
Transcription turns speech into text. Dictation turns your speech into text you intended to write. An ambient tool listens to a conversation without anyone addressing it. An AI scribe takes any of those and produces a structured document. They're four different products, and vendors use the words almost interchangeably.

I build in this category, so I've watched the vocabulary blur — sometimes carelessly, occasionally on purpose. Knowing which one you're being shown changes what questions you should ask, because the privacy implications and the failure modes are genuinely different.

![Hands typing on a laptop beside a document on a wooden desk](/images/article-body/pexels-artempodrez-8518848.jpg)

*Four words, four products. The differences matter more than the marketing suggests.*

## What does each one actually do?

**Transcription** converts speech to text and stops. The output is words in order, attributed to speakers if you're lucky. It makes no decisions about structure or relevance. Everything downstream is still your job.

**Dictation** is transcription of speech you're producing deliberately, for the document. You compose out loud; the tool types. The intelligence is entirely yours — the tool's job is accuracy and not making you fight it.

**Ambient** describes *how* audio is captured rather than what happens next: a tool listening to a naturally occurring conversation nobody is directing at it. Ambient capture usually feeds transcription, which usually feeds something else. It's a property of the front end, not a product category, which is exactly why it gets used loosely.

**AI scribe** is the one doing the most work and carrying the most ambiguity. It takes source material and produces a *structured document* — selecting, organising and phrasing. That's a different order of operation from the first three, because it makes editorial decisions.

## Why does the distinction matter for privacy?

Because the questions you'd ask are different for each.

For anything with **ambient capture**, the first question is about the room: who else is present, what they've been told, and how someone declines. That's a consent conversation rather than a software feature, and it's the reason clinicians tell me ambient tools get the most scrutiny from their colleges and their clients.

For **transcription and dictation**, the questions are about the pipeline: where the audio goes, whether it's retained, and what happens to the text afterward.

For an **AI scribe**, add a third: what did the tool decide, and can you verify it. A structured document is the product of choices, and you're signing those choices.

There's a general framework for the pipeline questions in [what to ask about where your data is processed](/articles/what-to-ask-about-where-your-data-is-processed).

## What are the different failure modes?

This is the practical reason to know which one you have.

**Transcription fails by mishearing.** Wrong word, wrong speaker, dropped phrase. Annoying, usually visible, and correctable because you can compare against what was said.

**Dictation fails by mistyping you.** Same category of error, and you'll normally catch it because you know what you meant.

**Ambient fails by capturing the wrong thing.** Side conversations, a colleague in the corridor, material nobody intended to be part of the record.

**An AI scribe fails by writing something confident and wrong.** This is the failure mode that matters, because a well-formed sentence that nobody said is much harder to notice than a garbled one. The output reads correctly, so the error hides inside fluency.

That difference should change how you review. Checking a transcript is proofreading. Checking a structured document means verifying claims against source material — a different task, and a slower one.

![Hands typing on a laptop showing a nearly blank document on a desk](/images/article-body/pexels-tima-miroshnichenko-6611932.jpg)

*A garbled transcript announces its error. A fluent invented sentence doesn't.*

## Which one solves an assessment write-up?

None of the first three, which is the point worth making.

A full assessment report isn't a record of a conversation — it's a synthesis of scores, history, observations and professional reasoning, most of which was never spoken aloud. Transcription has nothing to transcribe. Ambient has no conversation that contains the document. Dictation helps only if you were going to compose the whole thing out loud, which almost nobody does for a twenty-page report.

That's why session-note tools and assessment-report tools are genuinely different products even when the marketing is identical, and why I'd rather lose a therapist writing progress notes to something built for notes than pretend one tool serves both.

## What should you actually ask a vendor?

Three questions that cut through the vocabulary:

**"What is the input, and what is the output?"** Audio to text is one product. Structured source material to finished document is another.

**"What decisions does it make?"** If the answer is none, it's transcription with better branding. If it selects or organises, ask how you verify those choices.

**"Can I trace any sentence back to what I supplied?"** This is the one I'd press hardest, because it decides whether reviewing the output is a real check or a formality.

The wider criteria are in [how to choose documentation AI for a clinical practice](/articles/how-to-choose-documentation-ai-for-a-clinical-practice), and the protocol for testing any of them safely is in [trialling an AI report tool without risking a client file](/articles/how-to-trial-an-ai-report-tool-without-risking-a-client-file).

For the record, ours is a drafting tool — structured source material in, a first draft out, which you correct and sign. [That's what it does](/l/clinical-report-generator), and it's deliberately not the other three.

— Ian


AI Scribe vs Dictation vs Ambient vs Transcription: A Category Explainer


When I asked psychologists what actually mattered as they evaluated AI for psychological report writing, the answers came down to a short list of non-negotiables: they stay the author who reviews and signs, their client data stays in Canada under PIPEDA, the tool matches their own structure and length, and the drafts are defensible. Below are the vendor-neutral questions they told me to ask — criteria, not brand names.

I'll be upfront: I run one of the companies in this space, so I won't pretend I'm neutral about whether AI can help with the write-up. But I can pass along the *criteria* honestly, because the psychologists I talk to apply the same ones to any tool — including mine. They describe 10 to 15 hours of writing per assessment on top of the testing, so it's no surprise they're shopping for help. Here's how they told me they shop well.

