Fully, because an agent handling personal information is the organisation handling it. PIPEDA, Canada's federal private-sector privacy law, governs collection, use and disclosure in commercial activity through fair information principles, with consent and accountability at the centre. Each agent's purposes, consent basis, data limits and safeguards must hold up.
What does PIPEDA require, in substance?
PIPEDA is built on principles that, in substance, require organisations to:
- Be accountable for personal information under their control, including information passed to a third party for processing, using contracts to protect it.
- Identify purposes before or at the time of collection.
- Obtain meaningful consent, express or implied depending on sensitivity and reasonable expectations.
- Limit collection, use, disclosure and retention to what the purposes need.
- Keep information accurate and protect it with safeguards suited to its sensitivity.
- Be open about practices, give people access to their information, and let them challenge compliance.
Breaches that create a real risk of significant harm must be reported to the federal privacy commissioner and the affected people. Some provinces have their own private-sector privacy laws that apply instead in certain cases. Consent is unpacked in PIPEDA consent and AI agents.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where do agents test the principles hardest?
Purposes and consent. Customers consented to their data being used to service their account. An agent that also uses the same data to suggest products, or to decide eligibility for something new, may be outside what they reasonably expected. Each agent capability needs a purpose the customer would recognise.
Limiting use. Agents tend to read widely because more context often improves answers. PIPEDA pushes the other way: the agent should only reach the information its purpose needs.
Retention. Agents create new stores of personal information: memory, summaries, logs. Each needs a retention period tied to its purpose and deletion that works. Applying schedules to agent-held records is covered in how to apply record retention schedules with an AI agent.
Accountability for providers. If the agent's model runs with a provider, the organisation remains accountable and needs contractual protection, including where the provider processes data outside Canada.
Concrete examples: what may the agent do, and what stays with a person?
| Activity | Agent | Person |
|---|---|---|
| Use account data to answer the customer's own questions | Yes | Disputes |
| Use account data to target a new product | Only if purpose and consent cover it | Decides whether consent is adequate |
| Respond to an access request | Gathers data from all stores, including agent memory | Reviews and sends the response |
| Disclose information to a third party | Only through approved channels | Anything outside them |
| Assess a privacy breach involving the agent | Drafts facts and timeline | Decides whether harm is significant and reports |
An illustrative example
Take a Canadian online insurer that uses an agent to handle policy changes by chat. To help with renewals, the team gives the agent access to past claims and lets it remember conversations.
A privacy review asks three PIPEDA questions. Purpose: customers expect claims data to be used for claims and pricing, not surfaced in a chat about changing their address. Retention: chat memories had no deletion rule. Accountability: the model provider's contract did not address where data was processed. The insurer limits the agent's access to policy data for routine changes, sets a short retention period for chat memory, adds agent memory to its access request process, and updates the provider contract. Customer-data handling in more depth is covered in PIPEDA and AI agents handling customer data.
Which controls meet PIPEDA for agents?
- Purpose per capability, recorded and visible in privacy notices where needed.
- Source permissions limited to each purpose.
- Retention rules for agent memory, summaries and logs.
- Access and correction processes that reach every agent store.
- Provider contracts that keep the organisation accountable, including for cross-border processing.
- Breach assessment that includes agent actions, such as misdirected messages.
South Africa's approach makes an interesting comparison; see whether POPIA applies to AI agents. Canadian firms with EU customers should also check whether the EU AI Act applies to AI agents. The full map is in AI agent regulation, mapped to controls.
Questions people ask
Do customers have to consent to us using AI agents?
Not to the technology as such. They must consent to the purposes for which their information is used. If the agent uses information for a new purpose, consent for that purpose is the question.
Does PIPEDA give a right to human review of agent decisions?
PIPEDA gives rights of access and to challenge compliance, but not a specific right like the EU's on automated decisions. Some provincial laws go further; check those that apply.
Is sending data to a model provider a disclosure?
Transferring information to a provider for processing on your behalf is treated as a use by your organisation, provided you remain accountable and protect it by contract.