Probably yes. PIPEDA applies to a Canadian business that collects, uses or discloses personal information in the course of commercial activity, and typing customer details into ChatGPT or Claude is a use and often a disclosure. Check first whether you need the details at all, then consent, the vendor's data terms and retention. Not legal advice.
Two caveats matter immediately. Alberta, British Columbia and Quebec have their own private-sector privacy laws, which the Office of the Privacy Commissioner calls substantially similar, and a business covered by one of them is generally exempt from PIPEDA for information handled within that province. Quebec's law added extra duties, covered near the end. Confirm which regime applies to you before relying on anything below.
What counts as personal information when you only ask a question
Owners often assume a chat tool is like a search engine: you type a question and it goes nowhere. It is closer to emailing a contractor, because the text is sent to a company's servers, may be stored for a period and is processed by software you do not control. Personal information is information about an identifiable person, and the test is whether someone could be identified from it, not whether you typed a name.
A paragraph of a broker's client email with the surname removed but the street, policy number and a distinctive claim still identifies a person. A plumber's job note saying "customer at number 12 has a leak under the kitchen sink" with a first name and number may identify a household. A list of overdue invoice amounts with no names is much safer. The practical habit is to ask, before pasting: could a colleague work out who this is from what I am about to send?
The OPC's own advice on generative AI
The Office of the Privacy Commissioner of Canada and provincial counterparts published principles for generative AI in December 2023. For organisations using these tools, they say to use only tools that respect privacy law, to use anonymised or de-identified information in prompts where possible, to enter personal information into a prompt only where authorised, to make sure people using a public-facing tool know they are dealing with generative AI, and, unless otherwise required, not to retain prompts, use them for other purposes or disclose them. The document says the principles do not set out every compliance requirement, so read the page itself on priv.gc.ca.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThe seven checks, in the order I would run them
The checks below follow seven of the ten fair information principles that the OPC publishes for PIPEDA: accountability, identifying purposes, consent, limiting collection, limiting use, disclosure and retention, safeguards and openness. They are practical translations, not a legal opinion.
| Check | The question to ask | Evidence to keep |
|---|---|---|
| 1. Need | Does the task need personal information, or would a placeholder do? | A one-line note per task type |
| 2. Purpose and consent | Is AI processing within what customers were told when they gave their details? | Privacy notice wording and its date |
| 3. The plan you are on | Is this a business plan with data terms, or a personal account? | Plan name, terms page, screenshot of settings |
| 4. Training and retention | Is training on your content off, and how long are chats kept? | Settings screenshot, vendor retention statement |
| 5. Where the data goes | Which sub-processors and storage regions does the vendor name? | Vendor data-processing terms |
| 6. Safeguards | Who has access, with what sign-in protection, and what can staff paste? | A short written rule and the access list |
| 7. Breach handling | If something is pasted wrongly, who decides what happens next? | A named person and a breach log |
Each check is cheap on its own. Together they take a small firm about half a day, and the written answers are what you would show a customer or the regulator who asks.
Check one: do you need the personal details at all
This is the check that removes most risk. Many AI tasks need the shape of the problem, not the identity of the person. Drafting a polite reply to a complaint needs the complaint's content and the facts you will offer; it does not need the customer's surname or address.
A redaction habit, with a before and after
An illustrative case: a four-person insurance brokerage wants help summarising a client email. The original includes a full name, a postal address, a policy number and the line "my husband was diagnosed in March". The broker rewrites it before pasting.
BEFORE (do not paste)
Hi, it's [full name] at [full address], policy [number]. My husband was diagnosed in March and we may need to change the cover...
AFTER (safe to paste)
Hi, I'm a client with a household policy. A family member has a new health diagnosis and we may need to change the cover. Please summarise the question in two lines and list what information I need to ask for.
The summary comes back usable and the identity stays in the broker's own system. Health details are sensitive information, and the safeguard should match that sensitivity, so the second version also leaves out the diagnosis date. The rule: replace names, addresses, numbers and dates with generic words, and add them back in your own document.
Check two: purpose and consent
PIPEDA expects you to say why you collect personal information and to have the person's knowledge and consent for how you use it. The OPC's principle page also says any collection, use or disclosure should serve purposes a reasonable person would consider appropriate. Ask what your customers were told. If your form said "we use your details to provide your quote and contact you", then using an AI tool to draft that quote's email arguably falls within the purpose, but using the same details to train a model, or sending them to a vendor that keeps them for its own use, may not.