![A professional's hands typing at a laptop keyboard](/images/article-body/pexels-gustavo-fring-6699309.jpg)

*The right tool drafts; the clinician decides what the report says. That's the line they judge every vendor against.*

## Does the tool keep the clinician as the author who reviews and signs?

This was the first question every psychologist raised, and they treated it as non-negotiable. They're signing their name to a legal document, so any tool that positions itself as making the interpretation or the decision was, in their words, disqualifying. The frame they wanted was narrow: the tool assembles a draft, and the clinician reviews, edits, and finalizes every word.

The question they said to put to a vendor directly: *does your workflow assume I read and edit every draft before it's final, or is it built to hand off a "finished" report?* The honest answer, they told me, is the first one. A drafting tool should feel like starting from a first pass in your structure — not like [carving a blank template down every time](/articles/why-report-templates-dont-solve-the-problem), and not like receiving a black-box report you're expected to trust. If a vendor is vague about how much you're meant to review, they treat that as their answer.

## Where does the client data actually live?

For a Canadian clinician, this is where a lot of tools quietly fail — and it was the question the psychologists I spoke with were most insistent on. They're handling sensitive client information, so they want to know, in writing, where that data is stored, whether it crosses the border, how long it's kept, and who can access it.

The factual bar they ask for: **Canadian-owned, with data hosted in Canada, under PIPEDA** (and **PHIPA** where it applies to them). A tool built and hosted in the United States is subject to different laws about who can compel access to data, regardless of how good the encryption is. Their advice: ask every vendor to state their data residency plainly and put it in the agreement.

![A laptop showing a padlock and a secured connection](/images/article-body/pexels-dan-nelson-1667453-4973899.jpg)

*"Where is my data at rest?" should have a one-sentence answer, in writing, before you sign up — the clinicians I talked to were firm on that.*

One more question they raised, specific to these tools: if it works from a transcription of the session or intake, ask where that transcription is stored and whether it's retained. "We don't keep it" is a meaningfully different answer, they noted, from "it's stored indefinitely on a US server."

A note on what they told me *not* to accept: be wary of any vendor leaning on a US compliance acronym as if it settles Canadian privacy. Those frameworks are US law; for their college and their clients, PIPEDA and provincial rules are what matter. They ask for verifiable facts — where data lives, how it's encrypted, whether it's de-identified, what's retained — not a reassuring badge.

## Does it match the psychologist's structure and length?

A tool can be perfectly capable and still cost time, they told me, if it produces a report in a shape they don't use. Assessment reports aren't one format — a psychoeducational report, a neuropsychological report, and an eligibility report each have their own sections, order, and length, and most psychologists have a house style on top of that.

So they ask to see it produce a draft in *their* sections, *their* order, and *their* typical length, on a report type they actually write. If it only outputs a generic structure they then reshape, they've moved the work rather than removed it. The whole point of [structuring a report a certain way](/articles/how-to-structure-a-psychoeducational-report), as they explained it, is that it reflects their reasoning; a tool worth adopting should bend to that, not the other way around.

## Are the AI-drafted reports defensible?

Defensibility is the bar this audience lives by, because these reports are read by referrers, physicians, schools, and sometimes courts and lawyers. Two things, the psychologists said, make a draft defensible: it stays grounded in what they actually observed and scored without inventing content, and it reads like their clinical reasoning rather than generic filler.

The question they put to a vendor: *does the draft stay tied to the information I provide, or does it generate plausible-sounding content on its own?* They want the former. A test several described running during a trial: check the draft against their own notes on a real, de-identified case, line by line — does anything appear that they didn't observe? Does it read like them, or like a form? [What makes a psychological report hold up](/articles/elements-of-a-defensible-psychological-report) doesn't change because a draft was AI-assisted, they're clear; the standard is the same, and the tool has to meet it.

## Who owns the tool, and will a college or employer approve it?

The best tool is worthless, they reminded me, if a regulator or employer won't allow it. Some colleges have issued guidance on AI in documentation, and some employers have their own policies, so a vendor should be able to hand over the paperwork — how consent is handled, what security measures are in place, and where data resides.

The question they ask: *can you give me consent and security documentation I can show my college or my clinic's privacy lead?* A serious vendor has that ready. This is also where Canadian ownership earns its keep, they told me: a company operating under Canadian privacy law, with data in Canada, is far easier to get approved. Lead with clinician control and documentation, they said, and a lot of the "can I even use this?" barrier disappears.

## What questions did psychologists say to ask any AI report-writing vendor?

Here's the whole checklist in one place, the way they assembled it for me — copy it into your notes and ask every vendor the same things:

1. **Author control** — Do I review, edit, and sign every draft? Is the workflow built around that?
2. **Data residency** — Is data hosted in Canada, and is the company Canadian-owned? Will you put it in writing?
3. **Privacy framework** — Do you operate under PIPEDA (and PHIPA where relevant)? What's retained, and for how long?
4. **Transcription handling** — If it works from a transcription, where is that stored, and is it deleted?
5. **Structure fit** — Can it draft in my sections, my order, and my typical length, on my report types?
6. **Grounding** — Does the draft stay tied to what I provide, without inventing content?
7. **Defensibility** — Does the output read like my reasoning and hold up to a referrer or a court?
8. **Approvability** — Can you give me consent and security documentation for my college or employer?

