Data sovereignty for small business: when your models shouldn't leave Canada

“Where does our data actually go?” is the question that stops a lot of AI projects — usually late, after a contract is on the table. It’s the right question, and it’s better asked early. For most Ontario SMBs the honest answer is: it depends on the data, and you have more options than the default.

The three hosting postures

Most deployments land in one of three places:

1. Frontier API (US-hosted). The big commercial models, called over an API. Cheapest to start, most capable, fastest to ship. Your prompts and data transit to, and may be processed in, the provider’s infrastructure — often outside Canada.
2. Canadian-hosted commercial. The same class of model run from a Canadian region, or a Canadian provider, so data stays in-country. Slightly more setup, sometimes a premium, but it clears most data-residency requirements.
3. Self-hosted open models. Open-weight models running on infrastructure you control — cloud or on-premise. Full custody, no third-party processing, higher operating cost and engineering effort.

How to decide — before you sign

The deciding factor is rarely “AI is risky” in the abstract. It’s specific:

• Regulatory or contractual residency. Healthcare, legal, public-sector, or client contracts that mandate Canadian data residency push you to posture 2 or 3 — non-negotiable.
• Data sensitivity. Personal information, trade secrets, or anything you’d be unwilling to see in a breach raises the bar regardless of regulation.
• Sensitivity of the workload, not the company. Often only one workflow is sensitive. You can run that one self-hosted and keep everything else on the cheaper API — you don’t have to pick one posture for the whole business.

The mistake is treating this as all-or-nothing. The right architecture is usually mixed: frontier API for the low-stakes 80%, Canadian-hosted or self-hosted for the 20% that actually carries risk.

What it costs

The trade is predictable. Posture 1 is the lowest total cost and the least control; posture 3 is the most control and the highest fixed cost — you’re now paying for, and operating, inference infrastructure. Posture 2 sits in between and clears most residency requirements without the operational burden of self-hosting. For most SMBs, a Canadian-hosted setup for sensitive workflows is the pragmatic middle.

The point of doing it early

Data residency is an architecture decision, and architecture decisions are cheap to make at the start and expensive to retrofit. We map this in the readiness audit — which data is sensitive, what each posture would cost you, and where the line should sit — so the question is answered before you sign anything, not after.