Canada regulates AI agents through existing institutions, because no federal AI act is in force. OSFI supervises federally regulated financial institutions and expects model risk to be managed under Guideline E-23. FCAC enforces consumer protection for banks. Provincial securities regulators oversee dealers and advisers. PIPEDA governs personal information in the private sector, with some provinces running their own privacy laws.
What do the rules require, in plain words?
OSFI Guideline E-23. OSFI's model risk guideline expects federally regulated institutions to manage the risk of models throughout their life cycle: an inventory, risk rating, development standards, independent validation, approval, monitoring and decommissioning. OSFI has made clear that AI and machine learning methods are within its view of model risk. Its separate expectations on third-party risk and on technology and cyber risk also apply when an agent runs on an external platform.
FCAC. The Financial Consumer Agency of Canada supervises federally regulated financial institutions' compliance with consumer protection obligations, including fair treatment, clear disclosure and complaint handling.
Provincial securities regulators. Registered dealers and advisers must meet know-your-client, suitability and conflict of interest duties, and keep adequate records. Regulators coordinate nationally but supervise provincially.
PIPEDA. Federal private-sector privacy law built on fair information principles: accountability, identifying purposes, consent, limiting collection, limiting use, disclosure and retention, accuracy, safeguards, openness, individual access and challenging compliance. Some provinces have substantially similar laws that apply instead within the province.
Who does it apply to?
E-23 applies to OSFI-regulated banks, insurers and trust companies. FCAC's remit covers banks and other federally regulated financial entities. Securities rules apply to registrants in each province. PIPEDA applies to commercial activity across most of Canada outside provinces with their own private-sector law, and to cross-border transfers.
What does this mean when the system is an agent that acts?
An agent is a model, and often several. Under E-23, the question is not whether the agent is "AI" but whether it produces outputs that drive decisions. An agent that triages fraud alerts, drafts credit memos or recommends portfolio changes does. It belongs in the model inventory with a risk rating, and its validation must cover the whole agent: the model, its instructions, its data sources and the actions it can take.
Consumer outcomes are the institution's. If an agent gives a customer wrong information about fees or misdirects a complaint, FCAC treats that as the bank's failure. Complaint handling is a frequent pressure point: an agent must recognise a complaint and route it under the bank's procedure.
Suitability cannot be delegated to an agent's judgement alone. Securities registrants remain responsible for suitability. An agent can gather know-your-client information and draft recommendations; a registered individual approves them.
Consent under PIPEDA is purpose-specific. Using customer data in an agent for a purpose not identified at collection generally needs new consent, unless an exception applies.
For how payments firms apply this, see AI agents for compliance teams in payments in Canada and AI agents for operations leaders in payments in Canada.
How do you meet the rules?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| E-23: inventory and risk rating | Each agent capability recorded with its model components, data and actions | Model inventory entries |
| E-23: validation | Agent tested on cases from your own operation before use and on change | Validation reports |
| E-23: monitoring | Outcome monitoring with thresholds that trigger review or withdrawal | Monitoring dashboards; incident records |
| Third-party risk | Agent platform assessed, contracted and monitored | Due diligence; contract; reviews |
| FCAC: fair treatment and disclosure | Agent content on fees and terms approved; complaints routed | Approved content library; complaint logs |
| Securities: suitability | Recommendations drafted by agent, approved by a registrant | Approval records |
| PIPEDA: purposes and consent | Each capability tied to an identified purpose; consent where needed | Purpose register; consent records |
| PIPEDA: safeguards and retention | Scoped access; retention for agent memory and logs | Access maps; deletion logs |
What will OSFI, FCAC or a securities regulator ask?
- Is each agent in your model inventory, and how did you rate it?
- Who validated it independently, and on what data?
- What are the agent's limits, and what happens when it reaches one?
- How are complaints raised to an agent identified and handled?
- Who approves recommendations an agent drafts for clients?
- Which purposes did customers consent to, and does the agent stay within them?
What do these rules not cover?
There is no federal AI act in force, so there is no registration of AI systems and no list of prohibited AI uses at federal level. Provincial rules can differ, especially for privacy, and some public sector bodies face their own directives on automated decisions. Canadian firms serving EU customers face the EU's rules too, as set out in how AI agents are regulated in the European Union; for another common-law comparison, see how AI agents are regulated in South Africa. For the full control map, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
E-23 asks for validation and monitoring of each model. Pryme's Attest certifies each agent by replaying it against real cases from your own operation, scored across seven gates. Certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day, which gives the inventory a current, evidenced status rather than a one-off approval.
Questions people ask
Does OSFI expect agents to be validated like credit models?
The principle is the same: independent, proportionate to risk, and repeated on change. The method differs, because agents produce language and actions rather than a score. Validation should test the actions as well as the outputs.
Does PIPEDA require telling customers an AI agent is involved?
PIPEDA's openness and purpose principles require you to explain how personal information is used. If an agent changes how data is used or makes decisions about people, explaining that is part of meeting those principles.
Is a provincial privacy law stricter than PIPEDA for agents?
Some provinces have added specific duties on automated decisions and transparency. If you operate in such a province, check its law as well as PIPEDA, because the provincial law may apply instead.