If a vendor answers all eight cleanly, the psychologists told me, you've found a serious tool. If they get evasive on data residency or on how much you're expected to review, that tells you what you need to know.

That checklist is also the bar we hold ourselves to at Soma: you review, edit, and sign; data is hosted in Canada; drafts stay grounded in your own material. You can [see a report drafted for you to review and sign](/l/clinical-report-generator) and judge it against the eight questions above.

Thanks for reading, and thanks to the psychologists who told me, plainly, which questions actually matter. Whatever you choose — mine or anyone's — hold it to these criteria, and don't hand over your clients' data without knowing exactly where it lives.

— Ian


What Psychologists Told Me Actually Mattered When Choosing AI for Report Writing


If you're looking for the best AI for psychological report writing, there isn't one answer — there are five tests, and they decide whether a tool survives contact with your practice: whether it produces a full-length report or just a summary, whether the draft arrives in your structure and voice, whether you can trace where client data actually goes, whether you remain the author of every line, and whether the vendor will tell you plainly what they haven't earned.

I should be straight about my position: I built one of these tools. This isn't a neutral survey — it's an argument for the choices we made, and you should read it that way. If you want the version written without a product in mind, including the tests where we come out badly, I put that in [how to choose documentation AI for a clinical practice](/articles/how-to-choose-documentation-ai-for-a-clinical-practice). This piece is narrower: it's about the full-length assessment report, and what I'd demand from anyone who claims to help with it.

![A professional reviewing a thick stack of documents at a glass desk](/images/article-body/pexels-ekaterina-bolovtsova-6077497.jpg)

*Every tool in this category is aimed at this pile. What separates them is which part of it they actually take.*

## Does it produce a full-length report, or a summary?

This is the first filter and it eliminates most of the field for assessment work.

A great many documentation tools are built for the session note — a few hundred words, produced quickly, cleared off the desk. That's a real problem worth solving, and it is not the problem a psychoeducational or neuropsychological report presents. The psychologists I talk to describe **ten to fifteen hours** on a full write-up. Nothing about a note-length tool addresses ten to fifteen hours.

So ask what the output actually is. Does it carry every section of the document you have to deliver — referral question through background, procedures, and results by domain drafted from your material, with the integration, summary, and recommendations scaffolded for the judgment only you can supply — or does it produce a tidy paragraph and leave the structure to you? A tool that hands back a summary has moved the easy part and left the hard part exactly where it was.

That's what we built for, deliberately and narrowly: the long document, not the note. It's also why I'd rather lose a therapist writing progress notes to something purpose-built for notes than pretend we're the answer to both.

## Does the draft come out in your structure and your voice?

The second filter, and the one that decides whether editing is faster than starting fresh.

A generic draft in a generic house style isn't a time saving. You'll rewrite it into your own structure, and rewriting someone else's prose is often slower than writing your own. The test is blunt: **does the draft come back close enough to your report that editing it beats a blank page?** If it doesn't, nothing else about the tool matters.

Which means the thing to demand is that the tool bends to your document, not the reverse — your section order, your headings, your length, the phrasings you always use. And the way to check it is to bring your own material to a trial rather than watching a vendor's sample. A demo document is chosen to flatter the demo.

I've written separately about the worry underneath this, which is whether a draft can [still sound like you](/articles/will-an-ai-drafted-report-sound-like-me). It's the first question almost every psychologist asks me, and I think it's the right first question.

## Where does your client data actually go?

Not "is it secure." Which jurisdictions, at which step, and what's kept.

Here's ours, including the part that's inconvenient. Our documentation processing runs on edge infrastructure in **Montreal**. There is **one step**, in report drafting, where content transits a large-model provider outside Canada. Clients are aliased so we're not holding identifiers. Content is encrypted in transit and at rest. Data used to improve the product is **off by default** and requires both your opt-in and each individual client's separate approval before anything is included.

I write that single out-of-country step down every time, because a clinician doing a privacy assessment needs the real path, and because a maple leaf on a homepage shouldn't be doing work that a sentence should do. If I've learned one thing worth passing on, it's that **"Canadian" is three separate claims** — who owns the company, where it's incorporated, and where the servers sit — and a vendor who blurs them is telling you something.

So the demand is simple: ask any vendor to walk the path step by step, and notice whether they name an exception or produce a diagram with no awkward parts. There's more on why this lands on you rather than the vendor in [the plain guide to what PIPEDA actually says](/articles/pipeda-for-clinicians-plain-guide).

![A professional reviewing a printed document and taking notes at a desk](/images/article-body/pexels-cottonbro-6814522.jpg)

*Whatever produces the draft, the reviewing and the signing stay yours.*

## Do you stay the author of every line?

Every psychologist I've spoken to over the past year draws this line in almost the same words: the interpretation is what they trained for, it's what their signature means, and they will not delegate it.

We built to that, and it costs us the flashier demo. Our software drafts prose from material you supply and review. It doesn't decide what a pattern of findings indicates. It isn't built to.

