1990s and early 2000s could store, query, and exchange clinical information reliably. The structured record was not a natural representation of clinical reality. Clinical reality is deeply unstructured. A physician thinks in narratives, in patterns, in exceptions and edge cases, in the particular tone of a patient who says they are fine but whose body language says otherwise. None of that fits in a dropdown menu.

So we built the EMR as a translation layer. We took the rich, unstructured world of clinical care and forced it through a narrow pipe of structured fields, billing codes, and standardised terminology, because that was the only way to make it computable. The EMR was not a feature of good medicine. It was a workaround for the limitations of the technology available at the time. And we have been iterating on that workaround for thirty years.

Large language models change the fundamental premise. An LLM does not need the data to be structured at input. It can receive a physician's spoken notes, a patient's free-text history, a discharge summary written in clinical shorthand, an image interpretation, a pathology narrative, and it can understand the clinical meaning across all of it. Then, on output, it can produce whatever structured format is required. ICD-10 codes. CPT codes. DRG groupings. FHIR resources. HL7 messages. The downstream system gets exactly the structured data it needs, in exactly the format it expects, without the clinician ever having to perform the translation manually.

ICD, CPT, DRG, FHIR, HL7; and yes, if a payer tomorrow decided they needed data in some other format, the model would accommodate that too. The format is no longer the constraint. The format is just a parameter.

This Is Not Incremental. This Is Categorical.

I want to be precise about why this is a different kind of advance, because the healthcare industry has a long history of describing incremental changes in categorical language, and I want to do the opposite here.

The difference between "AI makes the EMR better" and "AI eliminates the need for the EMR as a primary data capture mechanism" is not a difference of degree. It is a difference of kind. It is the difference between building a faster horse and building a car. It is the difference between making the fax machine more reliable and building email. It is the difference between improving paper chart organisation and creating the concept of a searchable digital record.

When the primary constraint changes at the level of physics, the solutions that were designed around the old constraint do not need to be improved. They need to be reconsidered from first principles. And that is exactly what is happening right now with clinical documentation, except that most of the industry has not yet internalised what it means.

The clinician should be able to have a conversation with a patient. A full, present, eye-contact conversation. The AI listens. It understands the clinical meaning. It generates the note, the codes, the referral letter, the structured summary for the receiving specialist, and the payer authorisation request, all simultaneously, all correctly, all without the clinician ever touching a keyboard or clicking a single field. The structured outputs are a downstream artefact. The clinical encounter is the primary event.

That is not a better EMR. That is the end of the EMR as we know it.

The Quiet Confirmation Nobody Is Talking Loudly About

I try not to use other companies' strategic decisions as validation for my own thinking, because that is a lazy form of argument. But there is a data point worth noting, and I will mention it subtly enough that the people who have been watching the healthcare IT space will recognise it immediately.

The largest database company on the planet, which acquired one of the two dominant EHR systems in the world for nearly thirty billion dollars, did not spend the next three years trying to make that EHR better. It spent those years quietly building a new one from the ground up, AI-native, cloud-native, with intelligence embedded in every layer rather than bolted on top. The old architecture was not upgraded. It was scheduled for replacement. That is not a company that believes the future is a better version of the past. That is a company that has done the same analysis I am doing here and arrived at the same conclusion.

When a company of that scale, with that level of existing investment in a legacy architecture, decides that the right move is to start over rather than iterate, it is worth asking what they know that the rest of the market is still catching up to.

The Pattern Is Broader Than Healthcare Records

The EMR is the most viscerally obvious example of this pattern in healthcare, but it is not the only one. The same principle applies across the entire landscape of health IT, and I think it is worth naming a few of them explicitly, because the investment and innovation community tends to be sector-specific in its thinking, and the broader pattern is more instructive than any single example.

Prior authorisation, the process by which a physician requests permission from a payer to deliver care they have already determined is necessary, exists in its current form because humans needed to exchange structured information about diagnoses, procedures, and coverage criteria through systems that could not interpret unstructured requests. An AI system that understands clinical meaning can submit and respond to prior authorisation requests in natural language, with the structured fields generated automatically on both sides. The prior authorisation process, as currently constructed, dissolves. What remains is a policy question, not a technology problem.

Clinical trial matching, a process that currently requires specialist coordinators to manually review hundreds of eligibility criteria against patient records, exists because structured databases cannot reliably interpret the clinical nuance of both the trial protocol and the patient's history. An LLM that reads the protocol and the chart together can match patients to trials at a scale and accuracy no human team can match. The matching coordinator role, as currently constructed, dissolves. What remains is a governance question about who reviews the AI's recommendations.

Discharge summaries, referral letters, insurance claim narratives, clinical audit documentation: each of these is a structured translation of an unstructured clinical event, performed by a human, at a cost measured in time that could have been spent on care. Each of them dissolves in a world where the clinical event is captured once, in its natural form, and the downstream documents are generated as outputs.

