PHIPA and AI scribes: what an Ontario clinic has to have in place before the first recorded visit

AI scribes went from pilot to normal in Ontario in about eighteen months. The province now runs an Ontario AI Scribe Program, and Supply Ontario's Vendor of Record arrangement for AI scribes runs from April 27, 2025 to April 27, 2028, with more than two dozen pre-qualified vendors and a 2026 refresh open to new ones.

A vendor being on that list tells you the product met provincial requirements for clinical function, privacy and secure storage. It doesn't tell you that your clinic's use of it complies with PHIPA. That part is still yours.

The Information and Privacy Commissioner of Ontario (IPC) said so plainly in January 2026, when it released AI Scribes: Key Considerations for the Health Sector and a companion checklist. There's no new AI law here. It's PHIPA, applied to a microphone and a language model.

I spent about seven years building PHIPA-regulated clinical software, so this is how I read the guidance as an engineer, not a lawyer. Have your privacy officer or counsel confirm what applies to you.

The accountability doesn't move to the vendor

The IPC starts from PHIPA section 17: custodians are accountable for their agents and stay responsible for personal health information that agents and service providers handle for them. The vendor running the transcription model doesn't take that on. Neither does the provincial program.

The guidance says custodians should have an AI governance and accountability framework in place before implementing an AI scribe. For a ten-physician clinic, that doesn't need to be a committee with a charter. It needs to be a named person who decides which tools are approved, a written policy agents must follow, and a privacy impact assessment done before go-live and reviewed when the product changes.

Consent: get it, record it, and have a plan B

This is the section most clinics underestimate. The IPC notes there is currently no PHIPA provision that explicitly permits an AI scribe to collect, use or disclose personal health information without consent, so consent would generally be required. Its stated best practice is express consent before using the scribe, recorded in the patient's health record.

Consent has to be knowledgeable (PHIPA s. 18). In this context, the IPC says patients should know an AI scribe is being used and why, its benefits and risks, including accuracy and bias, and when their information goes to the vendor and on what authority.

Two practical consequences:

• Withdrawal has to actually stop the recording. When a patient withdraws consent, the custodian must make sure no further information is collected, used or disclosed that way. If your scribe auto-starts from the EMR schedule, that's a configuration problem, not a policy one.
• Patients who say no get the same care. The IPC says custodians must make sure people who withhold or withdraw consent get the same or a similar level of care, and should tell them about the alternative documentation method. You need a non-AI charting workflow that still works on a busy Tuesday.

The IPC also flags people who never consented at all: family members in the room, and neighbouring conversations in open-concept clinics such as some physiotherapy and chiropractic settings.

Nothing gets used until a human has reviewed it

The guidance says custodians must have policies to ensure every record created or altered by an AI scribe is reviewed for accuracy before it's used for any purpose or disclosed. As a best practice, review should happen before the note enters the EMR or EHR.

If a product can write to the chart before review, the IPC expects it to flag that content as AI-generated and unreviewed, and to prevent its disclosure. That's a requirement you can test in a demo: generate a note, don't sign it, and try to send it in a referral or show it in the patient portal.

Two more points engineers should build for. First, auto-translation doesn't shift responsibility, so translated notes need the same review. Second, the IPC says policies should require the custodian to stop using the system right away if performance falls outside an acceptable range. That means you need to define "acceptable" in advance and measure against it, for example a monthly sample of signed notes compared against the source transcript.

Read the vendor's data clauses before you sign

Many scribe vendors ask for the right to use recordings, transcripts or "de-identified" data to train and improve their models. The IPC is direct: custodians must not agree to disclose personal health information to a vendor unless PHIPA gives them authority to do so. Sharing recordings or transcripts for training would need patient consent, or data that is effectively de-identified.

It also warns that de-identification is harder than it looks. Removing names and health card numbers isn't enough. The IPC notes it may be virtually impossible to de-identify a voice recording given voiceprint and speaker-recognition technology. If you do share de-identified data, you have to say so in your public statement of information practices (PHIPA s. 16) and contractually require the vendor not to re-identify it.

On hosting: the IPC confirms PHIPA doesn't prevent storage outside Ontario or Canada, but custodians remain responsible and need to be satisfied the risks are appropriate, usually through strong contract terms.

Decide what you keep, and delete the rest properly

Audio, full transcript, draft summary and signed note are four different records. The IPC expects a defined retention period for any recordings and transcripts you keep, and secure disposal for anything you don't. Records you retain are subject to patients' access and correction rights, and patients should know that both the summary and the full transcript exist. Marking AI-generated or AI-altered records as such is called a best practice.

The breach that shows what goes wrong

In October 2025, the IPC published its response to a hospital breach (file HR24-00691). A physician who had left the hospital in June 2023 was never removed from a recurring virtual rounds invitation, which he had joined with a personal email address. In September 2024 he installed an AI transcription tool on a personal device. It joined the meeting from his calendar, recorded discussion about seven patients and emailed a summary to participants.

No vendor was hacked. The failures were offboarding, personal accounts and unapproved tools. The IPC's guidance now says custodians must make clear whether agents can use personally procured AI tools, and must monitor compliance.

What I'd check this month

1. List every AI scribe or transcription tool in use, approved or not, including meeting-bot integrations on Teams, Zoom and Google Meet.
2. Confirm consent is captured per visit in the chart and that withdrawal stops the recording.
3. Try to disclose an unsigned AI-generated note, and fix whatever lets you.
4. Read the vendor's terms for training, improvement and de-identified data rights, and check auto-update notice terms.
5. Set retention for audio and transcripts, and confirm deletion actually happens on the vendor side.
6. Audit leavers against recurring meeting invites.