There's an operational version of this worth demanding too. "You review and sign" is a phrase everyone in this category uses, so press on it: **is the draft traceable back to what you actually supplied?** Reviewing a draft you can check against your own source data is a genuinely different act from approving fluent prose you'd have to re-derive from scratch. The second one isn't saving you time — it's moving the work and adding risk to your name.

## Will the vendor tell you what they haven't earned?

I've come to think this is the most diagnostic question of the five, and nobody asks it.

Ours: we describe de-identification as a control that lowers risk, not a guarantee of non-identifiability — the privacy research is [clear about the limits there](https://iapp.org/news/a/the-myth-of-anonymization-why-ai-needs-a-new-privacy-paradigm), and overstating it is the most common thing I see around us. And we don't name competitors, which means this article can't tell you we beat anyone; it can only tell you what to measure.

A vendor for whom everything is a strength is reading you a brochure. Ask what their tool does badly and listen to whether the answer is specific.

## Where would we not be the right fit?

Worth saying plainly, because a list of demands written by a vendor is only useful if the vendor will fail some of them.

If you write short session notes rather than long assessment reports, we're not built for your day. If your practice is standardised enough that a well-built template already gets you most of the way, the gain from drafting is smaller — I dig into that in [AI-drafted reports vs report templates](/articles/ai-drafted-reports-vs-report-templates). And if you write only a handful of assessments a year, the arithmetic probably doesn't work in our favour — the tool earns its keep on volume.

Where I think we earn a serious look is narrower and specific: a psychologist or neuropsychologist producing full-length assessment reports, who wants the draft in their own structure, who cares where the data goes and wants the exceptions named, and who intends to stay the author. That's who we built for.

If that's you, you can see how [the drafting reads on a report you review and sign](/l/clinical-report-generator). Then go and run these same five demands on everyone else you're considering — including on us, harder than on them.

— Ian


The Best AI for Psychological Report Writing? Five Things I'd Judge It On


A report template saves you the setup — the headings, the order of sections, the formatting — but you still type every sentence yourself. AI drafting starts you one step later: a full first draft, in your structure, that you review, edit, and sign. For most psychologists writing full-length reports, that difference is where the real hours are.

I'm not a clinician, so I won't tell you what belongs in your report. But I've watched a lot of psychologists work through the same bottleneck, and templates versus drafting is one of the most common questions I get. Here's the honest version.

![Hands typing on a laptop showing a nearly blank document on a desk](/images/article-body/pexels-tima-miroshnichenko-6611932.jpg)

*A template gives you the outline. The blank space under each heading is still yours to fill.*

## What is the real difference between a template and AI drafting?

A template is a fixed scaffold; AI drafting is a first pass at the writing itself. With a template, you open a document that already has your sections, headings, and boilerplate in place, and you write the narrative into it. With AI drafting, you feed in the scores, notes, and observations you've already gathered, and you get back a draft narrative in your structure that you then correct, reshape, and approve.

Both keep the clinical thinking entirely with you. Neither interprets the findings or makes the call — that's the part that can't be handed off, and shouldn't be.

## AI-drafted reports vs report templates: a side-by-side

Here's the at-a-glance comparison.

| Factor | Report template | AI drafting |
|--------|-----------------|-------------|
| **What it gives you** | Structure, headings, formatting, boilerplate | A first-draft narrative in your structure, from your data |
| **Where you start** | A blank space under each heading | A draft you edit rather than write from scratch |
| **Setup effort** | Build or buy once; reuse | Provide your data each time; optionally supply past reports for style |
| **Consistency** | High — same skeleton every time | High — plus consistent phrasing you then adjust |
| **Handles the actual writing?** | No — you write every sentence | Yes, for the mechanical assembly; you review and revise |
| **Clinical judgment** | 100% yours | 100% yours |
| **Best when** | Reports are simple or highly standardized | Reports are long and narrative-heavy |

## Where do report templates genuinely win?

Templates win on simplicity, cost, and control. They require no new tool, no data leaving your practice, and no vendor questions to answer. For short, highly standardized reports, a good template may get you close to as fast as anything — you know exactly what goes where, and there's little narrative to write.

Templates are also a real improvement over a truly blank page. If you don't use one today, building one is the highest-leverage free change you can make. I'd never tell a psychologist to abandon a template that's working.

The catch is that a template solves the *format* problem, not the *writing* problem. That distinction is the whole reason I wrote [why a better template alone doesn't solve it](/articles/why-report-templates-dont-solve-the-problem).

## Where does AI drafting go further?

AI drafting attacks the part templates leave untouched: turning data into prose. A template hands you an empty "Test Results and Interpretation" heading; drafting hands you a first pass at the descriptive paragraphs under it — the scores reported, the observations placed — built from the material you provide. What the results *mean* is not in the draft's gift: the interpretive call is yours to make and word, and you correct and sign every line.

The psychologists I talk to describe the slow part as staring at that empty space, not as missing headings. One person told me the formatting was never the issue — it was writing the same kind of narrative for the hundredth time. Drafting turns that from composition into editing, and editing a near-right paragraph is usually faster than writing one from nothing.

Some tools also learn your phrasing from a few past reports, so the draft comes back closer to how you'd have written it. You're still the one deciding whether each sentence is accurate and defensible — the tool just spares you the retyping.

