
The clearest warning signs in this category are things a vendor *says*, not things a product does — which is useful, because it means you can spot most of them on a first call before you've committed a minute of trial time. The eight below came from clinicians describing what they wish they'd noticed sooner.

I build one of these tools, so I'm describing a field I'm standing in. Several of these are tests we'd pass and at least one is a place we'd have to give you an awkward answer. That's the point of writing them down.

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

*Most of these show up in the first conversation, long before a trial does.*

## 1. Everything is a strength

The most reliable signal, and the easiest to test: ask what the tool does badly and where they'd tell you not to use it.

A vendor who knows their product answers crisply and specifically. A vendor who deflects into a benefit is reading you a brochure. Clinicians told me this question changed their shortlist more than any feature comparison did.

## 2. "Compliant" used as a finished state

No software is compliant on its own. Compliance describes how information is handled in your practice — the obligation stays with you, and a vendor can supply safeguards but not carry the duty.

So "we're compliant, so you're covered" is either a misunderstanding or a convenient one. The better answer sounds like: here are the specific safeguards you can point to when you do your own assessment.

## 3. The demo runs on their document, not yours

A sample case is chosen to flatter the sample. If you can't run your own structure, your own section order and your own length before paying, you're evaluating a performance rather than a product.

This is the one clinicians said cost them the most: a tool that looked excellent on a tidy two-page example and fell apart on a real twenty-page report with an uneven profile.

## 4. Time savings quoted as a single number

A saving depends on your document length, your structure, how much you edit and how fast you already write. A figure without those anchors is describing somebody else's practice.

Ask what the baseline was and how it was measured. If there isn't an answer, the number is decoration. There's a method for producing your own in [how I'd measure whether an AI tool actually saves you time](/articles/how-to-measure-whether-an-ai-tool-saves-you-time).

## 5. Vagueness about where content is processed

Storage location and processing location are different questions, and a page can answer the second while implying the first. Ask for the path step by step, and notice whether the vendor names an exception.

Almost every real architecture has one. A vendor who volunteers theirs has thought about it; a vendor whose diagram is entirely tidy either hasn't looked or isn't telling you. The full question set is in [what to ask about where your data is processed](/articles/what-to-ask-about-where-your-data-is-processed).

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

*Ask for the path. Notice whether they name the awkward step or produce a tidy diagram.*

## 6. De-identification described as a guarantee

It's a control that lowers risk, not a mathematical promise that content can never be reconnected to a person. A vendor presenting it as absolute has told you how carefully they read their own field.

Related: if "we don't train on your data" is offered as the complete answer to privacy, ask separately about retention, sub-processors, and what happens on your device before transmission. Those are four claims, not one.

## 7. Output you can't trace back to your input

The one I'd weight most heavily, because it decides whether reviewing the document is a genuine check or a formality.

If you can take a sentence in the draft and find what it came from, review is real. If the draft is fluent and unverifiable without redoing the work, you've been handed liability with a time saving attached. Test this deliberately: leave a gap in your trial material and see whether the tool flags the absence or confidently fills it.

## 8. No clean way out

Ask on day one how you'd leave: what you can export, in what format, whether you get underlying content or only rendered files, and what's deleted when you cancel.

A vendor with a straightforward answer expects to be kept for good reasons. A vague one is relying on it being hard to go.

## Where would we set off one of these?

Number five, if you push far enough — and I'd rather say it here than have you find it.

Our documentation processing runs on edge infrastructure in Montreal, and there is one step in report drafting where content transits a large-model provider outside Canada. That's the awkward exception in our own path. I write it down every time because a clinician doing a privacy assessment needs the real answer, not a flag on a homepage.

Run all eight at us alongside everyone else. The full evaluation framework is in [how to choose documentation AI for a clinical practice](/articles/how-to-choose-documentation-ai-for-a-clinical-practice), and if you want to see the drafting first, [it's here](/l/clinical-report-generator).

— Ian


Red Flags When Evaluating Documentation AI


To find out whether an AI tool saves you time, measure your own baseline for two weeks first, then measure the same work with the tool for two weeks, and compare only the stage the tool actually touches. A vendor's headline number — mine included — is a claim about somebody else's practice.

I build one of these tools and I'd rather you held me to a number you produced than one I did. This is the method I'd use, and it's deliberately simple enough that you'll actually finish it.

![Hands typing on a laptop beside a wristwatch on a wooden table](/images/article-body/pexels-iamhogir-17781874.jpg)

*The only number that means anything is the change against your own baseline.*

## Why is a vendor's time-saving number close to useless?

Not because vendors are lying — mostly they aren't — but because the number is unanchored.

