To run AI on sensitive data in Canada, you need to know which privacy laws apply to your data, where every part of the AI pipeline stores and processes it, and who could be compelled to access it. For many organizations the cleanest answer is a private LLM: an open-weight model hosted in a Canadian data centre or on your own hardware, so prompts, documents and outputs never leave infrastructure you control. Data residency in a Canadian cloud region can also work, but only if you verify where inference, logging and support access actually happen. This article covers the main laws, the infrastructure options and a practical checklist. This is not legal advice; consult a privacy lawyer for decisions about your own obligations.
The legal landscape in brief
PIPEDA
The Personal Information Protection and Electronic Documents Act (PIPEDA) is the federal private-sector privacy law. It applies to organizations that collect, use or disclose personal information in the course of commercial activity, and to federally regulated businesses such as banks, airlines and telecommunications companies. In provinces with private-sector laws recognized as substantially similar (currently Quebec, British Columbia and Alberta), the provincial law generally applies to activity within that province instead, while PIPEDA still covers interprovincial and international transfers.
PIPEDA is built on ten fair information principles, including accountability, identifying purposes, consent, limiting collection, limiting use and retention, accuracy, safeguards and openness. For AI projects, several are directly relevant:
- Accountability for third parties. When you send personal information to a vendor for processing, including an AI provider, you remain responsible for it and must use contractual or other means to ensure a comparable level of protection.
- Cross-border transparency. PIPEDA does not generally prohibit processing outside Canada, but the Office of the Privacy Commissioner of Canada (OPC) expects organizations to tell people when their information may be processed in another country and could be accessed by authorities there.
- Purpose limitation. Using customer data collected for one purpose to train or tune a model for another purpose may require fresh consent.
- Safeguards and breach reporting. Security must match the sensitivity of the information. Breaches of security safeguards that create a real risk of significant harm must be reported to the OPC and to affected individuals, and records of all breaches must be kept.
Quebec's Law 25
Quebec's Law 25 (formerly Bill 64) modernized the province's privacy legislation in phases between 2022 and 2024. It is the most demanding privacy regime in Canada for AI projects. Key requirements include:
- A designated person in charge of the protection of personal information, by default the highest-ranking person in the organization.
- Privacy impact assessments for projects to acquire, develop or overhaul information systems that involve personal information, which covers most AI deployments.
- An assessment before communicating personal information outside Quebec, considering the sensitivity of the data, the purpose, the safeguards and the legal framework in the destination. This applies to other provinces as well as other countries.
- Transparency for decisions based exclusively on automated processing: the person must be informed, and on request told what personal information was used and the main factors behind the decision, with an opportunity to have it reviewed by a person.
- Reporting of confidentiality incidents that present a risk of serious injury to the Commission d'accès à l'information (CAI) and to affected individuals.
Penalties under Law 25 are substantial and can be calculated as a percentage of worldwide turnover, so organizations serving Quebec residents should treat it seriously.
Provincial health-privacy laws
Health information has its own laws in most provinces. In Ontario, the Personal Health Information Protection Act (PHIPA) governs how health information custodians, such as physicians, hospitals, pharmacies and laboratories, collect, use and disclose personal health information, and it places obligations on agents and electronic service providers who handle that information on their behalf. The Information and Privacy Commissioner of Ontario oversees it, and PHIPA has been recognized as substantially similar to PIPEDA for health information custodians.Alberta, Manitoba, Saskatchewan and other provinces have their own health-information statutes. If you build AI tools for clinics or health organizations, the custodian's obligations flow down to you as a vendor.
Public-sector bodies
Government bodies, municipalities, school boards and universities are covered by separate public-sector privacy laws. Some provinces have historically imposed stricter rules on storing or accessing personal information outside Canada, and these rules change; British Columbia, for example, relaxed its strict residency requirement in 2021. Public-sector procurement contracts often add residency requirements of their own.
Federal AI legislation: AIDA
The proposed Artificial Intelligence and Data Act (AIDA) was introduced as part of Bill C-27, alongside the proposed Consumer Privacy Protection Act that would have replaced much of PIPEDA. Bill C-27 died on the order paper when Parliament was prorogued in January 2025. As of that point, Canada had no federal AI-specific statute in force. Existing privacy, human rights and consumer protection laws continue to apply to AI systems, and new federal privacy or AI bills may have been introduced since, so check the current status before relying on this summary.