![A stack of printed report forms with reading glasses and a pen on a desk](/images/article-body/pexels-cottonbro-5971243.jpg)

*The goal isn't fewer reports. It's fewer blank-page hours per report.*

## Which one saves psychologists more time?

It depends on how much narrative your reports contain. For short, checkbox-style reports, a strong template and drafting land close together. For long, narrative-heavy reports — the ten-to-fifteen-hour kind — drafting pulls ahead, because that's where the blank-page time lives: not in the judgment itself, which costs what it costs, but in all the prose that has to be assembled around it.

The honest answer for most assessment-focused psychologists is that it isn't either/or. Drafting works best *inside* a good structure — your template is the skeleton, the draft fills it in, and you do the judgment throughout. If your templates are starting to feel like a workaround rather than a help, that's usually a sign, which I unpack in [5 signs your report templates are slowing you down](/articles/signs-your-report-templates-are-slowing-you-down).

Thanks for reading. If you'd like to see what "a draft in your structure that you review and edit" actually looks like, you can watch Soma [draft a report you review and sign](/l/clinical-report-generator) — and keep your template right where it is.

— Ian


AI-Drafted Reports vs Report Templates: Which Saves Psychologists More Time?


A template gives you the skeleton of a report — headings, section order, the scaffolding. It does not give you the writing. That's why clinicians who already have a strong template still lose 10 to 15 hours per report. The hours live in the prose and the assembly wrapped around your thinking, and a template can't touch that part.

I want to make the honest argument here, not the easy one. Templates are useful. Most of the psychologists I talk to already have one. The problem isn't that they're missing a tool — it's that the tool they have was never designed to do the hard part.

![A stack of printed report forms with reading glasses and a pen on a desk](/images/article-body/pexels-cottonbro-5971243.jpg)

*A template hands you an empty container to fill by hand.*

## Don't report templates already solve this?

They solve the structure, and that's real — but the structure was never where the hours went. When I sit with a clinician and walk through where the time actually goes, the template is already open. They're not staring at a blank page wondering what sections to include. They know.

What they're doing instead is reading through their findings, weighing them, and writing the narrative that ties it all together in their own clinical reasoning. That's the work. A template hands you an empty container — you still have to fill every inch of it by hand.

![An open blank notebook and pen beside a laptop on a desk](/images/article-body/pexels-negativespace-34601.jpg)

*The scaffolding repeats; the writing on top does not.*

## Why does a template still leave hours of work?

Because a template standardizes the scaffolding, not the thinking-onto-the-page. Roughly 60 to 70% of a report's structure repeats from one to the next — the same sections, the same order, the same bones. A good template captures that beautifully.

But the clinicians I work with still rebuild that scaffolding by hand every single time, then pour the findings in on top, one paragraph at a time. The repeating 60 to 70% gives you a head start on *layout*. It gives you almost nothing on the writing, which is the part that takes 10 to 15 hours even with the template right there.

It's worth naming what that time costs. In a survey of 93 health-service psychologists, most described themselves as working at or beyond capacity — 41% "at full capacity," 39% "over-extended" — with [non-billable clinical activities among the most commonly named sources of work-related stress](https://pubmed.ncbi.nlm.nih.gov/31728882/). A template doesn't move that load, because the load isn't structural — it's the writing. I dug into [why report writing takes so long](/articles/why-psychological-report-writing-takes-so-long), and into [the report-writing problem underneath all of it](/articles/report-writing-problem-private-practice).

## What's the difference between a template and a draft in your voice?

A template is an empty form you fill in. A draft in your voice is a report that's already laid out the way *you* would write it — assembled from your own material, ready for you to read, correct, and sign off. That's the difference between starting and finishing.

This is the distinction I care about most, because clinicians have told me plainly: they don't want a generic template that flattens their reasoning into someone else's house style. They want *their* voice — the way they phrase a recommendation, the way they sequence an interpretation, the clinical judgment that makes the report theirs.

A static template can't carry your voice. By design, it's the same for everyone who uses it. The goal — the only goal worth aiming at — is a report that reads like *yours*, not a generic template. The assembly done, the prose drafted, the judgment still entirely yours.

## Are templates still worth using?

Yes — and I want to be clear about that, because I'm not here to knock a tool that genuinely helps. A template is a fine starting point. It keeps your sections consistent, it stops you forgetting a heading, and it gives a new clinician a sane place to begin. If a template works for you, keep it.

What I'd gently push back on is the idea that a better template is the answer to the time problem. You can refine a template for years and it will still be a container. It will never hand you a draft to correct instead of a form to fill, because that was never what a template was for. So use it — just don't expect it to give you your evenings back.

![Macro of a fountain-pen nib resting on graph paper](/images/article-body/pexels-pixabay-261719.jpg)

*Moving from an empty form to a drafted report in your own voice.*

## What actually removes the hours?

What removes the hours is moving from structure to outcome — from an empty form to a drafted report in your own voice that you then refine. Not a faster way to fill the template. A way to skip the filling and land on a draft you're correcting instead of composing.

The way I think about it, the clinician's time should go where their expertise actually matters: reading the draft, sharpening the interpretation, catching the nuance, making the call. Not rebuilding the same scaffolding for the hundredth time and hand-typing prose into it.

