AI Data Residency in Canada: Why 'Our Cloud Is in Canada' Doesn't Mean Your AI Data Stays Here

Many Canadian IT leaders have made peace with the fact that their data lives in the cloud. The trickier question — one that most organizations have not fully answered — is *which* cloud, in *which* country, for *which* processing.

The assumption is common and understandable: our workloads run on AWS in Montreal, or on Azure Canada Central, so our data stays in Canada. For most workloads, that is true. For AI inference specifically, it often is not — and the gap between that assumption and reality has real implications under PIPEDA, Quebec's Law 25, and the OPC's ongoing guidance on transborder data flows.

How AI Inference Routing Actually Works

When your application sends a prompt to an AI model — whether through a direct API call, an AI assistant embedded in your SaaS platform, or a custom agent workflow — the text in that prompt travels to wherever the model is hosted and run. That hosting location is not necessarily the same as your cloud environment's region.

For Amazon Bedrock, AWS's managed AI model service, this distinction is explicit in the service documentation. Bedrock offers three processing modes:

• In-Region: the inference request stays within the single AWS region you specify
• Geo Cross-Region: AWS routes the request across regions within the same geography (e.g., North America) for higher throughput
• Global Cross-Region: the request can be routed anywhere worldwide

For Claude models on Amazon Bedrock, AWS documentation shows that ca-central-1 (Montreal) and ca-west-1 (Calgary) are not supported for In-Region inference. Cross-region inference — which routes requests outside Canada, typically to US East — is supported. In practice, this means that an organization running its application on AWS in Canada, using Bedrock to call Claude, is sending prompt data to a US region for processing, even though the rest of its workload stays in ca-central-1.

This is not a criticism of AWS or Anthropic. It reflects the current state of model deployment infrastructure, and the pattern is similar across most foundation model services: the models run in a small number of high-capacity regions, and regional availability expands over time. The point is that organizations cannot assume in-Canada processing simply because their cloud account is configured to a Canadian region.

The Direct API Case

Organizations calling Claude directly through the Anthropic API — rather than through Bedrock — send data to Anthropic's infrastructure, which is US-based. Claude for Teams and Claude Enterprise customers benefit from strong contractual data handling commitments: Anthropic does not train models on conversations from these subscriptions, and the terms include data handling and security commitments. But processing occurs in Anthropic's US infrastructure rather than in Canada.

The same general pattern applies to OpenAI, Google Vertex AI, and other foundation model providers operating primarily from US data centres, though each vendor's terms and regional infrastructure differ.

What PIPEDA Actually Says About Cross-Border AI Processing

Here is the part that surprises most Canadian business leaders: PIPEDA does not prohibit sending personal information to a foreign country for processing. Canada's federal private-sector privacy law takes an accountability-based approach. Under Principle 1 of Schedule 1, a Canadian organization that transfers personal information to a third-party processor — including a foreign one — remains accountable for how that data is handled. The transfer itself is permitted; the obligation is to ensure appropriate protection through contractual or other means.

The OPC's position on transborder data flows confirms this framework: organizations must be transparent with individuals about cross-border transfers and must use contractual measures to protect the data. The OPC has also expressed a view that consent may be required for transfers in certain contexts — an interpretation that not all legal experts agree with, and one that the OPC is actively consulting on.

The practical obligation is this: if your AI tools process personal information about Canadian individuals — employee data, customer data, health data, financial information — that processing creates accountability even when it occurs on foreign infrastructure. You need contractual data processing agreements, a record of your data flows, and, in most cases, a way to inform individuals that their data may be transferred outside Canada.

Quebec's Law 25 Sets a Higher Bar

For organizations subject to Quebec's *Act Respecting the Protection of Personal Information in the Private Sector* (Law 25), the cross-border transfer requirements are stricter than PIPEDA. Before disclosing personal information outside Quebec, organizations must conduct a Privacy Impact Assessment (PIA) to verify that the protection afforded in the receiving jurisdiction is equivalent to that in Quebec. They must also ensure by contract that the information is given equivalent protection.

This requirement applies to AI processing. If your organization has Quebec employees or customers whose data flows through your AI tools, you cannot simply accept a vendor's standard terms — you need to confirm that a PIA has been conducted and that contractual protections are in place.

The Commission d'accès à l'information (CAI) has been active in enforcing Law 25 since its phased implementation began in 2022, and AI-related complaints are a growing part of its workload.

The Specific Gap: What Your AI Vendor Contract Probably Does Not Say

Most standard AI vendor agreements — including the free and lower-tier plans for productivity tools with embedded AI features — do not explicitly address:

• The specific geographic locations where your prompts are processed
• Data retention periods for inference requests
• Whether your content is used in model training (important for Enterprise/Teams tiers, less defined for standard tiers)
• Security incident notification timelines for AI-specific events
• Subprocessor disclosure for the infrastructure the vendor uses

The OPC's 2026 guidance on third-party AI vendor compliance reinforces the point that PIPEDA's accountability principle extends to your AI vendors. A standard click-through terms-of-service does not constitute "appropriate contractual protections" under PIPEDA's accountability framework.