A time saving depends on the length of your documents, your section structure, how much you edit, how fast you write already, and what stage of the work the tool touches. Change any of those and the figure moves. A saving measured on short session notes tells you nothing about a twenty-page assessment report.

There's also a reporting asymmetry worth knowing about: comparisons tend to be drawn against a slow baseline, because that's the version that makes the difference visible. Your baseline is whatever you actually do now, which is usually faster than the one in the case study.

So the number to want isn't "how much time does this save?" It's "how much time does this save **me**, on **my** documents?" That has an answer, and you can get it in a month.

## What's the baseline, and how do you get one?

Two weeks of honest timing before you change anything.

Time only the stage you're evaluating. If you're assessing a report-drafting tool, time from *"I have everything I need and I'm starting to write"* to *"this is signed"*. Don't include testing, scoring or the feedback session — the tool doesn't touch them, so including them dilutes the measurement toward zero.

Record three things per document: the stage time, the document type, and roughly how long the finished piece is. A note on your phone is enough. What you want at the end is a small table, not a precise average — five or six reports gives you a usable picture of your normal range.

Most people find the baseline itself useful regardless of what they buy. Several clinicians told me the exercise was the first time they'd seen the write-up separated from the rest of the assessment, which is also the thing that makes the bottleneck arguable in a staffing conversation. The broader arithmetic is in [what an assessment report actually costs a practice to produce](/articles/what-an-assessment-report-costs-a-practice-to-produce).

## What do you time once the tool is in?

Exactly the same interval, on the same kind of documents.

The measurement is **draft to signed**, and it starts when the draft arrives — not when the tool finishes generating. Time spent reading, correcting, restructuring and rewriting is all part of the cost. A tool that produces a draft in ninety seconds and then takes you two hours to fix has not saved you two hours.

Keep timing for at least two weeks, and keep going past the first few. Nearly everyone gets faster with a new tool over the first several documents as they learn what to hand it and what to expect back. If you stop after three, you'll measure the learning curve rather than the tool.

![A digital clock glowing in the small hours on a desk at night](/images/article-body/pexels-introspectivedsgn-9582658.jpg)

*Draft to signed. The clock starts when the draft arrives, not when the tool stops generating.*

## What will make the number lie to you?

Three things, and all three are common.

**Timing a quiet fortnight.** A tool trialled during a light week is measured under conditions that don't resemble the week you needed it. Run the comparison across a normal or busy stretch or the result won't transfer.

**Changing two things at once.** New tool plus a new template plus a new protected writing block, and you'll never know which one moved the number. Change one thing.

**Counting the wrong stage.** If the tool touches drafting and you measure the whole assessment, a real 40 percent improvement in the write-up shows up as a modest overall change and looks like noise.

There's a subtler one too: **quality drift.** If the "signed" document is quietly worse than your baseline document, the time saving isn't real, it's borrowed. Which leads to the check that matters most.

## What does a real result look like?

Two numbers and one judgment.

**The time delta on the stage the tool touches**, expressed against your baseline range rather than as a single figure — "usually three to four hours, now usually two" is a more truthful result than "38 percent faster".

**Consistency.** Did every document improve, or one dramatically while the rest stayed flat? A single spectacular case is usually a case that suited the tool, not a pattern.

**And the judgment: would you have signed these anyway?** Pull two of the tool-assisted documents and read them cold against two baseline ones. If the assisted versions are thinner in the analytic sections, the tool moved work rather than removing it — and you've traded quality for hours without deciding to.

If the delta is real, consistent, and the documents hold up, that's a genuine result. If it isn't, you've spent a month and learned something true about your own practice, which is worth having either way.

Run this on us with everything else. The protocol for the trial that feeds it — synthetic case, no real client file — 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), and the drafting itself is [here](/l/clinical-report-generator).

— Ian


How I'd Measure Whether an AI Tool Actually Saves You Time


You can run a real trial of an AI report tool without a single real client file going through it. Build one synthetic case, run it through every tool you're considering, measure the same four things each time, and decide against a threshold you set beforehand. It takes an afternoon to set up and it will tell you more than any demo.

I build one of these tools, so read this knowing I have an interest in you trialling carefully rather than being persuaded quickly. That's genuinely my preference — a customer who tested is a customer who stays.

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

*The trial that tells you something starts before any real client information exists.*

## Why not just trial it on a real case?

Because of the order of operations. Evaluating a vendor means deciding whether you trust them with client information — and running the trial on client information means you've already answered that question in order to ask it.

