This is a plain-language orientation, not legal advice. If you handle health information, financial records or anything sensitive at scale, talk to a privacy lawyer.
Most Ontario businesses are covered by PIPEDA — the federal private-sector privacy law — for the personal information they collect in the course of commercial activity. Customer names, emails, phone numbers, addresses, purchase history. That is personal information, and pointing an AI tool at it does not change what the law expects of you.
The good news is that PIPEDA is principles-based rather than prescriptive. It does not ban AI. It asks you to be able to answer some reasonable questions.
1. Where does the data physically go?
When you paste a customer list into a free AI tool, that data leaves your control and lands on servers somewhere. PIPEDA does not prohibit cross-border transfer, but it does hold you responsible for the information regardless of who is processing it, and it expects you to be transparent about it.
Practically: know which country your provider’s servers are in, and prefer Canadian residency when you can get it. It is a simpler story to tell a customer who asks.
2. Is your data being used to train someone’s model?
This is the question that catches people out. Consumer AI products have historically defaulted to using inputs for model improvement; business tiers usually do not. The difference matters, because “we used customer records to train a third party’s product” is a hard sentence to defend.
Ask for it in writing. Any serious vendor will confirm it in the contract without hesitation.
3. Did you collect this for the purpose you are now using it for?
PIPEDA cares about purpose. If a customer gave you their phone number so you could deliver their order, using it for delivery updates is consistent with that. Feeding it into a system that scores them for an unrelated marketing campaign is a different purpose, and may need fresh consent.
The test is not “is it technically possible” but “would this customer be surprised?” If they would, do not do it without asking.
4. Are you keeping only what you need?
Minimisation is a real obligation and it also happens to be good engineering. An agent that helps you chase an overdue invoice needs the client name, the amount and the payment history. It does not need their date of birth, their full purchase history from 2019, or anything you have not looked at in three years.
Narrower data access also produces better output, because there is less noise for the model to get distracted by.
5. Can you explain a decision after the fact?
If an automated system did something a customer objects to, you should be able to reconstruct why: what data it saw, what it concluded, what it did. That means logging, and it means keeping a human in the loop for anything consequential.
This is why we default every Orbys agent to supervised mode. Beyond the compliance argument, there is a simpler one: a person clicking approve is a person who saw it.
A short checklist before you connect anything
- Data hosted in Canada, confirmed in writing
- Contractual guarantee that your data is not used for training
- Documented list of what the system can read and what it cannot
- A human approval step on anything that reaches a customer
- An audit log you can actually query
- A written exit: your data back, their copy deleted
- Your privacy policy updated to reflect that you use automated processing
None of this is exotic. It is the same diligence you would apply to any vendor holding your customer records — which, if you use accounting software or a CRM, you have already done at least once.
Orbys hosts client data in Canada, never uses it for training, and keeps a human approval gate on every outbound action by default. The specifics are in our privacy policy.