Canada AI Regulation Guide (post-AIDA): The Current Landscape

Canada still has no federal AI statute: the Artificial Intelligence and Data Act (AIDA) died with Bill C-27 in January 2025, and the privacy bill tabled in June 2026 deliberately leaves AI out. That does not mean AI in Canada is unregulated. Federal institutions are bound by the Treasury Board Directive on Automated Decision-Making, Quebec already regulates decisions made exclusively by automated processing, Ontario has a statutory AI framework for its public sector waiting on regulations, and privacy regulators have set out how existing law applies to generative AI. This guide maps the instruments that apply today, their status and penalties, what they mean for AI agents and the identities they run on, and the steps a security or AI governance team should take. It is for practitioners and executives at organisations that build, sell or deploy AI in Canada, including suppliers to Canadian governments.

Key takeaways

• AIDA is dead and has not been replaced. Parliament was prorogued on 6 January 2025 and Bill C-27, which contained AIDA as Part 3, died on the Order Paper. Bill C-36, tabled on 15 June 2026, reforms private sector privacy law but contains no AI-specific provisions.
• Bill C-36 would still touch automated decisions. As tabled it requires organisations to disclose information about automated systems and the criteria they use, and to give individuals a way to challenge decisions. At October 2026 it had completed first reading only.
• Quebec is the binding private sector rule today. Section 12.1 of Quebec's private sector privacy act, in force since 22 September 2023, requires notice of decisions based exclusively on automated processing, an explanation on request and a route to have the decision reviewed by a person.
• The federal government regulates its own AI hard. The Directive on Automated Decision-Making, last amended on 24 June 2025, requires an Algorithmic Impact Assessment and impact-level controls, and the May 2026 Guide on the Use of Agentic Artificial Intelligence sets controls such as unique agent IDs and external kill switches.
• Voluntary rules fill the private sector gap. ISED's September 2023 Voluntary Code of Conduct for advanced generative AI and the federal, provincial and territorial privacy commissioners' December 2023 generative AI principles are the reference points regulators and customers use.

Scope: who and what is covered

Canadian AI rules are a patchwork split by sector of the user (federal government, provincial public sector, private sector) and by province. Work out which layer applies to each system before you plan controls.

• Suppliers to government: The Directive binds federal institutions, not vendors, but institutions pass its requirements to suppliers through procurement, so vendors of automated decision systems should expect to support an Algorithmic Impact Assessment, peer review and testing evidence.
• Private sector outside Quebec: No AI-specific statute applies. Existing privacy, consumer protection, human rights and negligence law still does. The Air Canada chatbot ruling is the clearest example: in February 2024 the British Columbia Civil Resolution Tribunal held the airline liable for negligent misrepresentation by its website chatbot and rejected the argument that the chatbot was responsible for its own statements.

Key obligations

Quebec: decisions based exclusively on automated processing

• Notice (s.12.1): An enterprise that makes a decision based exclusively on automated processing of personal information must inform the person no later than when it informs them of the decision.
• Explanation on request: It must disclose the personal information used, the reasons and principal factors that led to the decision, and the person's right to have the information corrected.
• Human review: The person must be able to submit observations to a member of staff who is in a position to review the decision.
• Profiling technology (s.8.1): Before collecting personal information using technology that can identify, locate or profile a person, the enterprise must inform them of its use and of the means to activate those functions.
• Privacy impact assessment (s.3.3): Required for any project to acquire, develop or overhaul an information system or electronic service delivery system involving personal information, which covers most AI deployments. See the DPIA Guide.