There's also a consent problem. Putting a real file through a tool you haven't adopted, for the purpose of deciding whether to adopt it, is difficult to describe cleanly in a consent conversation. Clinicians raised this with me before I'd thought it through properly.

And practically, a real case makes a poor test. You can only run it once, you can't compare tools on identical input, and you'll be judging output you're emotionally invested in.

## What should the test material be?

**A synthetic case you construct yourself.** Invent the person. Invent the history. Make up the scores. Nothing traceable to anyone, including composites of real clients — a composite is still derived from real files.

Make it representative rather than easy. Specifically:

- **Full length.** If your real reports run twenty pages, a two-page test tells you nothing about the tool at the length where it matters.
- **An uneven profile.** A flat, tidy case flatters every tool. A scattered one is where they separate.
- **Something missing.** Omit a piece of information a real file would sometimes lack, and see whether the tool flags the gap or fills it in confidently.
- **Your own structure.** Your section order, your headings, your typical length.

Build it once, save it, and reuse it for every vendor. Identical input is what makes the comparison mean anything.

## What do you actually measure?

Four things, written down before you start.

**Time from draft to signed.** Not "did it produce something" — how long from receiving the draft to a document you'd put your name on, against your own current baseline for the same work. If you don't know your baseline, measure that first; there's a method in [how I'd measure whether an AI tool actually saves you time](/articles/how-to-measure-whether-an-ai-tool-saves-you-time).

**Edit distance in the sections that matter.** How much rewriting did the descriptive sections need — background, procedures, results by domain — versus the boilerplate? Heavy editing of boilerplate is fine. Heavy editing of the descriptive sections means the tool isn't doing the part you needed. The integration section is a different case: the judgment there is yours to write, and you should expect to write it. A tool that hands you a confident integration you didn't reason your way to is a warning sign, not a time saving.

**Fabrication count.** Did anything appear in the draft that wasn't in what you supplied? This is the one to be strict about. A single invented finding in a test case is a disqualifying result, not a rough edge — and the gap you deliberately left is how you'll catch it.

**Voice match.** Read it aloud. Does it sound like you, or like a competent stranger? Rewriting someone else's prose into your voice is often slower than writing your own.

![Hands typing on a laptop with handwritten notes beside it on a desk](/images/article-body/pexels-rdne-8124192.jpg)

*Identical synthetic input across every vendor is what makes the comparison mean anything.*

## What should a vendor let you do?

The trial itself is a test of the vendor, not just the product. What I'd expect:

- **Your own material, not their sample.** A demo document is chosen to flatter the demo. If you can't run your own structure through it before paying, that's an answer.
- **A trial without a card, or with a clean exit.** And a plain statement of what happens to your test content afterward.
- **A straight answer on what the trial version does differently.** Some trials run a smaller model or a shorter length limit. Fine — but you need to know, or you're evaluating a different product from the one you'd buy.
- **Someone technical to talk to.** Ask a specific question and see who answers.

## What did clinicians say went wrong in their trials?

Three things, all recognisable.

**They trialled during a quiet week.** Which means they measured the tool under conditions that don't resemble the week they actually needed help in. Trial it during a busy stretch or the result won't transfer.

**They didn't set a threshold beforehand.** Without a number decided in advance, the decision drifts into how the demo felt. Write down what "good enough to adopt" means before you see any output.

**Only one person tried it.** In a group practice this is a real trap — a tool that suits one clinician's writing style can be useless to a colleague. If more than one person will use it, more than one person runs the case.

## What would I say about trialling us?

Run exactly this protocol on us, with the same synthetic case you use on everyone else, and hold us to the same fabrication threshold. If our draft doesn't beat your blank-page baseline on a full-length uneven case, we haven't earned the seat.

The wider set of criteria worth scoring every vendor against — data path, authorship, pricing at your real volume, exit terms — is in [how to choose documentation AI for a clinical practice](/articles/how-to-choose-documentation-ai-for-a-clinical-practice). If you want to see the drafting before you build a test case, [it's here](/l/clinical-report-generator).

One synthetic case, four measurements, a threshold set in advance. That's the whole protocol.

— Ian


Trialling an AI Report Tool Without Risking a Client File — What I'd Do


Choosing documentation AI for a clinical practice comes down to six questions: who stays the author, where the data actually goes, whether the output matches your structure, what it costs at your real volume, how hard it is to leave, and what the vendor refuses to claim. Everything else in a demo is decoration.

