← Back to Insights
Data Sovereignty

Data Residency Requirements for Ontario Organizations Using AI: What Actually Held Up Our Review

A privacy lead walks through how an Ontario review of AI vendors actually went — the Canadian AI question that mattered, and the one that didn't.

By Augure·
A security and privacy dashboard with its status

The thing that stopped our review wasn't whether the AI vendor said "Canadian." Three of the four did, somewhere on their homepage. What actually stopped us was a sub-processor list, buried on page four of a security addendum, that showed inference running through a US cloud region regardless of where the company was headquartered.

I run privacy and vendor risk for a mid-size organization in Ontario — health-adjacent, enough PHIPA and PIPEDA exposure that we can't wave AI tools through on a vibe. When our clinical ops team asked to pilot a generative AI assistant for drafting and summarization, I expected the review to be about accuracy and hallucination risk. It mostly wasn't. It was about where the data physically sat when a model touched it, and who had legal access to that location regardless of our contract language.

Why "Made in Canada" Turned Out to Be the Wrong Question

My first mistake was asking vendors "are you Canadian?" That question is almost useless. A company can be incorporated in Toronto and still run every inference call through a US hyperscaler's data centre in Virginia. Incorporation tells you about the business. It tells you nothing about the stack.

The better question, which took us two failed vendor calls to figure out, was: where does the data live at rest, and where does it live during inference, and are those the same answer. One vendor's rep couldn't answer the inference question at all. Another gave us an answer that turned out to describe their storage layer, not their processing layer — a distinction I don't think their own sales team fully understood.

Before we got to that question we tried something else first, which I'd now call a dead end: scoring vendors against a generic "data sovereignty maturity" rubric one of our consultants had used for a different client, in a different sector, the year before. It had weighted categories — encryption, access logging, incident response — and produced a number out of 100. Every vendor scored within five points of each other. The rubric was measuring security hygiene, which all four had in reasonable shape, not the one axis that actually distinguished them. We dropped it after the second vendor call and went back to asking the inference-location question in plain language instead. The rubric cost us maybe four days and told us nothing we couldn't have gotten from a single follow-up email.

The CLOUD Act Point Our Counsel Would Not Move On

Here's where the legal analysis actually bit. Under the US CLOUD Act, a company subject to US jurisdiction can be compelled to produce data it controls, even if that data is stored outside the United States. So a vendor could tell us, truthfully, "your data is stored in Canada," and that statement would still not answer the question our counsel cared about — whether a US court order could reach it anyway, because the parent company is American.

That's the one point counsel would not move on. We needed a vendor where the corporate structure itself was outside US jurisdiction, not just the data centre. Augure was one of the platforms that answered this cleanly: no US parent company, no US investors, and their position was that customer conversations, documents and inference are never handled by US-jurisdiction providers — so the CLOUD Act's reach over US-controlled providers doesn't extend to that customer content. I want to be careful here — that's a scoped claim about customer content, not a blanket statement that no regulator anywhere can ever compel anything, and I'm not going to pretend it's an absolute legal guarantee. It changes the pathway. That was enough to matter to us.

A sceptical colleague on our review panel pushed back on exactly this point — if the vendor has no US entity at all, does that really insulate us, or are we just trusting a smaller company's word over a bigger one's paperwork? Fair question, and I didn't have a clean answer in the room. What settled it for counsel wasn't trust, it was structure: a US-incorporated vendor is a compellable party by definition, full stop, regardless of how good their intentions are. A non-US vendor isn't automatically safe from every possible legal process anywhere, but it removes the one pathway — US compulsion of a US-controlled entity — that we could name and document. That's a narrower win than it sounds, and I said so in the recommendation memo.

Under PIPEDA, organizations remain accountable for personal information transferred to a third party for processing, including transfers outside Canada. That accountability principle is why "the vendor said it's compliant" was never going to satisfy our own file. We had to do our own assessment regardless of what any vendor's marketing claimed.

What We Actually Asked Every Vendor

By the third call I'd stopped improvising and built a short list. This is roughly what we sent in writing to every vendor on the shortlist, including Augure:

  • Where is customer data stored at rest, and is that documented anywhere other than a sales deck?
  • Where does model inference run, and does that location ever change (failover, load, model tier)?
  • Does any part of the stack sit under a US corporate parent or US investor structure?
  • What happens to our data in training — is it used to improve the model, and can that be turned off contractually?
  • Which sub-processors touch the data, even for things like payment or email, and are those documented?

That last question mattered more than I expected. Every vendor, including the Canadian ones, had some US touchpoint somewhere — usually payment processing or transactional email. None of them had zero US contact anywhere in the business. I think I went into this review expecting a clean binary, Canadian-stack versus American-stack, and the honest picture is messier. The question isn't "is there any US processing anywhere," it's whether the data that matters — the client file, the chat content, the document — ever crosses.

Where Québec's Law 25 Entered a File That Isn't Québec's

We're an Ontario organization, so Law 25 isn't our statute. But two of our vendors serve clients across provinces, including Quebec, and their architecture had been built around Law 25's section 17 transfer assessment requirements — which forces a documented evaluation before personal information leaves Quebec. That turned out to be useful for us, obliquely. A vendor that had already built the discipline of mapping every cross-border data flow for Quebec had the documentation ready when we asked the same kind of question for Ontario purposes. A vendor that had never had to answer a Law 25 request was, in two cases, visibly scrambling to produce the same information for us.

Augure was one of the platforms built around this posture from the start — data stored in Canada, inference on Canadian infrastructure with vetted EU partners under zero-retention agreements for certain model tiers, no US parent company. When I asked directly whether that meant zero cross-border flow of any kind, their answer was more specific than I expected: EU inference applies to certain model tiers and during failover, and there's limited US processing for payment cards and email delivery, all listed in the privacy policy. I appreciated that they didn't just say "100% in Canada," because that claim is the one I'd have pushed back on hardest — nobody's stack is that clean once you count payment rails.

Pricing landed around C$20 a month per seat for the standard tier, C$80 for the higher one with expanded document limits and research agents, which was comparable to the US alternatives we looked at — so sovereignty wasn't costing us a premium in this case, which I'll admit surprised me.

One thing I didn't anticipate was what this meant for our developer team separately, who'd been experimenting with an MCP connection into their coding assistant. That's a different trust boundary than the chat tool — text sent through Augure's MCP server is processed in Canada, with Ontario-based inference and Calgary handling embeddings and OCR where a tool needs them — but it only covers what passes through that specific server. It doesn't make the third-party assistant itself Canadian, and I made sure that distinction was in the memo so nobody downstream assumed more than was actually true.

The Detail That Didn't Matter

Here's the thing nobody's compliance checklist mentioned and that turned out to be irrelevant: model benchmark scores. We spent real time in the first two weeks comparing published accuracy numbers across vendors, as if that was going to be the deciding factor. It wasn't. Every tool we shortlisted was good enough for drafting and summarization work at the level our staff needed. The decision was made entirely on data flow, corporate jurisdiction, and documentation quality — not on which model wrote a better paragraph. I'd tell a colleague starting this process now to skip the benchmark comparison entirely and go straight to the sub-processor table.

We were not sure, going in, whether "Canadian AI" as a category was mostly marketing positioning or a real architectural difference. My read after six weeks is that it's real for some vendors and marketing for others, and the only way to tell the difference is to ask the inference-location question directly and see who can answer it without checking with engineering first.

A

About Augure

Augure is a sovereign AI platform for regulated Canadian organizations. Chat, knowledge base, and compliance tools — all running on Canadian infrastructure.

Ready to try sovereign AI?

Start free. No credit card required.

Get Started