The guidance you are reading is probably American
Search for how to build compliant healthcare AI and almost everything you find is written for HIPAA. We know, because we published a version of that article ourselves and left it up for six months. It described Business Associate Agreements, a 60-day breach clock, and a six-year log retention rule — none of which is Canadian law.
If you run a clinic in Ontario, HIPAA does not apply to you. PHIPA does. It is structured differently in ways that change the architecture, not just the paperwork.
This is what we learned building Loquent, our AI voice platform, for Canadian dental and medical clinics. It is not legal advice — we build software, we are not a law firm, and you should have a privacy lawyer review anything that touches patient data before it goes live.
What actually applies in Canada
There is no single national health privacy law. What applies depends on where the clinic is and what kind of data is involved.
| Law | Where | What it governs | |-----|-------|-----------------| | PHIPA | Ontario | Personal health information held by health information custodians | | PIPEDA | Federal | Commercial handling of personal information, where no substantially similar provincial law applies | | Law 25 | Quebec | Personal information, with its own consent and transfer-assessment rules | | PIPA | BC, Alberta | Personal information in the private sector | | HIA | Alberta | Health information specifically | | PHIA | Manitoba, Nova Scotia, NL, NB | Provincial health information statutes |
For a Toronto clinic, the working answer is PHIPA plus PIPEDA. For a clinic in Montreal, add Law 25 and its rules on transfers outside Quebec. A vendor who only says "HIPAA compliant" has not thought about any of this.
Three differences that change how you build
1. There is no BAA. There is a written ESP agreement, and its contents are specified.
The HIPAA world runs on Business Associate Agreements. Canada does not have those. Under PHIPA's regulation, a custodian using an electronic service provider to handle personal health information needs a written agreement with that provider — and the regulation sets out what it has to cover: the permitted uses of the information, the provider's security obligations, breach notification procedures, what happens to the data when the contract ends, and the custodian's right to audit.
That last one is the item most AI vendors are not ready for. If a clinic has a contractual right to audit your handling of their patient data, your system needs to be able to answer an audit. That is an architecture requirement, not a legal one, and it has to exist before you sign.
The accountability point matters too: the custodian stays accountable for the information even when a service provider is handling it. You are not taking liability off their hands by being good at security. You are helping them discharge a duty that remains theirs.
2. There is no 60-day breach clock. There is "first reasonable opportunity" — and an annual report.
HIPAA gives you 60 days to notify. PHIPA does not set a day count at all. The standard is notification to the affected individual at the first reasonable opportunity, and Ontario's Information and Privacy Commissioner has made clear through its decisions that delays measured in months are not acceptable. Vaguer, and in practice tighter.
Custodians must also notify the IPC directly in defined circumstances. And there is a requirement with no HIPAA equivalent at all: health information custodians must file annual breach statistics with the IPC by March 1 covering the previous calendar year.
For an AI system, that has a concrete consequence. Your logging has to support a count — how many incidents, of what type, affecting how many individuals — produced on a fixed annual deadline, not reconstructed by hand in February. If your audit trail cannot answer that question quickly, you have built a system your customer cannot report from.
3. PHIPA does not mandate Canadian data residency. Everyone acts like it does.
This is the one most often stated wrongly, in both directions. PHIPA contains no explicit prohibition on storing personal health information outside Canada. Vendors claiming residency is legally required are overstating it.
But the accountability obligation does not travel. The custodian remains responsible wherever the data sits, which means cross-border storage adds a due-diligence burden, a lawful-access exposure, and paperwork that Canadian storage simply removes. Quebec's Law 25 goes further and requires an assessment before transferring personal information outside the province.
So: not a legal mandate, a very strong practical default. We keep Loquent's data in Canada because the alternative means explaining foreign lawful-access regimes to every clinic that asks, and they all ask.
The LLM provider question, asked correctly
In the HIPAA framing the question is "does this provider sign a BAA?" That question does not map cleanly onto Canadian law. The questions that actually matter:
- Where does inference physically run? Not where the company is headquartered — where the request is processed. Canadian or Canada-available regions exist on the major clouds; the default region usually is not one.
- Is there a written data processing agreement covering the items PHIPA's regulation requires, including deletion at contract end and a right to audit?