Residency versus sovereignty
Data residency means the data is physically stored in Canada. Data sovereignty is broader: it asks which country's laws can reach the data. A cloud provider headquartered in the United States may be subject to US legal orders, such as those under the CLOUD Act, regardless of where the servers sit. Whether that matters depends on your data, your contracts and your risk tolerance. For many commercial workloads, a Canadian region with strong contractual commitments is acceptable. For highly sensitive health, legal, financial or government data, many organizations prefer infrastructure operated by a Canadian company or by themselves.
Infrastructure options
Canadian cloud regions
The major cloud providers all operate Canadian regions: AWS in Montreal and Calgary, Microsoft Azure in Toronto and Quebec City, Google Cloud in Montreal and Toronto, and Oracle in Toronto and Montreal. Storing documents and running databases in these regions is straightforward.
Managed AI services need more care. Not every model is offered in every Canadian region, and some services offer "global" or cross-region inference options that route prompts to wherever capacity is available, which may be outside Canada. Abuse monitoring, logging and support access can also involve other regions. A Canadian storage region does not guarantee Canadian inference. Confirm the processing location for the specific model and deployment type you are using, and get it in writing.
Private and on-premises open-weight models
Open-weight models, such as the Qwen and Llama families, can be downloaded and run on infrastructure you control. Serving tools make this practical:
- vLLM is a high-throughput inference server for GPU machines, suited to multi-user production deployments.
- Ollama makes it easy to run models on a workstation or small server, useful for pilots and smaller teams.
- MLX is Apple's framework for running models efficiently on Apple silicon, a cost-effective option for some on-premises workloads.
With a private deployment, prompts, documents, embeddings and outputs stay inside your network or a Canadian data centre you choose. The trade-offs are that you own capacity planning, updates and security, and the most capable proprietary models are not available this way. For document question answering, extraction, classification and drafting, current open-weight models paired with good retrieval are often more than adequate. Licences differ between models and some carry conditions on commercial use, so check each one.
| Option | Data control | Effort | Good fit |
|---|---|---|---|
| Global managed AI API | Lowest; processing location may vary | Lowest | Non-sensitive data, prototypes on synthetic data |
| Managed AI in a Canadian region, in-region inference confirmed | Residency in Canada; foreign-law reach depends on provider | Low | Most commercial workloads with personal information |
| Open-weight model in a Canadian-operated data centre | High | Moderate | Sensitive data, sovereignty concerns |
| Open-weight model on-premises | Highest | Highest | Health, legal, government and air-gapped environments |
A practical checklist
- Classify the data. Identify whether the workload involves personal information, personal health information, Quebec residents' data, or confidential business data.
- Map the whole pipeline. List every component that touches data: ingestion, document storage, embeddings, vector database, model inference, logs, monitoring, backups and support access.
- Confirm processing locations. For each component, document where it stores and processes data, including any global or cross-region inference settings.
- Review vendor contracts. Check whether the provider uses your data for training, how long it retains prompts and outputs, and what notice it gives for legal access requests.
- Run a privacy impact assessment. Required for many projects under Law 25 and good practice everywhere.
- Apply least privilege. Enforce document-level permissions in retrieval so users can only get answers from content they are allowed to see.
- Plan for automated decisions. If the system makes or substantially drives decisions about people, design for explanation and human review.
- Log and retain deliberately. Keep enough to audit, not more than you need, and set retention periods.
- Update your privacy notice. Be transparent about AI use and any processing outside Canada or outside Quebec.
- Prepare for incidents. Make sure your breach response plan covers the AI components.
How Dassjay Labs helps
Dassjay Labs is an AI consultancy and product studio based in Kitchener, Ontario, serving organizations across Canada. We design and deploy private LLM systems using open-weight models on vLLM, Ollama or MLX, either on your hardware or in Canadian infrastructure. We pair them with permission-aware RAG and document AI so sensitive documents can be searched and cited without leaving your control. Our AI strategy work helps map data flows and choose the right deployment model, working alongside your legal advisers rather than replacing them. Contact us to discuss your requirements.