That's the whole reason Soma exists — and it's why I draw such a hard line between a template and a draft. A template standardizes your starting point. The outcome we care about is you opening a report that already reads like yours, with the assembly handled and the judgment left where it belongs — with you. And to be exact about the line: Soma drafts prose from the material you supply. It can suggest a possible phrasing; it doesn't interpret your findings, doesn't reach conclusions, and doesn't decide what a pattern means. That part can't be handed off, and we built it that way on purpose. You can see how we frame that on our [product page](/product).

I'll say the honest version once more: templates are good, and you should keep the one that works. They just solve the part that was never the bottleneck. The 10 to 15 hours don't live in the structure — they live in the writing, and that's the part worth getting back. See how Soma [drafts a report you review and sign](/l/clinical-report-generator).

Thank you for reading, and thank you to every clinician who's walked me through exactly where their hours go — it means a great deal, and it's shaped how we think about all of this.

— Ian


Why Report Templates Don't Solve the Problem


## The Elephant in the Room

Let's start with what everyone already knows: therapists, counsellors, and psychologists are using ChatGPT for clinical documentation.

They're copying session notes into ChatGPT and asking it to restructure them into SOAP format. They're dictating summaries and asking for polished progress notes. They're using it to draft treatment plans, generate psychoeducational handouts, and brainstorm intervention strategies.

This isn't hypothetical. It's happening in private practices, clinics, and hospitals right now. And for good reason -- documentation takes too long, ChatGPT is free and fast, and the output is often surprisingly good.

But "surprisingly good" isn't the same as "clinically appropriate, legally compliant, and ethically sound." And the gap between those two things is where clinicians face real professional risk.

![Abstract rendering of woven blue data structures on a pale field](/images/article-body/pexels-googledeepmind-17483906.jpg)

*General-purpose models are capable — the question is whether they fit clinical work.*

## What ChatGPT Gets Right

Credit where it's due. ChatGPT and similar large language models are genuinely capable tools:

- They can structure unformatted text into clinical note formats (SOAP, DAP, BIRP) with reasonable accuracy
- They understand clinical terminology and can generate clinically relevant language
- They can summarize lengthy session transcripts into concise documentation
- They're available immediately, require no training, and cost little or nothing

For clinicians drowning in documentation -- and 82% report that admin work contributes to their burnout (Google Cloud/Harris Poll, 2024) -- the appeal is obvious. Any tool that reduces the 15-20 minutes per note of after-hours charting feels like a lifeline.

![A row of server storage drives with cabling bathed in blue light](/images/article-body/pexels-cookiecutter-17489150.jpg)

*Paste clinical content into a general chatbot and it travels to servers you don't control.*

## The Problems Nobody Talks About

### Privacy and Compliance

This is the most critical issue, and it's often the least considered in the moment of "I just need to finish this note."

When you paste clinical session content into ChatGPT, that data is transmitted to OpenAI's servers. Under OpenAI's standard terms of service, this data may be used for model training (unless you've opted out through specific API configurations). Even with opt-out, the data is processed on servers that are not designed for regulated-data compliance.

In Canada, PIPEDA (Personal Information Protection and Electronic Documents Act) requires that personal information be protected with appropriate safeguards, used only for the purpose for which it was collected, and stored in a manner that ensures confidentiality. Provinces with additional legislation -- Ontario's PHIPA, Alberta's HIA -- add further requirements.

Entering identifiable clinical data into ChatGPT potentially violates:
- Your obligation to protect client confidentiality
- PIPEDA/PHIPA requirements for data handling
- Your regulatory college's standards of practice
- Your professional liability insurance terms

### No Audit Trail

Clinical documentation is a legal record. If a note is ever reviewed in a licensing investigation, liability claim, or court proceeding, you need to demonstrate how it was created, when, and by whom.

ChatGPT conversations have no clinical audit trail. There's no timestamp linked to your clinical record system, no chain of custody, and no way to demonstrate that the AI-generated draft was appropriately reviewed and modified before being finalized.

### No Session Context

Every ChatGPT interaction starts from zero. It doesn't know your client's history, their treatment goals, previous session content, or the therapeutic modality you're using. You have to provide this context every time -- which means either typing extensive prompts (adding time rather than saving it) or accepting generic output that lacks clinical specificity.

### No Clinical Note Standards

ChatGPT can approximate clinical note formats, but it doesn't enforce them. It may mix subjective and objective observations in a SOAP note, omit required elements in a DAP note, or generate recommendations that don't align with the actual session content. Without built-in clinical format validation, the clinician bears the entire burden of quality assurance.

![Close-up of hands typing on a laptop at a desk](/images/article-body/pexels-israwmx-17810854.jpg)

*Purpose-built tools fit the clinical workflow instead of sitting beside it.*

## What Purpose-Built Clinical AI Looks Like

The alternative to generic AI isn't going back to handwriting notes. It's using AI tools that are architecturally designed for clinical work. Here's what to look for:

### Consent-First Design

Clinical AI should operate within a consent framework from the ground up. This means:
- Client data is collected only with explicit consent
- The client controls what information is shared and can revoke consent at any time
- All data processing occurs within a compliant infrastructure
- Client anonymity is maintained — no personal identifiers are required to use the platform

### Client Anonymity by Design

This is a critical architectural difference. In purpose-built clinical AI, clients can be identified only by alias — no names, no emails, no dates of birth, no card numbers. The platform never needs to know who the client is. Identity information, when needed for EMR purposes, is stored in a separate encrypted vault and linked to clinical data only at the moment a professional needs to view it.