- Is there a contractual commitment not to train on your data, and does it cover the specific tier and endpoint you are using rather than the enterprise plan you are not on?
- What is the retention period on the provider's side, including abuse-monitoring copies that are often retained separately from your account data?
- Is audio treated differently from text? For a voice agent this is the one people miss. Transcription frequently happens on a different service, in a different region, under a different retention policy than the language model. Two of the vendors we surveyed for our AI receptionist comparison publish a residency claim that covers storage but says nothing about where processing occurs.
Architecture that holds up
Most of the technical practice is the same as it would be under any regime. It is the emphasis that shifts.
Map the data flow before writing code. Every place personal health information travels, is processed, and comes to rest. For a voice agent:
Patient call
↓
[Telephony] — encrypted, written agreement, Canadian region
↓
[Speech-to-text] — check the region separately, it is often not the same
↓
[Agent logic] — minimal PHI in prompts, no training on patient data
↓
[Action layer] — writes to the PMS or EMR over an authenticated API
↓
[Audit trail] — every access, queryable, retained to the custodian's schedule
Doing this exercise honestly surfaces most compliance gaps before they cost anything.
Send the model as little as possible. "Patient #12345 wants to reschedule Thursday" carries the same instruction as a full name and date of birth, and far less risk. Use identifiers, resolve them in your own database, and let function calling do the work against records the model never sees.
Encrypt in transit and at rest. TLS 1.2 or better, AES-256 at rest. Unremarkable, and still worth verifying rather than assuming.
Scope access properly. Role-based access, service accounts holding the minimum permission that works, credential rotation, MFA on any human-facing dashboard.
Build the audit trail as a product feature, not a log file. Who accessed what, when, what action followed, retained on the custodian's schedule rather than yours — retention periods come from provincial law and professional college requirements, and they vary, so the system has to be configurable rather than hardcoded to a number you read in an American article. Keep personal health information out of general application logs entirely.
Tell patients they are talking to an AI. Disclosure has never cost us a deployment. Patients care that the phone gets answered.
Canadian EMR and PMS integration
An AI agent that cannot write into the system the clinic actually runs is a demo. The systems that come up in Canadian practice:
- Oscar Pro, Juno, Accuro, TELUS PS Suite, Med Access — medical
- Curve Dental, Dentrix, ClearDent, AbelDent, Tracker, Open Dental — dental
- Jane App — allied health, and increasingly dental
Requirements are the boring ones: authenticated APIs with rotating credentials, validation at every boundary, sensible behaviour when the upstream system is down, and a defined answer for what happens when a record changes while the agent is mid-conversation.
Ask any vendor whether their integration writes appointments or only reads availability. The gap between those two is most of the value, and marketing language rarely distinguishes them.
Before it goes live
- Written agreement with every service provider touching personal health information, covering permitted use, security, breach notification, end-of-contract deletion, and audit rights
- Data residency decided deliberately, and documented — including where audio is processed, not just stored
- Personal health information encrypted at rest and in transit
- Role-based access, minimum-permission service accounts, MFA on dashboards
- Audit trail that can produce an incident count for the March 1 IPC filing
- Personal health information minimised in model prompts, and a no-training commitment in writing
- Retention and disposal configurable to the custodian's schedule
- Breach response plan written down, with "first reasonable opportunity" as the standard
- Patient-facing AI disclosure
- Escalation path to a human, and a measured rate for how often it fires
What we got wrong
We shipped Loquent to production in under eight weeks and it now handles thousands of clinic calls a month. The compliance work was not the hard part. Getting the framing right was.
The article this one replaces was written to HIPAA because that is what the available material described, and it sat on our site describing American law to Canadian clinics for six months. Meanwhile every Canadian competitor that started outranking us on healthcare queries led with PHIPA and PIPEDA by name. That was not a coincidence and it was not sophisticated — they simply named the statute their customers are actually subject to.
If you are building healthcare AI in Canada, the useful version of the lesson is narrower than "compliance is an architecture decision," true as that is. It is: the jurisdiction you write for is itself an architecture decision. Choose it before you choose a cloud region, and definitely before you write the marketing.
We publish what thirteen Canadian voice AI vendors say about PHIPA, PIPEDA and data residency, including where we come off worse than they do, in our AI receptionist comparison for Canadian clinics.
Building something similar? Tell us what you are working on.