I build one of these tools, so read this with that in mind. I've deliberately written it without naming any product, including ours, and I've included the criteria where we'd come out badly — because a buyer's guide that only contains the tests its author passes isn't a buyer's guide.

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

*Most of what separates these tools isn't visible in a demo. It shows up in month two.*

## Who stays the author?

This is the first question and it's close to disqualifying. Every psychologist I've spoken to draws the same line: the clinical judgment in a document has to be theirs, and their name on it has to mean something.

So look at what the tool does at the point of judgment. Does it produce a draft you edit and sign, or does it produce conclusions? Does it ever fill in an interpretation you didn't make? Can you tell, looking at the output, which parts came from you?

The Canadian Psychological Association's guidance on [artificial intelligence and psychology](https://cpa.ca/docs/File/CPD/Artificial%20Intelligence%20and%20Psychology%20EN%202024.pdf) is worth reading before any demo, precisely because it frames the professional-responsibility side rather than the feature side. And "you review and sign" is a phrase every vendor uses, so press on what it means operationally: reviewing a draft you can actually check against the source is a different act from approving prose you'd have to re-derive.

## Where does the data actually go?

Not "is it secure." Which jurisdictions, at which step, and for how long.

Ask for the path: where it's processed, where it's stored, whether any step leaves the country, what's retained, for how long, and whether anything is used to train or improve the product. Get the retention figure as a number and ask where that number is written down.

The reason this matters more than a security badge is that the obligation doesn't transfer, and it's worth reading [the plain guide to what PIPEDA actually says](/articles/pipeda-for-clinicians-plain-guide) before you take any vendor's word on this. The Office of the Privacy Commissioner [states plainly](https://www.priv.gc.ca/en/privacy-topics/airports-and-borders/gl_dab_090127/) that "an organization is responsible for personal information in its possession or custody, including information that has been transferred to a third party for processing." You are choosing on your own behalf, permanently.

Two specific things to be sceptical about. **De-identification is a technique, not a guarantee** — the privacy literature has been [blunt about the limits of anonymisation](https://iapp.org/news/a/the-myth-of-anonymization-why-ai-needs-a-new-privacy-paradigm), and a vendor who presents it as a guarantee is either not reading carefully or hoping you aren't. And **a Canadian company is not automatically Canadian hosting** — ownership, incorporation and where the servers sit are three separate claims.

If you're in Ontario, the Information and Privacy Commissioner's material on [AI scribes in the health sector](https://www.ipc.on.ca/en/resources/ai-scribes-key-considerations-health-sector) is the most directly relevant public document I know of, and it reads like a list of questions to bring.

## Does the output match how you already work?

This is the criterion most demos are designed to obscure, and the one that decides whether the tool survives contact with your practice.

Bring your own material. Ask to run the tool against the kind of document you actually produce — your section order, your length, your phrasing — rather than the vendor's sample. Then read the output as though you were about to sign it, and count the edits.

The number that matters isn't "does it produce something." It's **how long from draft to signed**, compared with your current baseline. If a vendor won't let you test that with your own structure before you buy, that's an answer.

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

*Test it on your own documents, at your own length. A vendor's sample is chosen to flatter the vendor.*

## What does it cost at your actual volume?

Pricing models behave very differently as volume grows. Per-seat is predictable and punishes small part-time teams. Per-document rewards low volume and gets expensive exactly when the tool is working. Per-minute makes long assessments cost more than short sessions.

Work out your real annual volume and price all the models against it, not against the headline. Then ask the questions the pricing page doesn't answer: what happens in a month where you do double, is there a minimum, what does a second reader or an administrator seat cost, and does anything change at renewal.

## How hard is it to leave?

Ask on day one, not on the day you want out. Can you export your documents in a format that's useful somewhere else? Do you get the underlying content or only rendered files? How long do they keep things after you cancel, and can you get that in writing?

A vendor with a clean answer here is telling you they expect to be kept for good reasons.

## What won't the vendor claim?

I've come to think this is the most diagnostic question in the whole list, and it's the one nobody asks.

Ask directly: what have you *not* certified? What does your tool do badly? Where would you tell me not to use it? A vendor who can answer those crisply is a vendor who knows their own product. A vendor for whom everything is a strength is describing a brochure.

Here's ours, so the question is fair. Our documentation runs on edge infrastructure in Montreal, with a single step in report drafting where content transits a large-model provider outside Canada — I'd rather write that down than let a maple leaf on a marketing page imply otherwise. And we don't name competitors, which means this guide can't tell you we beat them; it can only tell you what to measure.