This means that even in a breach scenario, clinical data and identity data are never stored together.

### Built-In Clinical Formats

Rather than prompting an AI to "write this in SOAP format," purpose-built tools have clinical note formats embedded in their architecture. The clinician selects SOAP, DAP, BIRP, or progress note format, and the output conforms to the structure automatically -- with appropriate section headers, required elements, and clinical language conventions.

### Session Context and Continuity

Purpose-built clinical AI maintains context across sessions. It knows the client's treatment history, previous session themes, treatment goals, and therapeutic modality. This means the AI can generate notes that are contextually appropriate without the clinician providing extensive background every time.

### EHR Overlay Architecture

Rather than requiring clinicians to switch to a new system, effective clinical AI works as an overlay on existing EHR platforms. Notes generated by the AI can be reviewed, edited, and integrated into the clinician's existing documentation workflow without a copy-paste step.

## AI Assists, Clinician Decides

One principle is non-negotiable across all clinical AI applications: the clinician is always the final decision-maker.

AI-generated notes are drafts. They require clinical review, editing, and approval before they become part of the clinical record. The AI doesn't make clinical judgments or treatment decisions. It handles the administrative task of structuring documentation so the clinician can focus their expertise on the clinical content.

This isn't a philosophical position -- it's a practical requirement. AI models can miss nuance, misinterpret tone, or generate language that doesn't accurately reflect the session. The clinician's review is what transforms an AI draft into a clinical document.

## The Decision Framework

For clinicians currently using ChatGPT for documentation, the decision isn't about whether AI helps -- it clearly does. The decision is about which AI tool manages the risk appropriately.

**Continue using ChatGPT if:**
- You never enter any identifiable client information (no names, dates, identifying details)
- You treat the output as a rough starting point and substantially rewrite it
- You're comfortable with the compliance risk under your jurisdiction's privacy legislation
- You don't need session context, clinical format validation, or audit trails

**Switch to purpose-built clinical AI if:**
- You work with clients whose data is protected under PIPEDA, PHIPA, HIPAA, or equivalent
- You need clinically structured output that doesn't require extensive reformatting
- You want session context maintained across encounters
- You need an audit trail for documentation
- You want to reduce documentation time without increasing compliance risk

## The Bottom Line

ChatGPT is a remarkable general-purpose tool. It's also a poor fit for clinical documentation when privacy, compliance, and clinical accuracy matter -- which is always.

Purpose-built clinical AI tools exist precisely because the requirements of clinical documentation exceed what general-purpose AI can safely provide. They're not more expensive than the compliance risk of using ChatGPT with client data. They're not slower than the prompting and reformatting workflow that ChatGPT requires. And they're not harder to use -- they're typically simpler, because the clinical workflow is built in rather than bolted on.

The 2025 research shows that AI scribing reduces burnout by 13 percentage points in 30 days and saves thousands of hours of documentation time. The benefits are real. (I broke down [that 2025 AI-scribing research](/articles/2025-09-22_AI_Scribing_Private_Practice_2025_Research) and [what an AI scribe for therapists actually does](/articles/ai-scribe-for-therapists-transcription) separately.) If you're ready to move past ChatGPT, here's [a clinician-first alternative that drafts in your voice](/l/chatgpt-for-therapists). The question is whether you access those benefits through a tool designed for the job or one that was designed for something else entirely.

---

*References: Google Cloud/Harris Poll (2024); OpenAI Terms of Service; PIPEDA (Personal Information Protection and Electronic Documents Act); PMC Ambient AI Scribe Study (2025); Permanente Medical Group/AMA (2025).*


ChatGPT and Therapy: What's Real, What's Noise, and What Clinicians Should Actually Use


## The Download Paradox

In 2023, wellness apps were downloaded over 500 million times globally. The market is booming. Yet behind the download numbers lies a stark reality: most users abandon these apps within weeks.

A systematic review by Baumel et al. (2019) examined engagement across 93 wellness apps and found a median 30-day retention rate of just 3.9%. That means for every 100 people who download a standalone app, fewer than 4 are still using it a month later. The median 15-day retention was only 3.3%.

This isn't a design problem. It's a structural one.

![Abstract network of glowing nodes](/images/article-body/pexels-lucasdc-10325707.jpg)

*Downloads are easy; sustained engagement is the harder problem.*

## Why Standalone Apps Fail

Standalone wellness apps face a fundamental challenge: they ask individuals who are often struggling with motivation, executive function, and consistency to maintain a self-directed practice without external support.

The research identifies several consistent failure points:

**No accountability loop.** Without a clinician reviewing the data, there's no external motivation to continue. The app becomes another thing on the to-do list that feels optional.

**No personalization over time.** Generic content and exercises don't adapt to the user's actual clinical needs. A journaling prompt designed for general wellness feels irrelevant to someone working through a specific clinical concern.

**No integration with treatment.** The data captured in standalone apps rarely makes it back to a clinician. Even when it does, it arrives as a PDF export or screenshot -- disconnected from the clinical workflow.

