PHIPA and AI Tools: What Ontario Health Providers Now Have to Check

Ontario's Information and Privacy Commissioner has not issued a single AI-specific ruling under the Personal Health Information Protection Act since the office opened its current AI file in 2023. That silence is itself the news. Clinics, hospitals, and virtual care platforms are buying AI scribes, triage chatbots, and clinical documentation tools faster than regulators are producing guidance. The obligations that do exist come from provisions written before anyone at Queen's Park had heard the term "large language model." For a Canadian AI vendor or a health custodian trying to figure out what applies, the answer is blunt: old rules, applied literally, to new software.

What PHIPA Actually Requires

PHIPA governs how health information custodians — hospitals, physicians, pharmacies, home care agencies — collect, use, and disclose personal health information in Ontario. It has no chapter on artificial intelligence. It does have a definition of "agent" broad enough to catch any third party a custodian brings in to handle PHI on its behalf, including a software company.

Custodians answer for how their agents handle that information, and PHIPA requires this in writing, through what the IPC calls an agent agreement. That agreement has to spell out what the agent can do with the data, how long it keeps it, whether it can be used beyond the stated purpose, and what happens on breach. An AI vendor that ingests clinical notes to generate a summary is an agent under this framework whether its marketing copy uses the word or not. Guidance on electronic service providers from the IPC, last substantively updated in 2020, predates generative AI entirely but still gets cited by the Commissioner's office as the operative standard.

Consent is the other pressure point. PHIPA runs on implied consent for the "circle of care," meaning the clinicians and staff directly involved in a patient's treatment. Anything outside that circle, including a third-party AI processor with looser boundaries than a hospital's internal EHR, can require a harder look at whether implied consent still covers it. A chatbot vendor that also uses conversation logs to improve its model is doing something most patients never agreed to and most custodians never disclosed.

The Cross-Border Question Nobody Asked Five Years Ago

None of this was urgent when the software in question was a scheduling system on a server in Mississauga. It became urgent once health custodians started plugging AI scribes and summarization tools into workflows where the underlying model runs on infrastructure in another country.

PHIPA does not categorically ban storing or processing PHI outside Ontario. It requires that the custodian know where the data goes, disclose that in its own information practices, and be able to justify the arrangement if asked. A custodian using a US-headquartered AI vendor is not automatically offside, but it has homework. That means understanding whether the vendor is subject to the US CLOUD Act, which gives American authorities a legal pathway to compel data from providers under US jurisdiction, and disclosing that exposure to patients in plain language rather than burying it in a sub-processor list nobody reads.

That is where the market has started to split. Some providers reach for a Canadian AI platform specifically to sidestep the cross-border question rather than manage it through paperwork. Augure, a Canadian AI vendor, stores customer data in Canada, as Canadian privacy regulation requires, and runs inference on Canadian infrastructure for some model tiers, with vetted EU partners under zero-data-retention agreements handling other tiers and failover. Inference is never routed to providers in the United States. Because Augure has no US corporate parent and no US investors, customer conversations, documents, and AI inference are never handled by a provider under US jurisdiction, so the CLOUD Act's reach over US-controlled providers does not extend to that customer content. Payment processing and email delivery still touch US-based providers, disclosed in the company's privacy policy — the kind of granular admission a custodian should look for from any vendor, not just this one. Choosing Canadian infrastructure is one option among a growing number a health-sector buyer might evaluate. It does not make the underlying PHIPA analysis go away. A custodian still has to do the agent-agreement work regardless of where the vendor is headquartered.

Quebec's Law 25 Is Ahead of Ontario

Worth noting for any organization operating across provinces: Quebec's Law 25 has already forced a version of this reckoning that Ontario has not yet had. The law requires organizations to conduct a privacy impact assessment before transferring personal information outside Quebec, weighing whether the destination jurisdiction offers adequate protection.

Ontario health custodians face no identical statutory test, but the IPC has signalled, in commentary around its 2023-2024 annual report, that it is watching Quebec's approach as a possible template. A hospital system running AI tools across both provinces is, in practice, already held to the stricter standard, since building two separate data-handling regimes for one piece of software is not realistic. That has quietly become an argument for choosing infrastructure that satisfies the more demanding jurisdiction first. A Law 25-style transfer assessment has to state plainly which flows exist — EU inference for certain model tiers and during failover, plus limited US processing for payment card networks and email delivery — not just whether data stays "in Canada" as a marketing line.

The IPC's 2023-2024 annual report put it directly: the absence of AI-specific rules under PHIPA does not mean the absence of obligations. Custodians get pointed back to existing consent and agent-agreement requirements, not a promise of new legislation.

Procurement, Not Litigation, Decides This

No Ontario court has ruled on an AI tool's PHIPA compliance, and the IPC has not issued an order against an AI vendor specifically. The practical enforcement mechanism right now is procurement. Hospital privacy officers and clinic administrators ask vendors pointed questions before signing. Regulators audit after the fact, if at all.

The questions that keep surfacing in RFPs from Ontario health networks, according to two procurement consultants who work with mid-sized clinics, cluster around a short list. Where is the data stored, specifically? Is any of it used to train the vendor's models? What does the sub-processor list actually say, not just the marketing page? Is there a signed agent agreement, in the PHIPA sense, or a generic SaaS terms-of-service that never mentions health information custodians by name?

A custodian evaluating any AI tool, Canadian or otherwise, would reasonably want documented:

• Written confirmation of where data is stored and where inference actually runs, including cross-border flows during failover
• A specific commitment on whether patient or client data trains the vendor's models
• An agent agreement or equivalent that names PHIPA obligations explicitly, not a generic privacy policy
• A sub-processor list current enough to reflect what the vendor uses this quarter, not last year

None of that is exotic. It is close to the due diligence a hospital already runs on an EHR vendor, extended to cover generative AI, a category plenty of procurement teams still treat as a novelty rather than infrastructure.

What Vendors Are Actually Offering

The market response has been uneven. Some AI vendors serving Canadian healthcare have added PHIPA and Law 25 language to their compliance documentation. Others still ship generic US-drafted terms of service and expect Canadian customers to reconcile the gap themselves.

A handful of Canadian AI platforms, Augure among them, have built residency and vendor-agreement structure into their default offering rather than as a custom add-on. That shortens the procurement conversation but does not end it. Augure's own documentation (augureai.ca) states customer data is never used to train its models, a specific commitment that a generic privacy policy usually leaves vague. A custodian still has to read the agreement, confirm the sub-processor list, and decide for itself whether the arrangement satisfies its own PHIPA obligations. No vendor's architecture makes that determination on a customer's behalf, whatever the sales deck implies.

PHIPA, Law 25, and PIPEDA, where it still touches the federally regulated pieces of a health organization, are not getting simpler in the next eighteen months. The IPC has given no public timeline for AI-specific guidance. Custodians making procurement decisions now are applying decade-old statutory language to tools built in the last two years. Some are betting that the right infrastructure partner makes that mapping easier rather than harder.