Title card: How Canadian public bodies should procure AI. Scope, Impact assessment, Residency, Supplier records, Synthetic test.

How Canadian public bodies should procure AI: what the rules actually require.

Short answer. A Canadian public body procures AI correctly by testing the rule that applies to its own level of government. A vendor’s account of its own policies does not settle that question. Federal departments check scope against the Treasury Board Directive on Automated Decision-Making and complete an Algorithmic Impact Assessment. Provincial and municipal bodies work from their own privacy statute. A Canadian address on the contract answers only one part of the review.

Government procurement teams already run a checklist for privacy, accessibility and security. AI adds one more question: which rule governs this system, and what record does it produce? The rules differ by level of government. The first job is working out which one applies to you.

Does the Treasury Board Directive on Automated Decision-Making apply to your project?

The Directive on Automated Decision-Making is a federal policy instrument. It took effect on April 1, 2019, with compliance required by April 1, 2020, and it is reviewed every two years. The version in force was last updated on June 24, 2025. You can read the current text on the Treasury Board of Canada Secretariat’s policy site.

It applies to “departments” as the Financial Administration Act defines them, the federal statute setting out how federal institutions are run. The Directive’s own scope guide puts the count at roughly 97 federal institutions, listed in three schedules of that Act. The Canada Revenue Agency is excluded by its own founding law. The Canadian Armed Forces sits outside the definition too, though National Defence is covered.

Four more conditions narrow the scope further. The system must be developed or procured after April 1, 2020, or significantly modified after that date. It must support or fully automate an administrative decision. That means a decision a government official makes about a specific person’s rights, benefits or obligations. It must replace or assist human judgement, through rules, statistical models or AI. And it must run in production. A pilot tested on sample data in a sandbox, with no real client affected, falls outside the Directive until it goes live.

None of this reaches provincial ministries, municipalities, health authorities or Crown corporations. A procurement team outside the federal public service needs a different starting point, covered further down.

What is the Algorithmic Impact Assessment, and when do you complete one?

The Algorithmic Impact Assessment, or AIA, turns the Directive’s scope test into a numeric result. It is a mandatory questionnaire, 65 risk questions and 41 mitigation questions, run through the Algorithmic Impact Assessment tool. The answers produce an impact level from I to IV, minimal to very high. That level sets how much client notice, human oversight and sign-off the Directive requires.

A department completes the AIA at the design phase, then again before production, and updates it whenever the system’s scope changes. Departments publish the finished assessment on the Open Government Portal, in both official languages. A completed AIA for a comparable system elsewhere in government is often something you can read before writing your own requirements.

The Directive names further obligations once the assessment is done, and they scale with the impact level. Clients must get notice that a decision will be made or assisted by an automated system. Every employee who builds or runs the system needs training. Before the system reaches production, someone signs off on it: an assistant deputy minister for lower levels, a deputy head for the highest. Clients also get told how to challenge the decision. Ask the supplier to help you produce these records as you build the system, ahead of when the auditor asks.

What does “Canadian data residency” actually settle?

A vendor who says the data stays in Canada has answered a narrower question than it sounds like. Residency tells you which country the servers sit in. It says nothing about who holds the administrative keys, or which model service actually processes a prompt. It says nothing about who can read the action logs, or what happens to your records the day the contract ends.

We cover the fuller test in a Canadian address is only one part of an AI supplier review and what is a sovereign AI platform. In short: ask where the data sits, who holds the keys, which model receives the prompt, who reads the logs, and what happens on exit. A residency clause answers the first question and leaves the other four for you to test.

What do the provincial privacy statutes add outside the federal Directive?

Health authorities, municipalities and most Crown corporations sit under provincial law. The federal Directive does not reach them. I have not found a provincial equivalent to the Algorithmic Impact Assessment; check your own province’s rules before assuming one exists. The general privacy statute is the rule that actually governs how these bodies procure an AI system touching personal information.

In British Columbia, the Freedom of Information and Protection of Privacy Act, known as FIPPA, sets these rules. It covers provincial ministries, health authorities, municipalities and other listed public bodies. The Office of the Information and Privacy Commissioner for BC oversees it. Every covered body needs a privacy management programme, and must report a breach when harm is possible.