**No relational anchor.** Clinical recovery is fundamentally relational. An app without a human connection on the other end lacks the working relationship that research consistently identifies as the strongest predictor of outcomes (Wampold, 2015).

![Laptop showing a video-call grid](/images/article-body/pexels-ivan-s-8117408.jpg)

*A clinician on the other end changes how a tool gets used.*

## The Therapist-Connected Difference

Research by Mohr et al. (2017) demonstrated that therapist-connected digital interventions show significantly higher engagement and retention than standalone alternatives. When a clinician is involved -- assigning activities, reviewing data, and adjusting treatment based on what the app captures -- client engagement fundamentally changes.

The mechanism is straightforward: the accountability and personalization loop.

1. **Clinician assigns** specific activities, journaling prompts, or tracking tasks tailored to the client's treatment goals.
2. **Client completes** the assigned work in the app, generating between-session data.
3. **Clinician sees** the data before the next session -- mood patterns, journal themes, activity completion, wearable signals.
4. **Treatment adjusts** based on real data rather than retrospective self-report.

This loop transforms the app from a standalone self-help tool into an extension of the clinical relationship. The client knows their clinician will see the data, which provides both motivation and meaning.

![Abstract translucent data blocks](/images/article-body/pexels-googledeepmind-17485706.jpg)

*Connected data gives the clinician a fuller view between sessions.*

## The Data That Matters

[Between-session data](/articles/2025-10-13_Between_Session_Journaling_Evidence) is where the real clinical value lies. Research shows that [clinicians detect only 1 in 20 deteriorating clients](/articles/2025-09-29_The_1_in_20_Problem) without algorithm-based monitoring (Hannan et al., 2005). The weekly snapshot provided by a 50-minute session simply isn't enough to capture the complexity of a client's week.

When clients use a therapist-connected app between sessions, clinicians gain access to:

- **Mood patterns** across the full week, not just the in-session report
- **Journal entries** that capture thoughts and emotions in real time
- **Activity completion** data that shows engagement with assigned homework
- **Wearable data** (sleep, heart rate variability, activity levels) that provides objective behavioral markers
- **Trend lines** that make gradual change — easy to miss in a single weekly snapshot — visible for the clinician to evaluate

This continuous signal replaces the retrospective recall that dominates traditional practice. And the research on retrospective recall is not encouraging: ecological momentary assessment studies show that retrospective self-reports correlate only r = 0.4 to 0.6 with real-time data (Shiffman et al., 2008). Clients don't accurately remember their week. Real-time capture is more reliable.

## Digital Homework and Treatment Outcomes

The connection between homework compliance and treatment outcomes is well-established. Research has shown that digital homework tracking improved completion rates from 49% to 71% -- a meaningful increase that translates directly to better outcomes.

Each 10% increase in homework compliance has been associated with a 2.6-point reduction on a standard clinical symptom-severity measure, a clinically meaningful change. When clients complete their between-session work, treatment works better.

Therapist-connected apps make homework tracking automatic. Instead of asking a client "Did you do the breathing exercises?" and relying on their memory, the clinician opens the dashboard and sees exactly what was completed, when, and how the client felt during and after the activity.

## Privacy and the Consent Question

One of the legitimate concerns about therapist-connected apps is privacy. If a clinician can see between-session data, how is client autonomy protected?

The answer lies in consent-first architecture. Effective therapist-connected apps should operate on explicit, revocable consent. The client controls what data is shared, with whom, and can withdraw consent at any time. Data flows only because the client has actively chosen to share it.

This approach aligns with PIPEDA requirements in Canada and represents a meaningful departure from many standalone apps, which monetize user data through advertising or sell anonymized datasets to third parties.

## SOMA's Approach

SOMA's model embodies the therapist-connected approach with a specific emphasis on anonymity and consent. The client app is completely anonymous -- no personal identifiers are required. Clients connect to their clinician through a consent code, and data flows only through that consent-driven channel.

Clinicians see between-session data in a pre-session dashboard: mood trends, journal entries, activity completion, and wearable signals. This data feeds into the AI Prep feature, which drafts possible themes and suggests patterns that may be worth a look before each session — suggestions the clinician reviews and confirms, revises, or discards. The clinical read of the week is always the clinician's; the software never makes it.

The result is a platform where the app isn't a standalone tool the client uses in isolation. It's a bridge between sessions -- one that keeps the clinical relationship present even when the clinician isn't in the room.

## The Bottom Line

The evidence is clear: standalone wellness apps have a retention problem that no amount of UX polish can solve. The missing ingredient is the clinical relationship.

When apps are connected to a clinician -- when there's an accountability loop, personalized content, and clinical integration -- engagement increases, homework compliance improves, and treatment outcomes get better.

The 3.9% retention rate for standalone apps isn't a failure of technology. It's a reminder that clinical support works best when it's relational. Technology should amplify that relationship, not replace it.

---

*References: Baumel et al. (2019), JMIR; Mohr et al. (2017), Annual Review of Clinical Psychology; Hannan et al. (2005), Clinical Psychology & Psychotherapy; Wampold (2015), The Great Psychotherapy Debate; Shiffman et al. (2008), Annual Review of Clinical Psychology.*


Comparisons