The pattern is remarkably consistent. A clinic hears about AI scribes, tries a demo, gets genuinely excited — the thing works — and then stalls for a year. Not out of ignorance. Out of three unanswered questions that feel bigger than they are: is the privacy clean enough for BC, which one actually fits us, and how do we roll it out without breaking the front desk?
Having helped stand up AI clinical documentation across a real health system, here is how we'd compress that year into a few weeks.
The privacy questions, in order
Marketing pages say “secure and compliant.” The answers you actually need are specific, and any vendor worth using can produce them in writing:
Where does the audio and transcript live? Which jurisdiction, which provider, how long is it retained, and can retention be set to zero after the note is produced?
Is your patients' data used to train anything? The answer must be no, in the contract — not in a settings toggle someone can miss.
Who is accountable under PIPA?BC's private-sector privacy law makes your clinic responsible for the personal information it hands to a service provider. You need the vendor's privacy terms to be signable by you, reviewable by counsel, and consistent with the conservative reading of PIPA and PIPEDA on cross-border processing.
What do patients get told?The College's guidance on AI in care points the same direction as common sense: be transparent. A one-line consent script at the start of the visit, and a poster in the waiting room, cost nothing and close the loop.
Fit is about your clinic, not the leaderboard
The “best” scribe is a category error. The right one depends on how your clinicians actually work: visit types and specialty mix, accents and languages in the room, whether notes are built on templates, how the note gets from draft into the EMR, and what it costs per clinician once the intro pricing ends. Two or three products, trialled on real clinic days against those criteria, separate quickly. The trial matters more than the shortlist.
The rollout is where the value is won or lost
A scribe changes the shape of the note, and the note is the input to half the back office. Templates need rebuilding to match the new dictation flow. The MOA loop — who checks what, who files what, what gets flagged back — needs one deliberate correction, or the time the physician saves reappears at the front desk as confusion.
And it needs a number. Pick one — time-to-completed-chart is the honest one — measure it before, measure it at 60 days, and let the number decide whether the tool stays. Adoption theatre is common in healthcare AI; a baseline and a re-measure make it impossible.
The stall is rational. Staying stalled isn't.
Every question above has a knowable answer. None takes a year. A clinic that assigns the work — to an owner inside the practice, or to someone like us — gets from “interested” to “deployed, measured, and keeping Sundays” in under a month. The charts finished at the kitchen table were never the cost of doing medicine. They were the cost of an unanswered question.
— Himanshu, for TOSC · July 2026