I've been rebuilding my clinical documentation tools over the last few weeks, and the work keeps pulling me toward the same question: why is the data inside a PT note so hard to use? I can write a note that tells a clear clinical story. Another therapist can read it and understand what happened. But no system I've ever worked with can answer a simple question like "show me this patient's knee flexion across 12 visits" without me opening 12 notes and looking with my own eyes.
That's not a technology problem. It's a format problem. And I think the format is going to have to change.
Clinical notes in outpatient PT are trying to do two things at once. They're a clinical narrative for the next person who reads the chart, and they're a defensible document that justifies billing. Those goals pull in different directions — the clinical narrative wants to be concise and specific, while the compliance narrative wants to be thorough and protective. What you end up with is a hybrid that clinicians resent writing and that nobody reads unless they have to.
The bigger issue is what happens to the data after the note is signed. It sits there. Every ROM measurement, every pain score, every functional milestone — it all enters the EMR as prose and essentially dies. You can read it. You can search for keywords if your EMR is feeling generous. But you can't query it, trend it, or feed it to any system that could do something useful with it. The clinical data is trapped in sentences.
The idea I keep coming back to is a dual-layer document. On top, the clinician writes and reads clean prose — the same narrative they'd produce today. Underneath, a structured data layer captures every clinically meaningful detail in a format machines can parse. Range of motion values anchored to the sentence they came from. Milestones tagged when they're achieved. Interventions logged with progression history. The clinician sees a note. The machine sees a dataset. They're looking at the same document.
The key is that the structured layer isn't something the clinician builds separately. It's generated from the same input as the narrative — either extracted by the model during generation or captured through a structured input that produces both layers simultaneously. The clinician's workflow doesn't change. They document the encounter the way they always have. The structure is a byproduct of good documentation, not a tax on top of it.
Once clinical data lives in a structured format that's anchored to the source narrative, the downstream possibilities start to stack. An agent could track a patient's trajectory across visits without anyone building a spreadsheet. Morning briefs could flag patients whose progress has stalled for three visits in a row — the kind of thing that currently gets lost in a caseload of 40 people.
The one that keeps me up at night is prior authorization. Right now, writing a prior auth request means opening a stack of notes, manually pulling out the data points that prove medical necessity, and drafting a letter that you hope the reviewer will actually read. With a structured annotation layer, that letter could assemble itself — cited, objective, pulling directly from the data. The payer's system could read the same structured format. Flexion went from 95 to 125 over 12 visits. Three functional milestones achieved. Progress either shows or it doesn't. There's nothing to argue about.
I'm not sure what this looks like at scale. The schema design alone — figuring out the right structured representation for ROM, strength, pain, functional status, interventions, and all the edge cases that come with real clinical documentation — is a significant problem. Getting it right for five data types might be manageable. Getting it right for the long tail of clinical scenarios is something else entirely.
There's also the question of who builds it. The EMR vendors have no incentive to make their data more portable or more useful outside their walls. A shared annotation schema between clinician systems and payer systems would be transformative, but it requires a level of cooperation that this industry hasn't historically been great at. I'm not naive about that.
What I do know is that the current model — narrative notes that no machine can meaningfully read — is not going to survive a world where AI agents are embedded in clinical workflows. The data has to be accessible. It has to be structured. And ideally, the clinician shouldn't have to do anything different to make that happen. One input, many outputs. The note stops being a document you write and forget, and starts being the infrastructure that everything else runs on.