Federal institutions: Directive on Automated Decision-Making

• Algorithmic Impact Assessment (AIA): Mandatory. The questionnaire has 65 risk questions and 41 mitigation questions; it is completed at the start of design and again before production, reviewed after changes in functionality or scope, and final results are published on the Open Government Portal in both official languages.
• Impact levels: Level I (little to no impact, 0 to 25 percent of the maximum score), Level II (moderate, 26 to 50), Level III (high, 51 to 75) and Level IV (very high, 76 to 100). Mitigation does not lower the level; only a new AIA with changed project facts does.
• Requirements by level: Appendix C of the Directive scales peer review and human involvement in decisions with the impact level. Baseline measures at every level include consulting legal services before development, testing and monitoring outcomes, and providing recourse.
• 2023 and 2025 amendments: The 2023 amendment added explanation, bias testing, data governance, GBA Plus and peer review measures. The 2025 amendment aimed to increase transparency, strengthen senior-level accountability and enhance testing and monitoring. A Guide to Peer Review of Automated Decision Systems (January 2025) supports subsection 6.3.5.

Voluntary Code of Conduct (ISED, September 2023)

The code lists 46 signatories. It creates no legal obligation, but it is the closest thing Canada has to a private sector AI baseline and customers use it in questionnaires. See the AI Red Teaming Guide for the testing measures.

Ontario public sector

Schedule 1 of Bill 194, the Enhancing Digital Security and Trust Act, 2024, lets the province require prescribed public sector entities to publish information about their AI use, develop an accountability framework, take prescribed risk management steps, avoid prohibited uses, and in prescribed cases disclose information and ensure a person oversees the AI system. These duties bite only once regulations prescribe them. The two regulations in effect from 1 July 2026, O. Reg. 51/26 (cyber security) and O. Reg. 52/26 (digital technology affecting individuals under 18), are not AI regulations. Ontario ministries and provincial agencies follow the province's Responsible Use of AI Directive.

Timeline and current status

Watch three things: whether Bill C-36 reaches second reading and committee, whether the federal government brings forward a separate AI bill (none had been tabled at October 2026), and whether Ontario makes AI regulations under Schedule 1. Even if C-36 passes, compliance obligations may take years to start.

Enforcement and penalties

For most private sector organisations the realistic exposure is privacy enforcement, civil liability for what AI tells customers, and failing a public sector buyer's assessment, not an AI-specific fine.

What it means for AI agents and non-human identities

The Treasury Board's agentic AI guide is one of the few government documents anywhere that turns agent governance into identity controls. Federal teams must follow it; everyone else can use it as a ready-made baseline that a Canadian public sector buyer will recognise.

• Unique agent ID and labelled permission level: The guide calls for each agent to have a unique ID and a labelled permission level such as "draft only", with read-only access by default. In identity terms that is one workload identity per agent (NIST SP 800-53 IA-4, IA-9) and least privilege (AC-6). See the Agentic AI Identity Guide.
• Human checkpoints before state changes: Review or confirmation before sending, publishing or updating records. This is also how a Quebec enterprise keeps an agent's decision from being "based exclusively on automated processing", though only if the human review is real.
• Accountable owner: Each agent needs a named owner and escalation path, and agents are paused or reassigned when owners leave. That is joiner, mover and leaver for non-human identities. See the Identity Governance for AI Agents Guide.
• Logs the agent cannot modify: Tool calls, actions, approvals and key inputs and outputs logged in a system outside the agent's control (AU-2, AU-9), kept in line with records and litigation-hold rules. These logs are also how you answer a Quebec s.12.1 request for the principal factors behind a decision.
• External kill switch: A way to stop the agent that does not depend on the agent, which in practice means revoking or suspending its credentials centrally.
• Untrusted content as data: External text must never be treated as instructions, the control OWASP describes under LLM01:2025 Prompt Injection.
• Directive check: Before using agentic AI, federal teams must determine whether the use case falls under the Directive on Automated Decision-Making, which would trigger an AIA.