A Practical Risk-Tiering Framework

Not all data processed by AI tools carries the same PIPEDA risk. A workable approach for most Canadian organizations starts with tiering:

Tier 1 — Low risk: Prompts that contain no personal information. Draft communications that contain generic placeholders. Summarizing publicly available documents. Marketing copy generation. These carry minimal PIPEDA exposure regardless of where inference occurs.

Tier 2 — Moderate risk: Prompts that contain employee names, internal job titles, project descriptions, or aggregated business information without individual identifiers. Standard contractual protections and an internal data handling policy are appropriate.

Tier 3 — Higher risk: Prompts that contain customer names, contact details, health information, financial data, Social Insurance Numbers, or any information that would constitute "sensitive personal information" under PIPEDA or Law 25. These require a documented data processing agreement with the AI vendor, a record of the cross-border transfer, and — for Quebec — a prior PIA.

Tier 4 — Highest risk: Prompts containing information subject to professional privilege (legal, accounting, medical), regulatory restrictions (OSFI requirements for financial institutions), or contractual confidentiality obligations. These require legal review before being sent to any external AI service.

Most organizations find that a significant share of their actual AI use involves Tier 1 and 2 data — and that a clear policy governing Tier 3 and 4 data, rather than a blanket restriction on AI use, is the proportionate response.

What You Should Have in Place Now

Regardless of where your AI tools currently sit, four actions reduce your PIPEDA and Law 25 exposure in a material way.

1. Map your AI data flows. Document which AI tools your organization uses, what types of data flow through each, where processing occurs (to the extent the vendor discloses it), and what contractual protections are in place. This is the minimum required to demonstrate accountability under PIPEDA.

2. Get data processing agreements with your key vendors. For any tool processing Tier 3 or higher data, a contractual DPA is required under PIPEDA and Law 25. For Canadian organizations using Claude for Teams or Claude Enterprise, Anthropic provides data processing terms that cover these requirements. For Microsoft 365 Copilot, Microsoft's Data Protection Addendum covers PIPEDA-relevant obligations. Ensure these agreements are executed rather than assumed.

3. Update your privacy notice. If your organization transfers personal information to foreign AI processors, your privacy notice should say so. PIPEDA's transparency principle requires that individuals be informed of the fact of cross-border transfers, though not necessarily the specific country or processor.

4. Conduct a PIA for Quebec-subject data. If you have employees or customers in Quebec and your AI tools process their personal information outside Quebec, Law 25 requires a PIA. The CAI has published guidance on PIA methodology.

Looking Ahead: In-Region AI Processing

The availability of in-region AI inference will continue to expand. AWS has publicly indicated plans to extend Bedrock's regional availability, and Microsoft's announced CA$19 billion investment in Canadian AI infrastructure includes data centre expansion in Canada. As more model capacity is deployed in Canadian regions, in-region inference will become a practical option for organizations with the most stringent data residency requirements.

For now, the data residency question for AI workloads is best approached through the accountability framework PIPEDA describes — clear data flows, appropriate contracts, and a documented understanding of where your data goes — rather than through the assumption that your cloud region governs your AI processing location.

Sources

• Office of the Privacy Commissioner of Canada. *Consultation on Transborder Data Flows.* priv.gc.ca
• Office of the Privacy Commissioner of Canada. *Privacy Commissioner tables 2025–2026 Annual Report.* priv.gc.ca (June 2026)
• AWS Documentation. *Claude Sonnet 5 — Regional Availability.* docs.aws.amazon.com
• Statistics Canada. *Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026.* statcan.gc.ca
• Commission d'accès à l'information du Québec. *Act Respecting the Protection of Personal Information in the Private Sector (Law 25).* cai.gouv.qc.ca
• Bennett Jones LLP. *Transborder Data Flow Consent Standard Is Reaffirmed.* bennettjones.com
• Hunton Andrews Kurth. *OPC Reframes Consultation on Transfers for Data Processing.* hunton.com

Understanding where your AI data goes — and what your contracts say about it — is one of the first outputs of a structured AI governance review. Cloud Forces' AI Advisory service helps Canadian SMEs and mid-market organizations map their AI data flows, assess vendor contracts against PIPEDA and Law 25 requirements, and build the documentation that demonstrates accountability to the OPC and provincial regulators. As an AWS Consulting Partner since 2019, we work with the infrastructure layer as well as the policy layer. Book a consultation to review your current AI data posture.