Measure visits to this site with Google Analytics? Only with your OK. Privacy
Acteurly

AI and Quebec's Law 25: what customer data you can give AI

By the Acteurly team · · 8 min read

In short

Quebec's Law 25 doesn't ban AI: it requires you to know where your customers' information goes, who keeps it and what it's used for. Use the traffic-light test: green (nothing identifiable) in any tool; yellow (name, email) only in a business tool with a written contract, training off and an assessment done; red (health, SIN, prospecting without consent) never in an AI assistant.

Thursday, 10:10 p.m.: the copy-paste that seems harmless

Thursday, 10:10 p.m. A clinic owner wants to know which patients to follow up with. She pastes her list, appointment notes included, into ChatGPT on her personal account. The answer is excellent. The problem is the route that information took to get it.

Quebec's Law 25 doesn't ban AI. The private-sector privacy act it amended doesn't talk about tools but about information: anything about a person that allows them to be identified, directly or indirectly (s. 2). For each piece, it asks three questions: where does it go, who keeps it, and what is it used for?

Our view: those three questions protect you better than any list of "compliant" tools. A tool is never compliant on its own; your use of it is, or isn't.

The traffic-light test

Before every copy-paste, give what you're pasting a colour.

The traffic-light test: what customer information can you give AI?

LightWhat you pasteThe rule
GreenNothing that identifies anyone, even indirectly: your own copy, your prices, aggregate numbers, reviews with no name or detailsAny tool
YellowOrdinary personal information: name, email, address, purchase historyBusiness tools only, under the four conditions below
RedHealth, biometrics, intimate details, SIN, bank numbers, or a customer list for prospecting without consentNever in an AI assistant

Green is the only zone outside the law, because there's no personal information left. Watch the word "indirectly": "the woman from Cap-de-la-Madeleine who sprained her ankle in March" identifies someone without naming her.

Yellow takes four conditions, each from a section of the act:

  1. A written contract. You can hand information to a service provider without the customer's consent if it's necessary for the mandate and the written contract covers confidentiality, use limited to the mandate, and no retention after it ends (s. 18.3). With AI vendors, that's the data processing addendum (DPA), which OpenAI and Microsoft offer business customers.
  2. Training turned off. By default, OpenAI doesn't train its models on data from ChatGPT Business, Enterprise or its API (OpenAI). Google makes a similar commitment for Gemini in Workspace (Google Workspace), and Microsoft for Copilot at work (Microsoft).
  3. A privacy impact assessment (PIA). It's required before personal information is communicated outside Quebec (s. 17) and to acquire an information system that handles it (s. 3.3), in proportion to how sensitive and how much information is involved. For a small project, the Commission d'accès à l'information's guide (in French) accepts meeting notes or emails; it names AI, "generative AI in particular", among technologies that raise specific issues.
  4. The same purpose as at collection. Information may only be used for the purposes it was collected for, unless the person consents or an exception applies (s. 12). Delivering a service the customer asked for is one. Commercial prospecting, on the other hand, can never count as a compatible purpose: the act says so in plain terms.

Red also comes from section 12: information is sensitive when, by its nature, "in particular its medical, biometric or otherwise intimate nature", or because of its context, it carries a high level of reasonable expectation of privacy. Using it for another purpose requires express consent. We add the SIN and bank numbers to be safe: your marketing never needs them. A genuine need for red data calls for a project supervised by a lawyer, not a copy-paste.

Example: a Trois-Rivières clinic, three tasks

Take a fictional physiotherapy clinic in Trois-Rivières: four therapists, 1,800 active patients, 240 Google reviews. Without ever choosing it, the owner is the person in charge of protecting personal information: the act gives that role to whoever has the highest authority in the business, who can delegate it in writing (s. 3.1).

Summarizing reviews: green, after a clean-up. In a spreadsheet, she deletes the name column and replaces "shoulder," "knee," "surgery" and "concussion" with [CONDITION]: about twenty minutes. Then: "Group these reviews into 5 themes, with a count for each." Result in this example: 60 of 240 reviews mention waiting on the phone, or 25%. No name left the clinic.

Writing follow-up emails: green. Her rule: have AI write the template, not the message. "Write a follow-up email sent 10 days after a first appointment, with {first name} and {date} fields." Her booking software then does the merge. The AI never sees a patient; having it personalize each email from clinical notes would be red.

Targeting a massage therapy promotion: red, made green. Pasting the list with diagnoses to "find interested patients" combines health information and commercial prospecting (s. 12). The safe version:

  1. She keeps only patients who agreed to receive offers: 620 of 1,800, or 34%.
  2. She asks AI for targeting rules based on made-up columns ("last visit," "package type"), with no real data.
  3. She applies the rule she picks, "last visit more than 6 months ago," in her own software: 140 of 620 patients, or 23%.