Wording a notice might use
An illustrative paragraph for a privacy notice, which an adviser should check before use: "We use software, including AI writing tools from third-party providers, to help prepare replies, summaries and documents. We limit the personal information we put into these tools to what the task needs, and we choose providers whose terms prevent them from using our business content to train their models. You can ask us how your information is handled by contacting [your privacy contact]." Put the real contact in yourself. The OPC's openness principle expects policies to be easy to find, so link it from forms and the footer.
Checks three and four: plan, training setting and retention
Vendors differ here, and the differences are the main reason one customer-data task is acceptable and another is not. As checked on the vendors' pages in November 2026, the pattern is consistent: consumer plans of ChatGPT and Claude let users switch off the use of their chats to improve models, whereas business plans (ChatGPT Business and Enterprise, Claude Team and Enterprise, as well as Microsoft 365 Copilot and Gemini in Workspace) do not train on business content by default.
Retention differs too. Anthropic's consumer-terms update keeps a 30-day retention period for users who decline training and extends it to five years for those who allow it, while OpenAI says it may keep a copy of a ChatGPT Temporary Chat for up to 30 days for safety purposes. Both are examples that change; read the vendor's current page and screenshot it with the date. Business plans have their own terms. On OpenAI's pricing comparison, choosing the region where data is stored (data residency) is listed for ChatGPT Enterprise, not Business, so ask before you assume where a small-business workspace keeps your data.
A small firm that prefers business plans meets two practical rules: ChatGPT Business and Claude Team both require at least two seats, so a one-person firm pays for two or uses an individual plan with training off. For help choosing, see who needs a paid AI licence and what to check in an AI tool's privacy policy and terms.
A short vendor questionnaire
- Is customer content used to train or improve your models, by default or ever?
- How long do you keep prompts and outputs, and can we delete them?
- Which sub-processors and regions handle the data?
- Who at the vendor can read our content, and in what circumstances?
- How do you notify us of a security incident?
Business plans usually answer these in a data-processing document. If the vendor answers vaguely or only by marketing copy, treat the tool as unsuitable for customer data. The questions to ask an AI vendor tutorial extends the list.
Check five: where the data goes, and connected apps
Pasting is only one route. ChatGPT's connected apps (formerly connectors) and similar features can read a drive or an inbox, which turns a harmless-looking convenience into access to every customer email. Before connecting, decide which folders it should see, and test with a folder that has nothing sensitive. What connectors can see in your drive and inbox goes through this. The same goes for browser agents: for Claude in Chrome, Anthropic's safety article strongly advises against using it to manage financial accounts, handle legal documents, process medical or health information or interact with sites containing other people's personal information.
Check six: safeguards your staff can actually follow
Safeguards under PIPEDA must be proportionate to sensitivity. For a small firm that means a short written rule, a named approved tool, protected sign-ins and a habit of not pasting certain things. An illustrative one-page rule for a plumbing firm:
AI USE RULE (draft for adviser review)
1. Approved tool: [business plan name]. Personal accounts are not allowed for customer work.
2. Never paste: full address, phone number, email address, payment details, photos of keys or alarm codes, or access notes.
3. Use placeholders: [first name], [street], [job number].
4. A person reads everything before it is sent to a customer.
5. If you paste something you should not have, tell [named person] the same day. Do not delete the evidence first.
Rule five matters more than it looks. People hide mistakes when they fear the response, and an owner who finds out three weeks late has lost the chance to act. For the review step, how to set up human review for AI work shows a light version.
Check seven: what to do when something goes wrong
PIPEDA requires organisations to report breaches of security safeguards to the Privacy Commissioner and notify the individuals when it is reasonable to believe the breach creates a real risk of significant harm, and to keep a record of every breach of security safeguards for 24 months whether or not it is reported. The OPC's guidance on mandatory breach reporting sets this out, and the OPC offers an online tool that walks you through the real-risk assessment. Whether a wrong paste to a vendor counts as a breach depends on the facts, such as sensitivity, who could see it and the probability of misuse.
A breach log entry, filled in
Date found: [date]
What happened: Staff member pasted a customer's full name, address and a medical-equipment note into a free personal chat account.