## Putting it together

A short, honest scorecard: authorship, data path, structural fit measured on your own documents, cost at your real volume, exit terms, and the vendor's own stated limits. Score every tool on the same six and the field usually sorts itself out quickly.

If you want to see how one of them answers, our [report drafting keeps the clinician as the author who reviews and signs](/l/clinical-report-generator). But run your own six-question scorecard first — including on us. I'd rather be chosen by someone who tested than persuaded by a demo.

— Ian


How to Choose Documentation AI for a Clinical Practice


AI drafting fits into a psychologist's assessment workflow in one specific place: after you've tested, scored, and interpreted, and before you finalize. It assembles a first draft in your structure, so your hours go to reviewing and refining instead of retyping. The interpretation and the signature stay entirely with you.

I'm not a clinician, so I won't pretend to tell you how to run an assessment. But I've spent the past year listening to psychologists describe where their time actually goes, and the pattern is consistent enough that I can at least map where a drafting tool does and doesn't belong.

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

*The thinking happens across the whole file. The retyping is what a draft can lift off you.*

## Where does AI drafting fit into a psychologist's assessment workflow?

It fits at the assembly step, not the thinking step. A typical assessment moves through a familiar arc: intake and history, testing, scoring, interpretation, and then writing it all up into a clear, defensible report. Drafting slots in near the end of that arc, once the clinical picture is yours and settled.

That placement is deliberate. The tool isn't there to decide what the results mean. It's there to take the structure and findings you've already worked out and hand you a first draft to work from, so you're editing prose instead of building it from a blank page. If you want a refresher on that structure, I wrote about [how to structure a psychoeducational report](/articles/how-to-structure-a-psychoeducational-report) separately.

## What stays the psychologist's job, start to finish?

All of the clinical work: the interpretation, the judgment, the conclusions, and the final sign-off. Nothing about a drafting step changes who is responsible for the report. You choose what the findings mean, which recommendations follow, and whether a sentence belongs in the document at all.

That's the part I'm careful never to overstate. The thinking is the thing that makes the report *yours*, and it isn't the part software can or should do. What software can do is give back the mechanical hours: the reformatting, the retyping of scores into sentences, the wrestling with a blank template at the end of a long week.

## What does the drafting step actually replace?

It replaces the assembly, not the analysis. The time cost of report writing is famously lopsided toward mechanics. In a [survey of 93 health-service psychologists](https://pubmed.ncbi.nlm.nih.gov/31728882/), most described working at or beyond capacity, with non-billable clinical work among their most common sources of stress — and clinicians tell me a single assessment report can swallow ten to fifteen hours on its own.

Very little of that is the interpretation, which you'd finished in your head before you opened the document. Most of it is turning a settled clinical picture into organized, readable prose. That's the specific chunk a first draft gives back.

![Two professionals reviewing a clipboard together in a modern office](/images/article-body/pexels-thirdman-7652132.jpg)

*A draft is a starting point for your review, not a substitute for it.*

## Does a draft box you into someone else's format?

No, and this is the fear I hear most, so it's worth answering directly. A useful draft comes back in *your* structure and voice, following the way you already organize a report, not a generic house style you then have to fight. The point is to save the retyping, not to impose a template.

If it came back sounding like a stranger wrote it, you'd spend your saved time undoing it, which defeats the purpose. I dug into that specific worry in [whether an AI-drafted report can sound like you](/articles/will-an-ai-drafted-report-sound-like-me), because it's the difference between a tool that helps and one that just adds a rewrite.

## How do you keep judgment in control of the final report?

By treating the draft as raw material for your review, exactly the way you'd treat a first pass you'd typed yourself. You read every line, you edit what doesn't fit, you add what's missing, and you sign only what you'd stand behind. The workflow ends where it always did, with your review and your name on it.

That's the whole design principle worth insisting on: the draft accelerates the start, and you own the finish. A tool that asks you to trust output unread is doing it wrong. A tool that hands you a strong first draft and then gets out of your way is doing exactly what the mechanical part of the job needs. That's the shape of it when Soma [drafts a report you review and sign](/l/clinical-report-generator).

Thanks for reading. The goal was never to take the clinician out of the report, it was to give the clinician back the hours that were never really about clinical work, and I'm grateful to the psychologists who keep helping me draw that line in the right place.

— Ian


How AI Drafting Fits Into a Psychologist's Assessment Workflow


An AI scribe for therapists transcribes your session so it can help draft your clinical notes — in your own voice, not a generic template. It doesn't replace your clinical judgment, and nothing from your session is retained that you haven't approved. You review every word. You sign off. Privacy is where the whole thing starts, not an afterthought.