Ontario runs two separate acts. Its own Freedom of Information and Protection of Privacy Act covers provincial ministries and agencies. The Personal Health Information Protection Act, or PHIPA, sets separate rules for hospitals and other health information custodians handling patient data. Neither act mentions AI by name. Both still govern how an AI system may collect and disclose personal information once it starts touching it. A procurement team in Ontario needs to read the vendor’s data handling against those acts directly.

If your jurisdiction is not British Columbia or Ontario, do not assume the pattern above holds. Check your own province’s freedom of information and health privacy statutes before you write requirements. Details differ by province.

What record should you ask the supplier for?

The table below pairs the question with the record that actually answers it. A verbal assurance is not a record.

Question The record to ask the supplier for
Does this system trigger the federal Directive? A written scope determination against the Directive’s own criteria, or a completed Algorithmic Impact Assessment with its impact level
Where does the data actually sit? The named storage region and the subprocessors involved, confirmed beyond a residency statement
Who can access our account? A list of vendor staff with standing access, and what a support ticket against a test account actually reveals
Which model processes a prompt? The named model endpoint and provider for the specific workflow, in writing
Can we produce the AIA-required notice to clients? A sample of the notice language the system generates, tied to a real decision type
Can we export the action logs ourselves? A raw log for one action, pulled directly from your own console
What happens to our data if we leave? The offboarding clause naming a deletion timeline, with a way to verify it independently

How do you run the evaluation, in order?

Run this sequence before you sign, using a scoped test account and synthetic data with no real client information.

  1. Confirm which rule governs your project: the federal Directive, a provincial FIPPA-type statute, or both, if a federally funded programme is delivered by a province.
  2. If the federal Directive applies, run the Algorithmic Impact Assessment and get the impact level before finalising requirements.
  3. Submit one synthetic task with no real personal data, and ask which model endpoint processed it. Get the answer in writing.
  4. Ask the vendor to generate the client notice language your impact level requires, and check it against a real decision type your team handles.
  5. Request the raw log for that synthetic task, pulled from your own account. A report the vendor prepares afterward does not count as the record.
  6. Confirm the offboarding process in writing: what happens to your data the day the contract ends.

Where do federal rules simply not reach?

The Directive binds federal departments and the agencies named in its schedules. It applies to the federal institutions named under the Financial Administration Act, so provinces, territories and municipalities sit outside it. The Financial Administration Act it rests on is a federal statute, with no reach into those other governments.

Health delivery is a provincial responsibility, so most health authorities sit outside the federal Directive by default. Many Crown corporations sit outside its schedules too. None of this means AI use at those bodies goes unregulated. A different rule applies: the relevant provincial privacy statute, and sometimes a sector-specific law. A procurement team that assumes the federal rule reaches everywhere will miss the rule that actually applies to its own body.

Questions buyers ask.

Does a Canadian address on the contract satisfy a federal AI procurement requirement?

No. Residency tells you where the servers sit. It says nothing about who holds the administrative keys, which model processes the prompt, who can read the logs, or the Directive’s own requirements around notice, human oversight and sign-off. Test those separately.

Do municipalities have to complete an Algorithmic Impact Assessment?

No. The Algorithmic Impact Assessment is a requirement under the federal Directive on Automated Decision-Making, which does not bind municipal governments. A municipality’s AI procurement is governed by its province’s freedom of information and privacy statute instead.

What happens if a federal system built before April 2020 gets a major upgrade?

The Directive’s own scope guide treats a significant modification, such as an upgrade to the system’s analytical capability or a change to who it affects, as creating a new system for scope purposes. That can bring an older system into the Directive’s requirements for the first time.

Can a procurement team read another department’s completed AIA before writing its own requirements?

Yes, in most cases. Departments publish finished Algorithmic Impact Assessments on the Open Government Portal, in both official languages. A comparable system’s published assessment is a useful reference point for what a supplier should be able to produce for yours.

This is the evaluation method behind the twelve tests in the Agentic AI Procurement Handbook. It is an evaluation method. It is not legal advice. Confirm which rules bind your specific body with your own legal and privacy office before you rely on any of it.

Josh Olayemi · Founder, Handvantage · September 2026 · About the author


Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *