Arcadion
AI governance framework
Close Icon

Stay up to date with the latest news in Managed IT, cybersecurity and Cloud Infrastructure.

AI Governance in Canada: A Practical Framework for PIPEDA, Law 25, and Microsoft Copilot


Friday, November 13, 2026
By Simon Kadota
Share

AI governance is the collection of policies, roles, controls, and records that an organization uses to decide which AI tools and uses are allowed, what data they can access, who reviews their outputs, and how it demonstrates compliance. Canada does not have a law specifically about AI, but privacy law already covers your use of AI.

Key takeaways

  • No federal AI Act in Canada; AIDA expired in January 2025.
  • PIPEDA and provincial privacy laws already cover AI that uses personal information.
  • Quebec’s Law 25 has a specific automated-decision rule.
  • Copilot follows existing permissions and first fixes oversharing.
  • ISO/IEC 42001 is the certifiable standard for organizations

What is AI governance, and how is it different from an AI policy?

AI governance is the entire system for managing AI, while an AI acceptable use policy is just one document within that system. The policy describes what staff can do, whereas governance assigns owners, reviews use cases, and produces proof.

Data governance is about how information is classified and accessed; AI governance adds which tools can use it.

Does Canada have an AI law? What applies in 2026

No federal law on AI is in effect. The Artificial Intelligence and Data Act (AIDA) died with Bill C-27 when Parliament was prorogued in January 2025, and ISED says the voluntary code is still in place until regulation is in place.

InstrumentBinding?Applies toWhat it means for AIControl in your framework
PIPEDAYes.Private-sector commercial activity (Alberta, BC, and Québec have similar laws)Ten principles, including accountabilityData rules, named owner
Québec Law 25 (P-39.1)Yes.Private-sector enterprises (Québec)Automated-decision notice, PIAs, profiling noticeLaw 25 workflow, human reviewer
Bill C-36Not yetUnknownTabled June 15, 2026; at second readingWatch list
OSFI Guideline E-23Yes, from May 1, 2027Federally regulated financial institutionsModel risk, including AI/MLModel inventory, validation
Voluntary Code on Responsible Generative AINoAdvanced generative AI developers and managersSix outcomesPrinciples
OPC and provincial gen-AI principlesGuidanceGenerative AI technologiesNine principles; existing law still applies.Privacy review
Algorithmic Impact AssessmentMandatory for federal departmentsAnyone can use it.65 risk and 41 mitigation questionsRisk tiering
ISO/IEC 42001VoluntaryAny organizationCertifiable AI management systemThe whole framework

What does Québec Law 25 require when AI makes decisions?

Where a decision of an enterprise is based solely on the automated processing of personal information, section 12.1 of Québec’s Act respecting the protection of personal information in the private sector requires the enterprise to inform the person of the decision and to explain it upon request. The rule was implemented in September 2023. A working process:

  1. Find the trigger. List decisions that use personal information and involve no person before they are final.
  2. Notify on time. Inform the person no later than when you tell them the decision.
  3. Prepare the explanation. Upon request, please provide the personal information used, the reasons, principal factors, and parameters, as well as the right to have it corrected.
  4. Route observations. Let the person submit observations to a staff member who can review the decision.
  5. Keep evidence. Log notices, requests, and reviews.

Three other sections of the same Act apply to AI projects:

  • Section 3.1 (accountability): the person with the highest authority is accountable, may delegate the role, and must publish the delegate’s title and contact details.
  • Section 3.3 (privacy impact assessment): required for any project to acquire, develop, or overhaul an information system involving personal information.
  • Section 8.1 (profiling technology): requires prior notice before using technology that can identify, locate, or profile people.

Administrative penalties reach $10 million, or 2% of worldwide turnover, and penal fines $25 million, or 4% (CAI).

Arcadion supports organizations across the Ottawa-Gatineau region; if you collect personal information in Québec, confirm with counsel how Law 25 applies.

An AI governance framework in 7 parts

The AI governance framework has seven parts, from inventory to monitoring. For principles, see the six outcomes of the Voluntary Code (see the FAQ).

