PIPEDA consent and AI agents

Meaningful consent under PIPEDA means a customer understands what an AI agent collects, why, who receives it, and the risks. Refresh it when an agent takes on a new purpose, a new kind of data, new inferences, or a new party using data for its own ends. A new supplier for the same purpose usually needs transparency, not fresh consent.

What does PIPEDA require on consent?

PIPEDA's third fair information principle requires knowledge and consent for the collection, use and disclosure of personal information, except where the law provides otherwise. Consent is valid only if the person would reasonably understand the nature, purpose and consequences of what they agree to.

Guidance from the federal privacy commissioner on meaningful consent, in substance, asks organisations to:

• emphasise the key elements: what information is collected, which parties receive it, for what purposes, and any meaningful risk of harm;
• let people make choices, with a clear yes or no for anything that is not necessary to the service;
• present information in a way people can understand and use, at the right moment;
• treat consent as an ongoing process, revisited as practices change;
• obtain express consent where information is sensitive, where the use would fall outside reasonable expectations, or where it creates a meaningful residual risk of significant harm.

People can withdraw consent on reasonable notice, subject to legal and contractual limits. For the full set of principles applied to agents, see PIPEDA and AI agents handling customer data.

Who does it apply to?

Private-sector organisations handling personal information in commercial activity in Canada, and federally regulated businesses such as banks, except where a substantially similar provincial law applies to intra-provincial activity. Our short answer on how PIPEDA applies to AI agents sets out the scope.

What does meaningful consent mean for an agent?

The test is reasonable expectation. A customer who opens a chat to change their address expects their address to be read and updated. They may not expect that the same conversation is used to score their likelihood of leaving, that a voice recording is kept, or that their messages go to a model provider overseas. The agent does not change the rules; it makes it easier to drift outside what people expected.

Three agent features need particular care.

Agents infer. An agent reading a customer's messages may conclude they are in financial difficulty or are vulnerable. Creating that inference is a collection of new, often sensitive, personal information. If it is used beyond the immediate conversation, consider whether it falls outside expectations and needs express consent, or whether it should be used only to route the customer to help.

Agents widen use by default. Data an agent sees for one capability is available, technically, for others. Consent tied to purposes only works if each capability is tied to a purpose and the agent cannot use data outside it.

Agents involve third parties. Prompts go to a model provider. Transfer to a service provider for processing on your behalf, for the purpose already consented to, is generally treated as a use rather than a disclosure. You remain accountable and should be open about it, but you do not usually need fresh consent for the transfer itself. Sharing with a party for its own purposes is different.

When must consent be refreshed?

Refresh consent, or at least re-examine it, when any of these happen:

• a capability is added that serves a new purpose, such as marketing from service conversations;
• the agent begins handling a new kind of information, such as voice recordings, documents with health data, or biometric checks;
• the agent starts creating new inferences that will be kept or used beyond the conversation;
• information will be shared with a party that will use it for its own purposes;
• the risk of harm materially changes.

Recording is a common trigger. Our answer on whether a support agent should ask before recording a conversation applies the same reasoning.

Take an illustrative case. A Canadian credit union's chat agent answers account questions under existing consent. The credit union then wants the agent to suggest products based on what members discuss. That is a new purpose outside reasonable expectations for a service chat. It adds a clear opt-in at the start of chats, records each member's choice, and configures the agent so the product suggestion capability only runs for members who opted in.

How do you meet PIPEDA's consent requirements for agents?

What will the privacy commissioner or an auditor ask?

• What did customers understand when they consented to the agent's processing?
• Which agent capabilities rely on express consent, and how is it recorded?
• What inferences does the agent make about customers, and how are they used?
• How does a customer withdraw consent, and how does the agent know?
• When did you last review consent after changing an agent?

What does PIPEDA consent not cover?

Consent does not cure excessive collection: PIPEDA still limits collection to what purposes need, even with consent. It does not settle security, accuracy or retention. India's law takes a more prescriptive line on notice and consent managers; see DPDP Act consent and AI agents for a comparison. For the full regulatory 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.