People shop “AI scribes” as if the category is one product. It is not. Most of the disappointment I hear is a mismatch: they bought a listener and still work like a typist, or they bought dictation and expected it to remember last March.

This is the comparison I walk through when a practice is trying to decide what to change first.

Typing it after the visit

You already know this one. The patient leaves. You reconstruct. If you are honest, the note is a mix of what happened, what you always write, and what you can still remember at 6:40.

It is private. No microphone, no consent speech, no vendor. It also punishes you for being a good listener. The better the session, the worse the note, because you were not taking minutes.

I still type when the visit is short, the patient declined a listener, or I need one paragraph and I already know what it is. I do not type a 50-minute intake at the end of a clinic day and pretend that is quality.

Dictation (including the “AI” kind that waits for you to talk to it)

Dragon-style dictation, and the newer tools that clean up your speech, are faster typing. You still do a second shift. You stare at the chart and narrate the visit to a machine.

That is a real upgrade if your bottleneck is fingers. It is not an upgrade if your bottleneck is that you cannot hold the whole hour in your head. Dictation will faithfully write down a confident, incomplete story.

It also keeps you in the note. Some psychiatrists like that. They think in the assessment while they talk. Therapists who already hate the EHR usually do not want another ten minutes of talking to it.

Use dictation when you already know the structure and you are filling a template you trust. Do not use it as a memory prosthetic. It cannot recall a sentence you did not say into the mic.

Ambient listening

A listener in the room (or on the telehealth call) drafts from what was actually said. The source material is the visit, not your recollection of the visit.

That helps on long, messy encounters: the patient who circles back to the sleep problem at minute 38, the med list correction, the exact wording of a safety plan. It does less for you if the visit was mostly silence, or mostly paperwork, or you spent the hour in a collateral call the patient was not on.

You pay for that with process. Consent. A BAA. A review step, because a fluent draft can still be wrong. If you skip the review, you have a well-written note you did not write. That is worse than a sloppy note you did write.

Ambient tools also inherit the visit you ran. If you never asked about side effects, the draft will not invent a side-effect review. Some people call that a bug. I call that the tool refusing to lie for you.

How I would choose, in a real week

Most clinicians I talk to should not pick one mode for every visit. They should pick a default.

If your evenings are eaten by intakes and med-management follow-ups, ambient first. Review before the next patient, or before you stand up. If you let those drafts sit overnight, you will edit them like they are someone else’s chart, because they will be.

If your panel is mostly 20-minute med checks you could write in four minutes, dictation or a tight template is fine. Buying a listener so you can narrate a refill is theater.

If a patient does not want a microphone, type. Do not sulk. The visit still counts.

Nextvisit is in the ambient camp: it listens, it drafts a structured note, you edit and sign, and the note can go into the EHR you already use. That is the job. It does not replace dictation for people who think by talking to the chart, and it does not make a bad visit look complete.

Decide when the note gets written, and what it is written from. Memory, your spoken summary, or the visit itself. Those are different records.