A Canadian address is only one part of an AI supplier review.
A Canadian software supplier answering a public-sector questionnaire can usually identify its hosting region. The difficult entries arrive a few rows later: who can access support data, which model service receives prompts, and what happens to the evidence when the contract ends.
I would turn those entries into an operating review before describing a system as sovereign. A map pin is useful, but the reviewer needs the arrangements behind it. Start with a synthetic task so the supplier can demonstrate the boundary without bringing personal information into an early evaluation.
Establish the applicable requirement.
The Treasury Board scope guide for the Directive on Automated Decision-Making describes a defined federal institutional and use-case boundary. It does not make every Canadian use of AI subject to the directive. Keeping a human involved also does not, by itself, put an administrative decision system outside that scope.
For a supplier, the useful first step is to ask the departmental owner which requirements apply to the proposed use and which obligations appear in the contract. Put that answer beside the technical scope. Have the appropriate legal and procurement reviewers resolve uncertainty rather than convert a framework badge into a compliance claim.
Follow one piece of work.
Suppose the evaluation involves drafting a response from an invented case file. Ask the operator to show the approved source and the identity reading it. Follow the content through any model endpoint, then identify where the draft and operational records are retained.
Repeat that exercise for support access. Who can see a diagnostic bundle? Does it contain source text? Can the customer authorise a bounded support session and establish when access ends? These questions need answers for the actual arrangement, including subcontractors. The firm’s Canadian address cannot answer them on its own.
Write down the gaps as you find them. An undocumented destination is an unresolved review item. Giving it an owner and a required evidence type is more useful than leaving a reassuring sentence in the questionnaire.
Observe the human boundary.
Prepare a proposed action against a controlled test system and reject it. Preserve the proposed content and the human decision. Inspect the source system independently to establish what changed, if anything. If that independent check is unavailable, record the limitation.
The accompanying matrix pairs each question with a record the reviewer can request. It is an evaluation method. It does not show a customer deployment, a government acceptance result or a product certification.
In a Vantage Workspace evaluation, I would use the same discipline: agree the permitted workflow and ask for evidence tied to its configuration. A template helps organise the review. It cannot decide the scope of a legal obligation or substitute for the source records.
Review the exit while the relationship is simple.
Ask the customer and supplier to retrieve the test evidence from outside the original conversation. Establish who holds the export, what remains in backups, and which retention obligations affect deletion. Record the agreed access and custody arrangements before the pilot grows.
Canadian operation can matter to a buyer’s contracting and operational choices. It does not establish immunity from foreign jurisdiction or imply government endorsement. Those claims need a different basis; they should not be smuggled into a hosting statement.
Questions buyers ask.
Does hosting in Canada make an AI system sovereign?
Hosting location settles one entry in the review. The reviewer still needs to know who can access support data, which model service receives prompts, where records are retained and what happens to the evidence when the contract ends. Canadian operation does not establish immunity from foreign jurisdiction.
Does the Directive on Automated Decision-Making apply to every Canadian AI supplier?
No. The Treasury Board scope guide describes a defined federal institutional and use-case boundary. Ask the departmental owner which requirements apply to the proposed use and which obligations appear in the contract. Legal and procurement reviewers should resolve any uncertainty.
How can a supplier show its data boundary without using personal information?
Run a synthetic task. Follow an invented case file from the approved source through any model endpoint to the place where the draft and the operational records are kept. Record each undocumented destination as an open review item with an owner.
For the wider review method, read the CISO evaluation guide for AI workspaces.
Bring one unresolved supplier question to the Agentic AI Procurement Handbook. Turn it into a bounded test with a named reviewer and an observable outcome. That gives both parties a concrete item to close before live work begins.
For a working definition to put in the evaluation brief, see What is a sovereign AI platform? on the workspace site.
Josh Olayemi · Founder, Handvantage · September 2026
