Where AI actually fits in an SLP's day
Strip away the hype and the practical use of AI for SLPs right now is documentation. For many clinicians, paperwork eats an hour or more a day: session notes, progress summaries, report drafts, the backlog that follows you home. That's the part of the job AI is currently good at helping with.
Two technologies matured at the same time to make this possible. Speech-to-text got accurate enough to transcribe a therapy session, and large language models got good enough to turn that transcript into a structured clinical note in your format. Put together, they let you finish a session and receive a draft note instead of a blank page.
What AI does not do: provide therapy, build rapport, interpret a swallow study, or decide whether a goal should change. The clinical work is still yours. The realistic promise is narrower and still worth having: a usable draft to edit instead of a note to write from scratch.
How the workflow works: record, transcribe, draft, review
Documentation-first AI tools follow the same basic shape:
- Record the session. With consent in place, you record audio on a phone, tablet, or computer while you run the session as usual. Many tools also accept typed or rough written notes if you'd rather not record.
- Transcribe. The audio is converted to text, often with speaker separation so the tool can tell clinician from client.
- Draft. The AI organizes the transcript into your note format — activities, targets, cues, client responses — following a template like SOAP or DAP.
- Review and finalize. You read the draft, correct anything the transcript got wrong, add what only you observed, and then export or paste the note into whatever system your workplace requires.
The clinician owns the note
AI drafts; you review, edit, and sign. That last step is not a formality — it's where clinical judgment enters the record. Be skeptical of any tool that implies its notes need no review.
As a concrete example: in SLPFlow (that's us, so read accordingly), you record a session or upload written notes, pick a note format such as SOAP or DAP — or define your own sections — and review the draft before copying it into your documentation system. Other documentation-first tools follow the same general pattern; templates, editing, and export are where they differ.
What a drafted objective section can look like
Objective: Client produced /r/ in the initial position of single words with 70% accuracy (14/20 trials) given moderate verbal and visual cues. Accuracy decreased to 40% (4/10 trials) at the phrase level. Client self-corrected on 3 occasions following a pause cue.
A draft like this comes from what was said aloud during the session. Your review adds what the recording can't capture — affect, attention, parent report, and your interpretation of the data.
Realistic benefits — and real limits
The honest case for AI documentation:
- Notes get drafted in minutes, not written from scratch at 9pm. If reviewing a draft takes three to five minutes where writing from scratch took ten to fifteen, a 25-session week returns roughly two to four hours.
- Notes get done the same day, while the session is fresh, instead of reconstructed from memory on Friday afternoon.
- Structure stays consistent across a caseload, which makes progress easier to track and reports easier to assemble later.
And the honest limits:
- Transcription isn't perfect. Disordered speech, articulation targets, young children, and noisy rooms are hard for speech-to-text. Expect to correct target words and phoneme notation in some drafts.
- The AI only knows what was said out loud. Nonverbal observations, affect, and anything you noticed but didn't verbalize have to come from you at review.
- Drafts can be wrong or over-general. A plausible-sounding sentence is not the same as an accurate one. Review every note before it enters the record.
- It doesn't reason clinically. The quality of your assessment and plan still comes from your training, not the software.
| Task | What AI can do today | What stays with you |
|---|---|---|
| Session notes | Draft a structured note from the recording or your rough notes | Verify accuracy, add observations, sign |
| Progress summaries | Summarize what's already documented across sessions | Interpret trends, decide what changes |
| Goal updates | Propose draft revisions based on recent session data | The clinical decision to adopt, adjust, or reject them |
| Diagnosis and treatment | Nothing | Everything |
Consent, privacy, and HIPAA
Session audio contains identifiable patient information. Before you record anything, three things need to be squared away.
Consent to record
Recording consent rules vary by state — some require consent from all parties — and your employer, district, or clinic may have its own policy on top of that. For minors, a parent or guardian consents. The clean way to handle it: add recording consent to your intake paperwork, explain that recordings are used to produce the session note, and document the consent itself.
A Business Associate Agreement
If HIPAA applies to your setting — it does for most clinics and private practices — any vendor that stores or processes patient audio, transcripts, or notes on your behalf must sign a BAA. School-based SLPs often fall under FERPA instead, but requiring a BAA is still the safe default. Either way: no signed agreement, no patient data.
A note on consumer AI tools
General-purpose chatbots and consumer transcription apps typically do not sign BAAs, and some use what you submit to improve their models. Without a BAA, they are not appropriate for anything containing patient information — even content that feels loosely identifying.
What happens to the data
Beyond the BAA, ask how data is protected in practice: encryption in transit and at rest, accounts isolated from one another, access controls and audit logging, and — the question worth asking out loud — whether patient data is ever used to train AI models. You should also be able to delete recordings and sessions when you no longer need them.
How to evaluate an AI documentation tool
Documentation tools are one slice of the broader question of choosing software for your practice. Whatever tool you're considering, put it through these questions:
- Will you sign a BAA? If the answer is anything but an immediate yes, stop here.
- Is patient data ever used to train AI models? Get the answer in writing.
- How is data secured? Look for encryption in transit and at rest, per-account isolation, and audit logging.
- Who reviews the note? The workflow should make your review and editing the default step, not a buried option.
- Does it speak your format? Templates that match your setting — and ideally custom sections — matter more than a long feature list.
- What happens with imperfect input? Can it work from written notes when recording isn't practical? How does it handle noisy audio?
- Can you get notes out easily? You'll likely be pasting or exporting into an EMR or district system every day.
- Is there a real trial? Test it on a few of your own consented sessions and compare the draft against what you would have written, before you pay anything.
Will AI replace SLPs?
No. Drafting a note is not delivering care. AI can't build a therapeutic relationship, adapt mid-session when a child shuts down, counsel a family, or take responsibility for a clinical decision. What it can do is take a real bite out of the administrative load that follows the clinical work — which is exactly the part most SLPs would happily hand off.
Getting started
You don't need to overhaul your practice to find out whether this helps. Sort out consent language first, pick one tool that clears the evaluation questions above, and run it on a handful of sessions during a free trial. Compare the drafts to your own notes, time the difference, and decide with your own data. If the drafts are close and the review is fast, you'll know within a week whether the hours are real.