Canada's privacy regulators published ten expectations for health care providers that adopt AI. Here is what each means for a clinic with one to ten providers.
The short answer: Canada's privacy commissioners expect clinics to govern AI like any other handling of patient information: explain it plainly, offer an opt-out where practicable, collect only what is needed, vet and contract vendors carefully, train staff, and keep a person checking outputs. Their 15 September 2026 resolution lists ten such expectations. It is not law, but PHIPA and PIPEDA still are.
On 29 October 2026, Ontario's Information and Privacy Commissioner announced that Canada's federal, provincial and territorial privacy regulators had "issued a resolution calling on their respective governments and interested parties in the public and private sectors to take a responsible approach to the development, procurement, and adoption of artificial intelligence systems (AI systems) in health care and in health research." It was passed at the regulators' annual meeting, hosted in Ottawa. (IPC Ontario news release, 29 October 2026)
The full text, "Responsible use of artificial intelligence in health and health research", was adopted in Ottawa on 15 September 2026. It says AI systems in health include "generative AI tools, such as chatbots and AI scribes." It urges governments to legislate so that AI systems meet six core principles: privacy protective, valid and reliable, safe, human rights affirming, transparent, and accountable. (resolution on the Privacy Commissioner of Canada's website)
Ontario's Information and Privacy Commissioner, Patricia Kosseim, put the tension plainly: "AI is rapidly changing the health care landscape. While it has potential to reduce administrative burdens so that physicians can place a greater emphasis on care, it comes with increased risks such as the unauthorized collection, use, and disclosure of personal health information and potential built-in biases that disproportionately affect marginalized and underrepresented groups."
For a clinic owner, the useful part is a list of ten things that providers who procure, adopt and use AI systems "should" do. The same day, the Government of Canada committed to an equity investment of up to $1.6 billion in Cohere, a reminder that Ottawa is pushing Canadian-built AI while regulators push for safeguards (Prime Minister of Canada, 29 October 2026).
No. A joint resolution of privacy regulators is not a statute and creates no new penalties. It asks governments to act, including to "legally require all relevant parties who develop, procure, and use AI systems in health care to establish effective AI governance and accountability frameworks."
What binds your clinic today is existing law. In Ontario, that is PHIPA for health information custodians and the agents and service providers who act for them. Elsewhere it is PIPEDA or the relevant provincial law. Read the resolution as a preview of how regulators will think when a complaint or breach arrives. A clinic that can show it followed these ten expectations is in a stronger position than one that cannot.
This is general information, not legal advice.
The resolution is written for hospitals and researchers as much as for a three-chair dental office. Here is each expectation, paraphrased, translated for a clinic with one to ten providers, with a question for any AI vendor.
| Provider expectation (paraphrased) | What it means for a 1–10 provider clinic | What to ask your AI vendor |
|---|---|---|
| 1. Governance and policies follow the privacy law that applies | Write a one-page AI policy on what is collected, used, disclosed and retained. The clinic stays accountable for its vendors. | Which privacy laws do you build to, and how does your service fit them? |
| 2. Clear communication, and an opt-out where practicable | Tell patients before the tool touches their information. For calls, say so at the start and offer a person. | Can the opt-out route to our front desk, and is it logged? |
| 3. A designated contact person, and a way to challenge | Name one staff member for AI concerns and publish how to reach them. | How do you help us find, correct or delete an individual's record? |
| 4. Limit collection, use, disclosure and retention | Decide what the tool needs. Set retention periods you could defend to a patient. | What is stored by default, for how long, and can we shorten it? |
| 5. Weigh privacy, security and AI assessments of the vendor | Ask for the vendor's assessments before signing. Let known risks, such as inaccuracy or bias, change the decision. | Which security and privacy assessments exist, and what did they find? |
| 6. A risk management framework, kept up to date | Keep a short risk register and review it when the tool changes. A privacy impact assessment is a natural anchor. | How do you tell us when models, subprocessors or features change? |
| 7. Utmost diligence in contracting and sub-contracting | Contracts should cover training, record keeping, auditing and limits on use, and bar unauthorized commercialization. | Who are your subprocessors, what does each see, and what do the contracts say? |
| 8. Plain-language explanations, including access rights | Put a readable notice on your website and at the front desk, including how patients ask for their information. | Can you supply editable wording and a way to export a patient's data? |
| 9. Train staff and prevent unapproved “shadow AI” | Approve tools by name and train everyone who uses them. | What training material and admin controls do you provide? |
| 10. Human oversight and verification of outputs | A person reviews anything consequential before it reaches a chart or changes care. | Where does a human review the output, and what does the audit trail show? |
The resolution says developers and service providers should take a privacy-by-design approach, including data minimization and the most protective privacy settings by default. It also says they should "provide clients with evidence of how the AI system is trained and a general explanation of how the AI system works," "demonstrate to the client that the use of the AI system complies with the relevant privacy law," and "share ongoing risk assessment findings with clients." Those lines are fair to quote back to a sales rep.
