A psychologist using a scribe for session notes, a general chat tool for letters and summaries, and a separate report writer is running three privacy postures, not one. Each tool sends client text somewhere different, keeps it for a different length of time, and has its own answer on training, so every question a careful clinician asks about client data has to be asked and answered three times.
It's an easy setup to fall into, and it usually isn't a plan. A scribe arrives first because notes are the daily pain. A chat tool comes next because it's quick for a letter. A report tool comes last, because the long assessment is the hardest problem. Each choice was reasonable on its own. Together they make a privacy picture that's hard to hold in your head.
I'm not a clinician and I'm not a privacy lawyer, so I won't tell you what your obligations are. What I can describe is the practical shape of running several tools at once, because we designed Soma around avoiding it, and I'll be specific about our own data path so you can check it.
Three tools means three places client text can end up.
Why do so many psychologists end up with three AI tools?
Because the work has three different shapes, and most tools were built for one of them.
Session notes are short and frequent: a transcription of the session becomes a note in your format. Chat is ad hoc: summarize this, draft that letter, find where the teacher said something. Reports are long and rare: dozens of documents, a full battery, a 30- or 40-page result.
A tool built for notes isn't built to read a 184-page file, which is part of why scribe reviews tell you little about report writing. A general chat tool wasn't built around client files at all. So psychologists stack them, and the stack is where the privacy picture gets complicated.
What is a privacy posture, and why does each tool have its own?
By privacy posture I mean the full set of answers about what happens to client text once it leaves your hands:
- Where it's processed and stored, by country and by step.
- How long it's kept, and who decides.
- Whether it's used to train a model, by default and by setting.
- What reaches the AI: full text, or text with identifiers removed first.
- How it's deleted, and whether you can do it yourself.
- Who can get into the account: sign-in protection, sharing, staff access.
Every vendor answers those differently, and the answers live in different places: a privacy page, a settings screen, a terms document, a sales call. I went through how to get honest answers to the first of these in what to ask about where your data is processed. Multiply that by three and it's a real afternoon of reading, repeated every time a vendor updates their terms.
What does running three postures cost in practice?
Three costs show up in practice.
The review work triples. Each tool needs its own read of the privacy terms, its own settings check, and its own answer when a client, a clinic owner or a college asks how you use AI. This lands hardest on clinic owners, because every tool a team member adds is another set of answers someone has to stand behind.
Your protections are only as strong as the loosest tool. If one tool removes identifiers and one doesn't, the client file is exposed wherever the weaker one touches it. Careful settings in the report tool don't help if the same text is pasted, identifiers and all, into a chat tool with different defaults.
The copy-and-paste gap. This is the one I think about most. The moment the tools don't talk to each other, client text moves between them by hand: a paragraph from the report pasted into a chat window to draft a letter to the school, a session note pasted into another tool for a summary. Each paste is a copy you're now responsible for, sitting under another vendor's rules. Nobody plans that. It's just the shortest path at 9pm.
A registered psychologist told me what she cared about when she chose a tool:
"What matters most to me is that it doesn't retain my data, and that it's actually built for the complex, long assessments I do."
That's two requirements in one sentence, and with a three-tool setup they often end up met by different tools.
What does one privacy posture look like instead?
Here is ours, stated plainly enough to check. Soma does reports, session notes and chat in one place, under one set of answers:
- Identifiers are removed before any AI sees the text, and you review and approve the redactions. I describe that as a tool that lowers risk, not a promise, for reasons I set out in why de-identification isn't a promise.
- Data at rest is stored in Canada. Session transcription and document reading run in Montréal.
- Drafting can stay fully in Canada. On paid plans you choose this in Settings. The other option sends the drafting step, with identifiers already removed, to a model provider outside Canada that doesn't retain it or train on it. I'd rather you know that and choose than find it later.
- Client data is never used to train models.
- You set the retention window, you can see what's scheduled for deletion, and one button deletes everything.
- Sign-in supports multi-factor authentication, and data is encrypted in transit and at rest. Soma is PIPEDA compliant.
Because chat has already read the case, it answers questions with page citations back to your files, and the letter to the school is drafted from the report inside the same workspace. Nothing gets pasted anywhere. And if you only want one or two of the three jobs, you can switch Reports, Notes or Chat off in Settings and it disappears.
One workspace, one set of answers to check.
Is one tool always better than three?
No, and I'd be suspicious of a vendor who said it was.
If you only write session notes, a good notes tool may be all you need, and adding report features you'll never use isn't a win. If the tools you already have answer every question above in a way you're comfortable defending, consolidation is a convenience, not a necessity.
Where it matters is when the long reports and the daily notes and the ad hoc letters all touch the same clients. Then the number of postures you're responsible for is a real part of the decision. I put data paths near the top of what I'd demand from any AI report tool for the same reason.
Quick answers
What is a privacy posture? The full set of answers about what happens to client text: where it's processed, how long it's kept, whether it's trained on, what reaches the AI, and how it's deleted.
Is it a problem to use several AI tools? Not inherently. It means every privacy question has to be answered for each tool, and client text tends to get copied between them by hand.
Does Soma keep data in Canada? Data at rest is in Canada, and transcription and document reading run in Montréal. Drafting can stay fully in Canada on paid plans; you choose in Settings.
Can I use only part of Soma? Yes. Reports, Notes and Chat can each be switched off, and you pay for what you use.
If you'd like to see the whole posture on your own material, bring one of your past reports, identifiers removed, and see the draft Soma produces. I'm glad to walk through any of the answers above, line by line.
Thank you for asking these questions on your clients' behalf. It's the right instinct.
— Ian