Compliance steps

1. Map your layer. For each AI system record whether it serves federal institutions, Ontario public sector entities, people in Quebec or the private sector elsewhere. One system can sit in several layers.
2. Find exclusively automated decisions. Identify decisions about people in Quebec that no person reviews before they take effect. Add the s.12.1 notice, an explanation process that can name the personal information and principal factors, and a staff reviewer route.
3. Run privacy impact assessments. Use the Quebec s.3.3 assessment for any new or overhauled system handling personal information; use the AI Impact Assessment Template to cover non-privacy harms.
4. Prepare for federal procurement. If you sell automated decision systems to federal institutions, be ready to supply information for the AIA questionnaire, peer review and testing evidence proportionate to the likely impact level.
5. Adopt the agentic controls. Give every agent a unique identity, a permission label, an owner, human checkpoints for state-changing actions, tamper-resistant logs and an external kill switch. Pilot on one low-risk agent in a controlled environment, then extend.
6. Benchmark against the Voluntary Code. Even if you are not a signatory, use its six outcomes as the control set for generative AI, starting with identifying AI systems that could be mistaken for humans.
7. Control customer-facing answers. Draw agent answers from approved policy sources, limit what agents may commit to, and retain conversation logs, the lessons the Air Canada ruling points to.
8. Track change. Assign someone to follow Bill C-36, any new federal AI bill and Ontario regulations, and review this plan each quarter.

Small private sector organisations outside Quebec can usually focus on steps 6 and 7. Large organisations operating nationally, or selling to governments, need all eight, and should design for the strictest layer (Quebec and the federal Directive) once rather than per province.

Detection and measurement

• Automated decision coverage: Share of decisions about people in Quebec that are either reviewed by a person or carry the s.12.1 notice and review route.
• Explanation turnaround: Time to answer a s.12.1 request with the personal information used and principal factors.
• Agent identity hygiene: Share of agents with a unique identity, permission label, named owner and logs held outside the agent.
• Kill switch test rate: Share of production agents whose credential suspension was tested in the last quarter.
• Assessment currency: Share of AI systems with a privacy or algorithmic impact assessment completed or updated since their last material change.

Common failure patterns

• Reading "no AI law" as "no rules". Privacy, consumer and negligence law apply, as Air Canada learned when its chatbot invented a refund policy.
• Rubber-stamp human review. A reviewer who approves every automated outcome without the information to challenge it does not make a decision less automated in substance, and will not satisfy a regulator asking how the decision was made.
• Waiting for AIDA's successor. Teams that paused AI governance in 2025 now face Quebec, federal procurement and customer questionnaires without evidence.
• Shared credentials across agents. Without per-agent identity, logs cannot show which agent took an action, so neither the agentic guide's accountability nor a s.12.1 explanation is possible.
• Kill switch inside the agent. A stop command the agent itself must obey fails exactly when the agent misbehaves; revoke credentials externally.

Framework mapping

Practitioner checklist

• Record, for every AI system, which Canadian layer applies: federal, Ontario public sector, Quebec or general private sector.
• Find decisions about people in Quebec made exclusively by automated processing and add notice, explanation and staff review.
• Complete a privacy impact assessment for each new or overhauled AI system handling personal information.
• Prepare AIA, peer review and testing evidence before bidding for federal automated decision work.
• Give each agent a unique identity, permission label, named owner and read-only default access.
• Require human confirmation before agents send, publish or update records.
• Log agent actions outside the agent's control and test an external kill switch each quarter.
• Use the Voluntary Code's six outcomes as the generative AI control baseline.
• Track Bill C-36, any new federal AI bill and Ontario AI regulations.

Standards and references

• Parliament of Canada LEGISinfo: Bill C-36, Protecting Privacy and Consumer Data Act
• Government of Canada: Algorithmic Impact Assessment tool
• Treasury Board of Canada Secretariat: Guide on the Use of Agentic Artificial Intelligence
• Government of Canada: Progress on AI in government
• LégisQuébec: Act respecting the protection of personal information in the private sector
• ISED: Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems
• Office of the Privacy Commissioner of Canada: Principles for responsible, trustworthy and privacy-protective generative AI technologies
• Legislative Assembly of Ontario: Bill 194, Strengthening Cyber Security and Building Trust in the Public Sector Act, 2024

This guide summarises Canada's federal and provincial AI rules and guidance as at October 2026 and is not legal advice. Related NHI Mgmt Group resources: Global AI Regulation Map · Canada PIPEDA and Bill C-27 Guide · UK AI Regulation Guide · South Korea AI Basic Act Guide