The resolution does not mention voice agents by name. But a receptionist that answers your phone hears names, health complaints and appointment reasons, so every expectation applies. These details are our interpretation, not text from the resolution.
Ask whether calls are recorded, whether transcripts are kept, and where each lives. Many clinics only need the booking outcome. If you keep recordings or transcripts, tie a retention period to a purpose and confirm the vendor enforces deletion.
Patients should know within seconds that they are speaking with an AI. A short opening works: who is answering, that it is an automated assistant, and that the caller can ask for a person at any time. If calls are recorded or transcribed, say that too.
Expectation two asks for an opt-out where practicable. On the phone, that means a clear route to the front desk, and a rule for when nobody is there: voicemail, a callback list or an on-call line. Test it yourself by calling and asking for a person in three different ways.
A voice agent is not one product, and the published number is only the front door. Behind one phone number there is typically a telephony carrier, speech-to-text, a language model, text-to-speech, an orchestration layer and a database. Expectation seven asks for utmost diligence, so ask for the full list, what each one sees, where it runs and what it retains.
Data minimisation applies to the conversation itself. A receptionist needs a name, a callback number, an appointment type and a preferred time, not a diagnosis. Tell the vendor what the agent may ask, and what it should do if a caller volunteers clinical detail: capture only what is needed and hand off to a person. Our guide to deploying AI agents safely in a Canadian business covers how to write these limits into the agent's scope.
AI scribes listen to a visit and draft a note, a more intimate tool than a phone agent. The resolution recalls existing guidance by name, including two Ontario documents, "Ontario's Principles for the responsible use of artificial intelligence" and the guidance "AI Scribes: Key Considerations for the Health Sector", British Columbia's "PIPA and AI scribes" best practices, and Alberta's AI Scribe privacy impact assessment guidance. In Ontario, read those IPC documents before trialling a scribe.
Three expectations weigh most for scribes: two, because patients should hear about the recording beforehand and have a real choice; five, because inaccuracy and bias are harms providers must weigh before adopting; and ten, because the clinician must verify the note before it enters the chart. Ask any scribe vendor where recordings go and when they are deleted.
Expectation nine asks providers to train staff and prevent use of AI without approval or oversight, which the resolution labels "shadow AI". In a small clinic it looks like a front-desk employee pasting a patient email into a free chatbot to tidy the wording. The information still leaves the clinic, and the clinic cannot answer a patient who asks where it went.
The fix is dull and effective. Publish a short list of approved tools. State what must never be entered into anything else. Name the person staff ask before trying something new, and answer quickly, because slow approvals push people toward shadow AI.
Maya is our AI voice receptionist for Ontario healthcare clinics, including dental and physiotherapy, and for other small businesses. We offer this as a description of how we work, not a certificate. No PHIPA certification exists.
Maya's stack uses Retell AI for orchestration, with a business associate agreement signed (a contract form borrowed from U.S. health privacy law), the Claude API for language, Deepgram Nova-3 for speech-to-text, Cartesia Sonic Turbo for text-to-speech, and Twilio or Vonage for telephony. Application data (call logs and bookings) is stored in Supabase's Canada region (ca-central-1), with Postgres row-level security separating each clinic's records. Mihron has conducted PHIPA compliance auditing across that vendor stack.
Several of those subprocessors, including Retell, Deepgram, Cartesia and Twilio, are not Canadian-only services, so we do not claim that all call data stays in Canada. The resolution does not address data residency either. Instead, we map the vendor chain subprocessor by subprocessor, which is expectation seven in practice. Decisions such as call recording, retention, the opening disclosure and the path to a human are settled with each clinic during scoping, because the clinic is accountable to its patients. See also our pages for AI receptionists for dental clinics, AI receptionists in Ontario and healthcare, and in our earlier post on AI receptionists in healthcare.
A few decisions, written down, go a long way.
For a second pair of hands, our AI Readiness Sprint is built for this mapping, and an AI implementation partner can keep the register current. More guides are on the Mihron blog.
The AI Readiness Sprint scopes your tools, vendors and patient-facing wording before anything goes live.