Why the Industry Keeps Choosing the Cage

If this analysis is correct, and I believe it is, then the obvious question is why so much of the investment in healthcare AI is still flowing toward "better EMR" projects rather than "post-EMR" ones. The answer, I think, has several components, and understanding them is important for anyone trying to navigate this space strategically.

The first is procurement inertia. Health systems make decade-long commitments to their core clinical systems. The sales cycle for an enterprise EHR is measured in years, and the switching cost is measured in hundreds of millions of dollars for large networks. The people making procurement decisions have every incentive to improve what they have already committed to, and every institutional reason to avoid conclusions that would require them to acknowledge that the commitment was made to a technology paradigm that is now being superseded.

The second is regulatory familiarity. The structured data formats that EMRs produce (ICD codes, CPT codes, FHIR resources) are baked into regulatory requirements, payer contracts, and accreditation standards. Regulators know what a structured record looks like. They know how to audit it. An AI-native documentation system that produces the same structured outputs from unstructured inputs is technically compliant, but it challenges the mental model of what "compliant documentation" means, and regulators are not typically early adopters of mental model changes.

The third, and most important, is that solving a problem is a much easier story to tell than eliminating it. A pitch that says "our AI reduces documentation time by forty percent" is comprehensible to every CIO in the room. A pitch that says "our AI makes structured documentation an automatically generated output rather than a manually performed input, which means your clinicians never need to interact with the documentation layer again" requires the listener to reconstruct their entire mental model of how clinical information flows. It is a harder sale. It is also a much better outcome.

What Problem Elimination Actually Looks Like in Practice

I want to be concrete about what this means in practice, because it is easy for this kind of argument to stay at the level of abstraction and become another form of the healthcare AI hype it is trying to critique.

Problem elimination in clinical documentation means a physician walks into an examination room. They have a conversation with their patient. The AI, listening with clinical understanding, captures the encounter in real time. At the end of the appointment, the physician reviews a generated note, makes any corrections needed, and approves it. Every code, every billing entry, every structured data field required by the downstream systems (the payer, the referral specialist, the population health database, the quality registry) is populated automatically from the same captured encounter. The physician's interface with the information layer is a review function, not a data entry function.

The difference between this and "AI that makes the EMR faster" is that in this model, the EMR as a data entry system has been eliminated. The physician still has access to the patient's record. The structured data still exists. The downstream systems still get what they need. But the human cost of generating that structured data (the two hours of after-hours documentation per day that drives physician burnout, the cognitive interruption of switching between clinical attention and data entry mid-appointment, the information loss that occurs when a clinician abbreviates a nuanced clinical observation to fit a structured field) is all gone. Not reduced. Gone.

That is what I mean by problem elimination. The problem was never "physicians don't have good enough tools for data entry." The problem was "physicians shouldn't be doing data entry at all." We built tools to help with the former because we could not solve the latter. Now we can solve the latter.

Where I Am Spending My Attention

I run EthicaLabs, an AI governance advisory, and I am the Director of Partnerships at Partners in Digital Health, publisher of Telehealth & Medicine Today. I also built CIGMA, an AI wellness platform for vulnerable populations. Across all three of these roles, the work that interests me most right now is not the work of optimising legacy architectures. It is the work of identifying, with precision, which problems in healthcare are genuine problems and which are workarounds for constraints that no longer exist.

The distinction matters because the governance question looks different depending on which type of problem you are addressing. Governance for a better EMR is governance for a more efficient version of an existing process. Governance for an AI-native clinical documentation system is governance for a fundamentally new way of capturing and structuring clinical information, with new questions about accuracy, auditability, accountability, and the appropriate role of human review.

Those are harder governance questions. They are also more important ones. And they are the questions I am increasingly focused on, both through EthicaLabs and through the peer-reviewed work we support at Telehealth & Medicine Today.

The Invitation

If you are building in health AI, I want to invite you to ask a different question before you start. Not "how can AI make this process better?" but "does this process need to exist at all in a world where AI can do what it can now do?"

The answer will sometimes be yes. Some processes exist for reasons that are independent of technological constraints: regulatory, ethical, relational. A physician should review and approve a clinical note. A human should be accountable for a treatment decision. There are elements of the clinical encounter that should not be automated, and good governance requires knowing which ones they are.

But many healthcare processes exist for no better reason than that the technology of the 1990s required them. Those processes are not sacred. They are not best practice. They are technical debt, dressed up in workflow familiarity. And the opportunity in front of the healthcare AI field right now is not to make that debt more manageable. It is to pay it off.

That is where the genuinely interesting work is. I am doing it. If you are too, I would like to compare notes.