PIPEDA treats an AI agent the way it treats any process that handles personal information: the organisation stays accountable for every collection, use and disclosure the agent makes. Its ten fair information principles apply in full. The agent-specific work is limiting what it collects and sees, keeping its purposes tied to consent, and securing what it sends to model providers.
What does PIPEDA require, in plain words?
PIPEDA, Canada's federal private-sector privacy law, sets out ten fair information principles:
- Accountability: an organisation is responsible for personal information under its control, including information passed to others for processing, and names someone accountable.
- Identifying purposes: purposes are identified at or before collection.
- Consent: knowledge and consent are required, except where the law says otherwise.
- Limiting collection: collect only what the identified purposes need.
- Limiting use, disclosure and retention: use and disclose only for those purposes; keep only as long as needed.
- Accuracy: information is accurate, complete and up to date enough for its use.
- Safeguards: security appropriate to the sensitivity.
- Openness: policies and practices are available.
- Individual access: people can access their information and challenge its accuracy.
- Challenging compliance: people can complain about compliance.
PIPEDA also requires organisations to report breaches of security safeguards that create a real risk of significant harm, notify affected people, and keep records of all breaches. For a short answer on scope, see how PIPEDA applies to AI agents.
Who does it apply to?
Private-sector organisations collecting, using or disclosing personal information in the course of commercial activity in Canada, except where a province has a substantially similar law covering activity within that province. It also applies to federally regulated businesses, such as banks, across the country. Cross-border processing does not remove PIPEDA duties; the Canadian organisation stays accountable.
What does PIPEDA mean when the system is an agent?
Four principles change most in practice.
Accountability reaches the model provider. An agent sends prompts, which often contain personal information, to a model provider, sometimes in another country. PIPEDA allows transfers for processing but keeps the organisation accountable. That means contracts that give comparable protection, limits on the provider's own use of the data, and telling customers that their information may be processed outside Canada.
Limiting collection applies to what the agent pulls in. An agent that retrieves a customer's whole file to answer a question about one payment collects more than the purpose needs. Scope retrieval to the fields the task requires.
Limiting use applies to new purposes the agent creates. An agent built to answer service queries should not start profiling customers for marketing because the data happens to be in front of it. Each capability should be tied to an identified purpose, and new purposes need fresh consideration of consent. Our companion piece on PIPEDA consent and AI agents covers when consent must be refreshed.
Accuracy matters more when the agent acts. An agent that updates records based on a misread email spreads inaccuracy into every system that relies on them. Changes to personal information should rest on reliable sources, with the source recorded.
Individual access also has an agent angle. If an agent made or contributed to a decision about someone, the record of what information it used helps you answer an access request and explain the outcome.
Take an illustrative case. A Canadian insurer's agent helps customers update their policy details by chat. Its privacy review limits the agent's view to policy fields needed for updates, blocks access to claims history, records the source of each change, and confirms the model provider contract bars use of customer data for model training. The privacy notice is updated to say an AI agent handles chat and that data may be processed outside Canada.
How do you meet PIPEDA for agents?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Accountability | Named privacy owner per agent; provider contracts with comparable protection | Accountability register; contracts |
| Identifying purposes | Each agent capability tied to a stated purpose | Purpose map per capability |
| Consent | Consent basis recorded per purpose; new purposes reviewed | Consent records; review notes |
| Limiting collection | Retrieval limited to fields the task needs | Data map; scoped access configuration |
| Limiting use and retention | Agent memory and logs kept only as long as needed | Retention schedule; deletion records |
| Accuracy | Record changes only from reliable sources, with source recorded | Change log with source labels |
| Safeguards | Least privilege, output checks, protection of prompts and logs | Security configuration; test results |
| Openness | Privacy notice explains agent use and cross-border processing | Published notice |
| Access and challenge | Records of what data the agent used per decision | Retrievable action records |
| Breach reporting | Agent incidents assessed for real risk of significant harm | Breach register |
What will the privacy commissioner or an auditor ask?
- What personal information does each agent collect, use and disclose, and for what purpose?
- Where is it processed, and what protects it outside Canada?
- How do you stop an agent using data for purposes customers did not expect?
- How would you answer an access request about a decision an agent contributed to?
- How are agent incidents assessed against the breach reporting test?
What does PIPEDA not cover?
It does not regulate AI as such, and Canada has no federal AI act in force. It does not cover public-sector bodies or, for intra-provincial activity, provinces with their own private-sector laws. For financial institutions, it sits beside OSFI's model risk and technology expectations. Firms operating in the US face a different structure; see GLBA and AI agents handling customer financial data and whether GLBA applies to agents. For the full map, see our guide to AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Does PIPEDA require customers to be told an AI agent is handling their data?
PIPEDA requires openness about policies and practices and consent based on understanding. Where an agent's processing would not be obvious to a customer, explaining it is part of meaningful consent and openness.
Can a Canadian firm use a model provider outside Canada?
PIPEDA does not prohibit it. The firm stays accountable, must use contracts or other means to provide comparable protection, and should be transparent that information may be processed abroad.
Are agent logs personal information under PIPEDA?
If they contain information about an identifiable individual, yes. They need purposes, retention limits and safeguards like any other record, and may fall within access requests.