The Transcript That Keeps Pace

A dictated note used to be a thing you waited for. You spoke into a recorder, the recording went somewhere, and later a transcript came back. Every step in that chain was a place where the audio sat still: on a device, in a queue, on a server, waiting for the next step to pick it up. For a clinical note, a recording sitting still is the worst state it can be in. It is the most sensitive form of the information and the least useful one.

This week we finished the third of three pieces that, together, stop the recording from ever sitting still. I want to describe the three as one motion, because that is how it feels to use, and because each piece only makes sense in light of the others.

The transcript keeps pace with the speaker

The capture is now chunked. The recorder sends the audio in pieces, cut at the pauses, each piece transcribed inside its own upload. The clinician sees the transcript filling in, about a minute behind their own voice, while they are still speaking; the final piece closes the session and the stitched note is the landing. There is no moment where a whole recording is waiting for a whole transcription. A long dictation is not a bigger risk than a short one, because the system never holds more than the piece in front of it.

The practical effect is that the note is finished a few seconds after the clinician is. They can read what they said, correct it, or tap stop and walk away.

The note proposes the work

A clinical note is full of instructions aimed at the future: book a follow-up, order a bitewing, call the patient about the result. In the old world those instructions were read out of the note by a person at the desk, later, by hand. Now the note proposes them — a short list of actions it found in the clinician's own words, each one a verbatim quote so the source is visible — and a person confirms each one before anything happens.

I want to be precise about that last clause, because it is the part I would be most skeptical of if I were reading this about someone else's product. The note does not act. It proposes, and each proposal is executed only when a person approves it, one at a time. A proposal that nobody approves does nothing. I wrote in The Hard Line about where autonomy stops; this is that line, drawn through a dictation.

The audio is destroyed when the record is filed

The default is that the audio lives only until the note is filed or discarded. Filing destroys every piece, and the audit trail carries a card that names how many were destroyed. A clinic that wants to keep audio for a defined number of days can elect that, per clinic, and the election is stamped on every dictation it governs. But the default is the one I would choose as a custodian: once the record exists, the recording has done its job, and the safest recording is the one that no longer exists.

Our nightly check runs this whole motion on a synthetic clinician dictation: two pieces in, a stitched note out, an action proposed and approved through the same gate a person would use, the record filed, and then the storage location listed to confirm it is empty. It has run every night since the end of August. On the first night with an action in it, one of our two synthetic clinics rejected the action and the other did not. A rule meant to keep phone numbers out of a task was counting the wrong thing, and the night found it before a clinician could.

Who finishes the motion

The piece that is still ahead of us is the one that matters most for a working clinic: the doctor taps stop and walks away, and the desk reviews the draft, files it and runs the actions. Today the owner of the dictation does that. The next step is a clinic-level election that lets any member of the practice finish what the clinician started, with the trail showing who did which part. I wrote about what the desk already does in What ARAGS Does for Administrative Staff; this is the next thing it will do.

A dictated note was a recording you waited on. It is becoming a sentence you say once, and the system keeps up.