I'm Ian — I founded Soma, and I want to be plain about what an AI scribe is and isn't. There's a lot of noise in this space, and clinicians have every right to be skeptical. So let me walk through it the way I'd want it explained to me.

## What is an AI scribe for therapists?

An AI scribe is software that listens during your session, transcribes what's said, and uses that transcription to help you draft a note afterward. That's the whole job.

Think of it like having a careful typist who hands you a first draft — one you then shape into the note you'd have written anyway. The clinician stays in charge of every clinical decision. The software just clears the blank page.

The reason this matters is simple: documentation is the near-universal #1 complaint we hear from the clinicians we work with. Notes pile up and follow people into their evenings. One survey of health-service psychologists found most were working at or beyond capacity — 41% "at full capacity," 39% "over-extended" — with [non-billable clinical work among their most commonly named sources of work-related stress](https://pubmed.ncbi.nlm.nih.gov/31728882/).

A good AI scribe gives some of that time back. That's the outcome we're after — not a flashy demo, just fewer late nights finishing notes. I've written about [how therapists actually cut documentation time](/articles/how-therapists-cut-documentation-time), and about [why purpose-built tools beat pasting into ChatGPT](/articles/2025-10-06_ChatGPT_and_Therapy).

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

*A scribe hands you a first draft to shape — it never makes the clinical call.*

## Does an AI scribe write my notes for me?

No — it drafts, you decide. This is the distinction I care most about, so I'll be blunt about it.

The software produces a starting draft from the transcription. You read it, correct it, add the clinical reasoning only you can add, and you sign off. Nothing leaves your hands without your review. The note is yours, and the responsibility for it stays yours.

I think the word "scribe" actually helps here. A scribe writes down what's happening so the expert in the room can think. A scribe doesn't make the call. The judgment — what's clinically significant, what to flag, what the plan is — that's you, and it should always be you.

We built Soma's [notes side and reports side](/product) around that idea. The draft is a convenience, never an authority. If a tool ever asks you to trust its clinical conclusions, that's the moment to walk away.

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

*Privacy as the starting posture, not a feature bolted on later.*

## Is it safe for client privacy?

Privacy has to be the starting posture, not a feature you bolt on later. For us that means a few concrete things, and I'd hold any tool you evaluate to the same bar.

First — and this is critical — a responsible scribe transcribes in the moment. The valuable thing is the transcription that helps you draft, and nothing from the session is retained beyond what you approve.

Second, de-identified data matters. When clinicians set Soma up, what they share with us is de-identified — no client's name or identifying details. Stripping identifiers out of clinical text reliably is genuinely hard; [research on automated de-identification of clinical notes shows it's a real technical problem](https://www.nature.com/articles/s41598-024-81170-y), not something to wave away. Treating it seriously is the whole point.

Third, consent is non-negotiable. Your client should know a scribe is being used and agree to it. More on how I'd introduce that below.

And one honest note: some employers and clinics restrict AI tools, sometimes for good reasons. If you're not solo, check your organization's policy before you bring any scribe into a session. I'd rather you ask first than assume.

![A hand writing notes in a spiral notepad with a pen](/images/article-body/pexels-artempodrez-6779632.jpg)

*A draft should follow your structure and phrasing — the way you'd write it yourself.*

## Will the notes sound like me?

They should — that's the entire point, and it's the thing clinicians ask us about most. Nobody wants their charts to read like they came off an assembly line.

The therapists, counsellors, and psychologists we work with don't want a generic template. They want notes that sound like *them* — their structure, their phrasing, the way they'd actually describe a session. A note that reads like someone else's is almost worse than a blank page, because now you have to unpick it.

So the goal we hold ourselves to is a draft that reads like *yours*. The draft follows your shape, not ours — so it reads the way you actually write, not some anonymous house style. I won't get into the mechanics here; what matters is the outcome. If the draft doesn't sound like you, the tool hasn't done its job, and you should expect better.

Your voice is part of your clinical work. It carries your reasoning and your relationship with the client. An AI scribe that flattens that into boilerplate isn't saving you time — it's just handing you editing work in a stranger's words.

## How do I introduce it responsibly?

Start with consent, keep your judgment, and review everything before it's final. Here's the simple way I'd approach it.

Tell your client plainly. Something like: "I use a software tool that helps me transcribe our session so I can write my notes faster and stay more present with you. Nothing is stored that I haven't reviewed myself. Is that alright with you?" Plain language, honest framing, their choice.

