You can ask an assistant that has read the whole case file anything the file itself can answer — where a finding appears, what each informant said about the same behaviour, what changed between two assessments — and it answers with the document and page number so you can check it. You can also ask it to draft from the file: a summary of thirty documents, a timeline of the history, or a letter to the school, which you then edit and sign.
That's the short version. The longer version is about why this turned out to matter more than I expected when we built it, and where I think the line sits.
I'm not a clinician, so I won't tell you what to make of anything in a file. What I can describe is what psychologists ask once the reading is done, because I've watched a lot of them do it.
The report is one thing you make from the file. It isn't the only thing you need from it.
Why would you want to ask a file a question at all?
Because the file is bigger than anyone's memory of it. A full assessment file is rarely one document. It's intake forms, school records, teacher and parent questionnaires, prior reports, score reports, collateral letters and your own interview notes — often dozens of documents and hundreds of pages. Soma reads every document in the file for that reason.
When the report is due, most of the time isn't spent typing. It's spent finding things: going back through a stack to confirm who said what, and when. That retrieval is a big part of why psychological report writing takes so long, and it's invisible on any timesheet.
A clinical psychologist described it to me better than I can:
"I have notebooks stacked that I would have to spend hours reading through for an insurance claim or a longer term report. Soma allows for me to quickly get that data in front of me."
"Get that data in front of me" is the job. Not deciding what it means. Putting it where you can see it.
What kinds of questions does the file actually answer?
The useful questions are the ones you'd otherwise answer by flipping pages. A few patterns come up again and again:
- Where is it? "What did the cognitive testing say about working memory?" "Find every mention of the 2019 assessment."
- Who said what? "What did the teacher report about homework, and how does it compare with what the parent reported?"
- What changed? "Compare the scores from the previous report with this one." "What's different in the school history since last year?"
- What happened when? "Build a timeline of the developmental and school history from these documents."
- What's in this pile? "Summarize the collateral documents in a page." "What does the referral letter actually ask?"
Every one of those has an answer sitting somewhere in the file. The assistant's work is to read, retrieve and lay it out. Yours is to decide whether it matters.
Why does the page citation matter more than the answer?
Because an answer you can't check isn't worth much with your name on the report.
When the assistant answers, it points to the document and the page it drew from. That changes what you're doing. You aren't trusting a fluent paragraph; you're opening page 47 of the school records and confirming the sentence is there. A checked answer takes seconds. An unchecked one is a risk you've taken on without noticing.
I've come to think this is the most important design choice in the whole chat feature, more than how good the answers sound. A tool that answers confidently without telling you where it looked is asking for a kind of trust that a careful clinician shouldn't give. I've written more about that line in why the clinician stays the author.
What can it draft from the file, not just answer?
This is where it stops being a search box. Once the case has been read, you can ask for a document, not just a fact:
- A records-review summary of everything that came in, so you start the report knowing what's there.
- A response to the referral source, drafted from what the file contains.
- A letter to the school drafted from the finished report, so the people acting on it get the part they need in plain words.
- Answers to a referral or intake form, drafted from the file for you to correct.
Drafts open beside the conversation. You edit them inline, download them as a Word file, and nothing is saved back to the case unless you choose to. You don't need to start a report to do any of this — some psychologists use the chat only on the documents, as a way into a new file.
And once a report exists, the same conversation works on it. "Tighten the recommendations." "Move the test scores up in the results section." "Pull Tuesday's session notes into the background." Each change shows up in the document, marked as updated from chat, so nothing changes quietly.
Ask in plain words. Check the page it points to. Keep what holds up.
What will it not do?
Three boundaries, all deliberate.
It doesn't go to the internet. Web search is switched off by design. The assistant answers from the documents in the case, and that's all. If the file doesn't say it, the honest answer is that the file doesn't say it.
It doesn't make the clinical call. It can lay out what each informant reported and where they differ. It doesn't settle what the difference means, and a draft can suggest a wording without settling one. Every psychologist I've worked with draws that line in nearly the same place, and so do we. Ontario's Standards of Professional Conduct say technology may not be used instead of a registrant's own professional knowledge, and that's the right frame for a tool like this.
It doesn't see who the client is. Identifiers are removed before any AI model sees the text, and you review and approve those redactions before the work starts. You set how long a case is kept, and one button deletes everything. I'd describe de-identification as a strong control, not a promise — I've explained why in why de-identification isn't a promise.
Why one assistant instead of a separate chat tool?
Because the chat is only useful if it has read the same file the report came from.
A lot of the psychologists I speak with juggle three things: a scribe for sessions, a general chat tool for quick questions, and whatever they use for reports. Three logins, three places their client material goes, and three different answers to "where does this data go?" The chat tool hasn't read anything, so every question starts with pasting.
One assistant that has read the case can answer the question, draft the letter and work on the report from the same set of documents, under the same privacy settings. Notes and chat are part of it, not something extra you buy. That's less clever than it sounds. It's mostly about not making you carry the file from one tool to the next.
Where does this fit in an assessment week?
It fits most naturally at the two ends of the work.
At the start, when the file lands, you can ask for the summary and the timeline so you know what you have before you plan the battery or the interview. At the end, after you've reviewed and signed the report, you can ask for the school letter or the referral response, so the people acting on it get something they can use without reading forty pages.
In between, it's the quick question that would otherwise mean ten minutes of flipping: "Did the parent mention sleep anywhere?" Small, but those add up across a file.
If you'd like to see it on a case, the best way is to bring one of your own past reports, identifiers removed, and we'll show you the draft and the questions it can answer about the file behind it. You can book that walkthrough here. You'll review and sign everything, as always.
Thank you to the psychologists who showed me how much of this job is finding things.
— Ian