Who: [role], [number] customer(s) affected
Action: chat deleted; training switched off in the account; staff reminded of rule 2
Assessment: sensitivity medium-high; vendor retention up to [period]; probability of misuse judged low; reasons: [short]
Decision: [report / not report], decided by [name] on [date]
Review: rule updated; business plan bought for the team
Notice that the log records the reasoning, not only the outcome. When the OPC inspected the breach records kept by seven telecommunications companies, it found that 40 percent of the sampled records did not contain enough information for it to assess whether the breach created a real risk of significant harm. Write down why you decided as you did.
The same checks on three more jobs, in short
Once you have the table, each new idea takes ten minutes to pass or fail. Three quick illustrations, all invented.
A roofing company drafting survey reports
The report needs the roof's condition, measurements and photos' descriptions. It does not need the homeowner's name. Result: pass with placeholders. One flag: photos can show faces, number plates or the inside of a home through a window, so describe photos in words instead of uploading them to a consumer account.
An insurance broker summarising a call transcript
A transcript is dense with identifiers, health information and finances. Result: not on a personal account; on a business plan only after the vendor questionnaire is answered, and only with names and numbers stripped by a person first. Many brokers will decide this job is not worth the risk until a vendor can give written answers on retention.
A plumbing firm writing reminder texts
The task needs a first name, an appointment time and a service type. A template with merge fields in the booking system does it with no AI at all, and the same plumbing firm that wanted a chat tool for this job drops the idea. The best privacy decision is sometimes to use a tool that never sees the data.
Briefing staff in five minutes
A written rule is not enough on its own, so brief staff once and repeat it in a month. Here is a script you can read aloud, using plain words.
"We use [tool] for drafting and summarising. Three rules.
One: use the company account, never your own.
Two: before you paste, swap names, addresses, phone numbers and policy or job numbers for placeholders like [first name].
Three: if you paste something you should not have, tell [name] today. Nobody gets in trouble for telling us; the trouble is finding out late.
Questions? Ask [name]."
Test understanding with two real examples from your week. Show a made-up email and ask staff to mark what they would remove. If different people mark different things, the rule is too vague; add a list of what counts as identifying in your business, such as job numbers that appear on invoices.
A short message to a vendor, ready to send
If a vendor's public pages do not answer the questionnaire, ask in writing and keep the reply. An illustrative message:
Subject: Data handling questions before we use [product] with customer information
Hello,
We are a small business and would like to use [product] for drafting customer correspondence. Before we do, please confirm in writing:
1. Is content we submit used to train or improve any model?
2. How long are prompts and outputs retained, and can we delete them?
3. Which regions and sub-processors handle our content?
4. How will you notify us of a security incident, and how quickly?
Thank you, [your name]
A clear reply goes in your file. A refusal to answer, or an answer that says "see our website" when the website does not say, tells you the tool is not ready for customer data. File the date with the answer, because terms change.
A worked pass through a real-looking job
Take an illustrative electrical contractor who wants to use AI to turn inspection notes into customer letters. The checks produce this result.
- Need: the letter needs the faults found and the recommended work, not the customer's identity. Use "the customer" and add the name afterwards.
- Purpose: the privacy notice covers "preparing documents"; add the AI-tools sentence.
- Plan: two business seats; settings screenshot filed.
- Retention: vendor terms read; note in the file with the date.
- Where it goes: no connected apps enabled.
- Safeguards: one-page rule signed by all four staff.
- Breach: the office manager is the named person; the log is a shared sheet.
That took the owner roughly a day, and the first draft letters took less time than before. The compliance work is a one-off; the saving repeats.
Making the file you would show someone who asks
If a customer, an insurer or the regulator ever asks how you handle customer information in AI tools, the best answer is a single folder, not a memory. Keep five items in it: the one-page staff rule with the date staff were briefed; the plan name and a dated screenshot of the training setting; the vendor's data terms or written answers; the breach log, even if it is empty; and your privacy notice wording with the AI sentence in it. Add a line in your diary to reread the vendor's terms every six months and after any announced change.
An illustrative contents list for a four-person firm might read: "01 staff rule, signed by all four", "02 plan and settings, screenshot taken in March", "03 vendor reply, 12 February", "04 breach log, nil entries, last reviewed June", "05 privacy notice, version with AI line". Nothing here is complicated. Its value is that on a bad day you can show you thought about it before it happened, and that is what a reasonable accountability principle asks of an organisation your size.