1-AI inventory. List every AI tool in use, including Shadow AI, with its owner, purpose, and data.

2- Intake and risk tiers. Tier each new use case:

TierData involvedImpact on peopleTypical controls
LowPublic or non-sensitive internalNoneAcceptable use policy
ModerateConfidential internalIndirectApproved tool, logging
HighPersonal or regulated informationAffects individuals; a person decidesPIA, named reviewer, audit trail
RestrictedSensitive personal informationDecides with no human stepLaw 25 workflow, executive sign-off, or not allowed

3-Ownership. Name an executive owner, privacy officer, IT/security lead, and business owner for each use case.

4-Data rules. Define which data classifications may enter which tools. Arcadion’s AI data security and governance services include AI-specific data classification and policy definition.

5-Security controls. Apply identity controls, least privilege, data loss prevention, and logging. Custom AI needs permission-aware retrieval for RAG.

6-Human oversight. Decide where a person must approve AI output; use the reviewer’s Law 25 process names.

7-Monitoring and review. Set re-review dates and monitor custom models for drift. Arcadion’s managed AI services cover audit logging, compliance reporting, quarterly reviews, and model version tracking.

Before you scale Copilot: the permission checks that matter

Microsoft 365 Copilot only uses content the user is already authorized to access (Microsoft Learn), so overshared content becomes easy to find. Run these checks first (Microsoft Learn):

  1. Run data access governance reports for the site permissions baseline, “Everyone except external users,” and sharing links.
  2. Tighten sharing defaults. Microsoft notes SharePoint sharing defaults to the most permissive option.
  3. Apply Restricted Access Control to limit sensitive sites to a named group.
  4. Use Restricted Content Discovery to keep sites that must stay accessible out of Copilot and organization-wide search.
  5. Apply sensitivity labels with encryption. Copilot requires the user to have EXTRACT and VIEW rights to access encrypted content.
  6. Turn on the Purview audit. Audit records capture Copilot prompts, responses, and referenced content.

SharePoint Advanced Management features that support Copilot require at least one Copilot license. Limited app-based site creation requires SAM Plan 1 (Microsoft Learn).

In practice: this meant that when Arcadion set up the Canadian Nurses Protective Society (CNPS) for Copilot, existing Microsoft 365 permissions were reviewed and aligned so that staff could only pull down information they were already authorized to access. But AI-written memos still need a professional’s eye. The case study indicates that more than 150 hours are regained per month.

Not sure what Copilot can see?
Arcadion conducts a Microsoft 365 assessment that reviews SharePoint and OneDrive permissions, external sharing, DLP, and data classification over the course of one to two weeks. Arcadion then delivers a Microsoft 365 Health & Optimization Report and strategic roadmap.
Request an Assessment

ISO/IEC 42001 and AI governance certification

ISO/IEC 42001, published in December 2023, is a certifiable AI management system standard.

Individual credentials, such as the IAPP’s AI Governance Professional (AIGP), certify a person, not your organization’s controls. ISO/IEC 42001 is worth pursuing when customers or regulators ask for evidence, or when you already run ISO/IEC 27001.

How to start in 30 days

  • Build the AI inventory.
  • Publish an acceptable use policy.
  • Risk-tier your top 10 use cases.
  • Fix Copilot oversharing.
  • Assign owners and set review dates.

FAQ

What is AI governance in simple terms?

AI governance means deciding which AI tools staff may use, with what data, and who checks the output, then keeping proof.

What are the main principles of AI governance?

Frameworks vary. Canada’s Voluntary Code lists six outcomes: accountability, safety, fairness and equity, transparency, human oversight and monitoring, and validity and robustness.

Does Law 25 apply to AI chatbots?

Section 12.1 applies only to decisions based exclusively on automated processing of personal information. A chatbot that collects personal information is still subject to Law 25’s general rules.

Does Microsoft Copilot see everything in our tenant?

No. It uses only content the signed-in user can access, which may be more than intended.

Put your AI governance on solid ground.

Arcadion’s managed AI services keep AI systems monitored and secure.

Talk to our AI Data Governance experts