Then give it a small run before you lean on it. Try it on a few sessions, see whether the drafts actually read like yours, and check that the time saved is real for your workflow. No rush — you've got plenty of runway to decide if it fits.

Keep your review step sacred. The draft is a first pass. You read it, you correct it, you add what only you can add, and you sign off. If a note doesn't reflect your clinical thinking, it's not done — and the software should make that easy to fix, not hard.

Finally, mind your setting. Solo practice gives you more latitude; a clinic may have policies. Check first, and pick a tool that's transparent about how it handles data so you can answer questions if they come up.

## The short version

An AI scribe for therapists transcribes your session to help you draft notes in your own voice. It doesn't write your notes for you, it doesn't hold on to your session data, and it doesn't replace your judgment. Consent first, de-identified data, your review, your sign-off — that's the responsible shape of it.

We built Soma for the clinicians drowning in documentation, and I never forget that you're trusting us with the most careful work there is. Thank you for taking the time to think it through carefully — it means a great deal to us, and it's exactly the kind of skepticism this work deserves.

If you want to see how the notes side and the reports side actually feel in practice, give it a run and see if it sounds like you. Here's [a clinician-first alternative to ChatGPT for your notes](/l/chatgpt-for-therapists). I'd love to hear how it goes.

With gratitude,

— Ian


AI Scribe for Therapists: Transcription Done Right


## The Documentation Crisis

If you're a therapist in private practice, you already know the problem. The work that drew you to this profession -- sitting with clients, holding space, facilitating change -- is increasingly crowded out by documentation. Notes, treatment plans, progress updates, insurance forms. The administrative burden is real, and the research confirms it.

A 2024 Google Cloud/Harris Poll study found that administrative work contributes to burnout for 82% of clinicians. The National Council for Mental Wellbeing reported in 2023 that 93% of clinicians experienced burnout. And a 2022 MHCC study found that 61% of Canadian clinicians report burnout, with 50% intending to leave the profession.

The question isn't whether documentation is a problem. It's whether AI scribing actually helps -- and the 2025 research is now clear enough to answer.

![Abstract rendering of a double-helix strand built from coloured spheres](/images/article-body/pexels-googledeepmind-25626516.jpg)

*The 2025 research is now clear enough to answer whether AI scribing helps.*

## What the 2025 Research Shows

### 15,791 Hours Saved

The Permanente Medical Group, in partnership with the American Medical Association (AMA), published results in 2025 showing that their ambient AI scribe deployment saved 15,791 hours of documentation time. That's equivalent to 1,794 eight-hour workdays -- nearly five years of full-time documentation work eliminated.

While this study covered multiple medical specialties, the implications for clinical practice are direct. Therapy documentation involves the same core tasks: capturing session content, structuring it into clinical note formats, and ensuring accuracy and compliance.

### Burnout Reduction in 30 Days

A 2025 study published in PMC (PubMed Central) examined the impact of ambient AI scribes on clinician burnout. The results were striking: burnout rates dropped from 51.9% to 38.8% after just 30 days of AI scribe use.

A 13-percentage-point reduction in burnout in a single month is clinically significant. For context, most burnout interventions in healthcare show modest effects over much longer timeframes. The immediacy of this result suggests that documentation burden is a primary driver of burnout -- and that removing it produces rapid relief.

### High Clinician Satisfaction

The same PMC clinician study found that 84% of clinicians reported a positive effect on client communication when using AI scribes. Clinicians showed the highest adoption rates across all specialties studied.

This finding challenges the concern that AI scribes would make clinical encounters feel impersonal or technology-driven. Instead, clinicians reported that removing the documentation task allowed them to be more present with clients -- exactly the outcome that therapy demands.

### Time Savings Per Provider

A 2025 University of Wisconsin study found that AI scribes saved an average of 30 minutes per day per provider. For a therapist seeing 5-8 clients daily, that translates to 2.5 to 4 hours per week -- or 130 to 208 hours per year.

Tierney et al. (2024), published in NEJM Catalyst, reported approximately 50% reduction in documentation time across their study population. For a therapist spending 15-20 minutes per note, this means notes completed in 7-10 minutes.

![A laptop, notebook, glasses and coffee on a wooden desk by a window](/images/article-body/pexels-arina-krasnikova-5951553.jpg)

*Not all scribes are equal — format support, privacy, and workflow fit matter.*

## What Solo Practitioners Should Look For

The research validates AI scribing as a category. But not all AI scribes are created equal, and the differences matter significantly for private practice therapists.

### Clinical Note Format Support

General-purpose AI tools produce unstructured text. Clinical practice requires structured formats: SOAP (Subjective, Objective, Assessment, Plan), DAP (Data, Assessment, Plan), BIRP (Behavior, Intervention, Response, Plan), and progress notes. A tool that can't produce these formats creates more work, not less -- you'll spend time restructuring the output into the required format.

### Privacy and Compliance

This is the most critical differentiator. In Canada, PIPEDA governs the handling of personal health information. In provinces with additional legislation (like Ontario's PHIPA), requirements are even more stringent. Any AI tool that processes clinical session data must comply with these requirements.

Key questions to ask:
- Where is the data processed and stored?
- Does the platform maintain client anonymity?
- Is there an audit trail for data access?
- Does the tool have a BAA (Business Associate Agreement) or equivalent?
- Can data be exported and deleted on request?

### EHR Integration

Most therapists in private practice use an Electronic Health Record (EHR) system. An AI scribe that exists as a standalone tool creates a copy-paste workflow that adds friction. Look for tools designed as EHR overlays that work alongside your existing system rather than requiring a rip-and-replace migration.

### Full Workflow vs. Transcription Only

This is where the biggest differentiation exists. Most AI scribes on the market focus exclusively on transcription and note generation -- they transcribe what happens during a session and produce a draft note.

But documentation is only one part of the clinical workflow. Session preparation (reviewing between-session data before the client arrives) and follow-up (delivering homework, tracking completion, maintaining the therapeutic loop between sessions) are equally important and equally time-consuming.

A full-workflow tool supports prep, session documentation, and follow-up; a transcription-only tool touches one of the three. On the prep side, the most a tool should do is draft possible themes and suggest patterns worth a look from the week's data — suggestions the clinician confirms or discards. Reading the week clinically is the clinician's call, never the software's.

## Why Purpose-Built Tools Beat Generic AI

Many therapists have experimented with ChatGPT or similar general-purpose AI tools for documentation. The appeal is understandable: they're accessible, flexible, and already familiar.

But the limitations are significant:

- **No session context.** ChatGPT doesn't know your client, their history, or their treatment goals. Every note requires extensive prompting to produce clinically appropriate output.
- **No clinical formats.** You'll need to specify SOAP structure every time and hope the output is correctly organized.
- **No compliance.** Data entered into ChatGPT is processed on OpenAI's servers under OpenAI's terms of service -- not under PIPEDA or PHIPA requirements. There's no BAA, no client anonymity, and no audit trail.
- **No integration.** Every interaction is a manual copy-paste workflow.
- **No continuity.** ChatGPT doesn't remember previous sessions or track longitudinal patterns.

Purpose-built clinical AI tools address all of these: session context is maintained, clinical formats are built in, compliance is architected from the ground up, and the tool integrates into the clinical workflow rather than existing alongside it. The clinician's judgment and final sign-off remain with the clinician -- the software drafts, the professional decides.

![An hourglass with running sand against warm sunset light](/images/article-body/pexels-dominiquemel-37113105.jpg)

*Time recovered each week is where the economics turn favourable.*

## The ROI for Solo Practitioners

For a solo practitioner seeing 20 clients per week:
- At 15 minutes saved per note, that's 5 hours per week recovered
- At an average billing rate of $180/hour, those 5 hours represent $900/week in potential capacity
- Monthly cost of most AI scribe tools: $50-150
- Monthly value recovered: $3,600+
- ROI: 24x to 72x

Even conservatively -- counting only the time savings without considering the burnout reduction, improved clinical quality, and capacity to see additional clients -- the economics are clear.

## The Bottom Line

The 2025 research removes the guesswork. AI scribes save meaningful time, reduce burnout, and improve clinical communication. For solo practitioners, the remaining question isn't whether to adopt one, but which one to choose.

Look for clinical note format support, privacy compliance, EHR integration, and full-workflow capability. The tool should make you more present in session, not create a new set of technical tasks to manage. I compared [ChatGPT against purpose-built clinical AI](/articles/2025-10-06_ChatGPT_and_Therapy) and laid out [how therapists cut documentation time](/articles/how-therapists-cut-documentation-time) if you want to go further. And here's [a clinician-first alternative to ChatGPT for your notes](/l/chatgpt-for-therapists).

---

*References: Permanente Medical Group/AMA (2025); PMC Ambient AI Scribe Study (2025); PMC Clinician Communication Study (2025); University of Wisconsin (2025); Tierney et al. (2024), NEJM Catalyst; National Council for Mental Wellbeing (2023); Google Cloud/Harris Poll (2024); MHCC (2022).*


AI in Practice