Quebec's Law 25: the extra step for customers there
Quebec's private-sector privacy act was amended by Law 25. Section 17 of the act, as published on LegisQuebec, says that before communicating personal information outside Quebec a business must carry out a privacy impact assessment that weighs the sensitivity of the information, the purposes, the protection measures (contractual ones included) and the legal framework where the information would go. The information may be sent if the assessment shows it would receive adequate protection, and the communication must be covered by a written agreement. The same applies when a business entrusts a person or body outside Quebec with collecting, using or keeping the information on its behalf, which is the position of many cloud AI services. It is not a ban on such transfers. Whether and how it applies to a small business with some Quebec customers is a question for the guidance of the Commission d'accès à l'information or a Quebec privacy adviser. If you serve Quebec customers or employ people there, add a line to your seven-check file: "Law 25 assessed, date, by whom".
Customer data only: what these checks leave out
This tutorial deals with customer and client information. It does not cover employee records, health-sector rules for regulated professions, sector codes of practice (a regulated broker or a trade body may have its own), or intellectual-property questions about what a tool produces. If you are in a regulated trade, ask your regulator whether AI tools are addressed. And because the law and vendor terms both change, put a recheck date in the diary for six months out. If the work feels larger than a day, what an AI consultant can't do for you sets out which duties stay with the owner; a lawyer must answer the legal ones.
Where to confirm each point
- PIPEDA's principles and who it covers: the Office of the Privacy Commissioner of Canada, priv.gc.ca.
- Generative AI principles: the OPC page "Principles for responsible, trustworthy and privacy-protective generative AI technologies".
- Vendor terms: the data-use or privacy pages of the vendor, with a dated screenshot.
- Quebec: section 17 of the Act respecting the protection of personal information in the private sector, and the Commission d'accès à l'information du Québec.
Questions small Canadian businesses ask about AI and privacy
Can I paste a customer's name into ChatGPT or Claude?
Possibly, but treat a name as personal information and ask whether the task needs it. Often it does not: use a placeholder such as [first name] and add the real name after. If you do include it, use a business plan, tell customers in your privacy notice and make sure it fits the reason you collected the name.
Do I need customer consent to use AI on their data?
PIPEDA requires the person's knowledge and consent for collecting, using and disclosing personal information, except where that would be inappropriate. If AI processing falls outside what customers reasonably expect when they gave their details, you may need to tell them and ask. Check this with the Office of the Privacy Commissioner's guidance or an adviser.
Does a free chat account change the risk?
Yes. Consumer plans on both ChatGPT and Claude let users switch off use of chats for training, but business plans do not train on your content by default. For customer data, a business plan with documented terms is the safer starting point. Read the current terms before you rely on either.
What happens if staff already pasted customer data into a free account?
Find out what was pasted and by whom, delete the chats where the tool allows it, switch off training in that account and record what happened. Then decide whether it was a breach of security safeguards needing assessment. Keep a note of your reasoning either way.
Does Quebec's Law 25 change anything for a small business?
It can. Section 17 of Quebec's private-sector privacy act requires a privacy impact assessment and a written agreement before personal information is communicated outside Quebec, and the same applies when a body outside Quebec keeps or uses it for you, which covers many cloud AI services. If you have customers or staff in Quebec, take advice before using their data with AI.
Further reads
- Where Is Your Data Stored When You Use AI Tools? — See where each tool keeps what you type.
- How to Manage Your Inbox With AI: Triage, Drafts, and Follow-Ups — Handle inbox work with AI while keeping client details out.
- A Simple AI Risk Register for Small Businesses (With Template) — Keep a register of AI risks, including privacy ones.
- How to Brief a Developer on a Custom AI Workflow You Need — Brief a supplier on data limits before they build anything.
- How to Evaluate an AI Implementation Proposal or Quote — Check what a supplier says about data before you sign.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Office of the Privacy Commissioner of Canada pages on PIPEDA in brief, the fair information principles, mandatory breach reporting, its breach record inspection findings and its December 2023 principles for generative AI; section 17 of Quebec's Act respecting the protection of personal information in the private sector (LegisQuebec); OpenAI's ChatGPT business pricing page and Temporary Chat help article; Anthropic's consumer terms update and its Claude in Chrome safety article, all checked November 2026. This is general information, not legal advice.