Bottom line: three tasks done with the same AI, and no health information left the clinic.

What most advice gets wrong

Two camps dominate. "Never use AI with customer data": unworkable, and it pushes staff onto personal accounts behind your back. "It's fine, just turn off training": that setting deals with the vendor's problem, not yours. It doesn't create a contract (s. 18.3) or an assessment (s. 17), and doesn't make a promotion compatible with the original purpose (s. 12).

Three more traps:

  • "Hosted in Canada" doesn't settle section 17. The section covers any communication outside Quebec, and Toronto is outside Quebec. OpenAI offers storage in Canada to some customers (Enterprise, Edu, API), not all (OpenAI): a point in your favour in the assessment, not an exemption.
  • Swapping the name for [CUSTOMER] isn't anonymizing. Without the name, information is "de-identified": it no longer identifies the person directly (s. 12). Anonymizing means the person can no longer be identified, irreversibly, directly or indirectly (s. 23). De-identified text that keeps details is still yellow.
  • Your employee's personal account. On ChatGPT, chats can be used for training as long as the "Improve the model for everyone" setting is on (OpenAI). In consumer Gemini, some chats are read by human reviewers and kept for up to three years (Google). And in May 2026, Quebec's privacy regulator and three other commissioners concluded that ChatGPT's initial training and deployment didn't comply with privacy laws (Gouvernement du Québec, in French).

AI agents: when the tool does more than write

Soon, an agent will read your inbox, update your CRM and answer quote requests. Giving it access means acquiring an information system: the section 3.3 assessment happens before you connect it.

Above all, the agent decides. If it turns down an out-of-area request or drops a customer from a list on its own, based on personal information, section 12.1 likely applies. You must tell the person no later than when you inform them of the decision, tell them on request what information and main factors led to it, and let them present their side to a staff member who can review it. This week, list the decisions your tools make on their own: for each one, a person signs off, or you write the notice and name who reviews.

The honest limit: this is general information, not legal advice. The traffic-light test doesn't replace an assessment; it tells you when you need one. For health data, large volumes or minors, talk to a lawyer, and to your professional order if you belong to one.

The 8-question test: is your AI use compliant with Quebec's Law 25?

Each yes scores 1 point, each no scores 0.

  1. You know which AI tools each employee uses, and with which account.
  2. Tools that receive personal information are on a business plan with no training.
  3. You have a written agreement with each of those vendors (s. 18.3).
  4. You know where that data is hosted and have documented an assessment, even a one-page one (ss. 3.3 and 17).
  5. Before pasting, your team removes names, contact details and details that aren't needed.
  6. No health information, SIN or bank number has ever been pasted into an AI assistant.
  7. When a customer's history is used for a new purpose, such as a promotion, you have their consent (s. 12).
  8. The title and contact details of the person in charge of protecting personal information are posted on your website (s. 3.1).

How to read your score (out of 8):

  • 0 to 3: red light. Starting today, no personal information goes into AI; stick to green and follow the plan below.
  • 4 to 6: yellow light. Fix questions 3 and 4 first: without a contract and an assessment, it isn't really yellow.
  • 7 or 8: green light. Retake the test with every new tool or new hire.

Special rule: a no on question 6 puts you in the red, whatever your total.

Your plan for the next 7 days

  1. Day 1: ask each employee which AI tools they use, and with which account. No blame: you want the truth.
  2. Day 2: post the traffic-light table near the workstations.
  3. Day 3: pick one business tool and confirm in writing: training, hosting location, data agreement.
  4. Day 4: write your one-page assessment (information, tool, location, access, safeguards), the same way as for a CRM.
  5. Days 5 to 7: turn your three most frequent AI tasks into templates, post your privacy lead's contact details, then retake the test.

Frequently asked questions

Can I use free ChatGPT with customer data in Quebec?

Not with personal information. A personal account doesn't come with a written agreement limiting the vendor to your mandate, as section 18.3 of Quebec's private-sector privacy act requires, and your chats can be used for training if the setting is on. Keep it for green tasks: drafting, rewording, or analyzing aggregate numbers and text with no identifying details.

Do I need my customers' consent to use AI?

Not necessarily for the tool itself: handing information to a provider to carry out a mandate is allowed without consent if the contract is in writing and covers confidentiality. What triggers consent is a new purpose, such as using purchase history for a promotion. If the information is sensitive, like health data, that consent must be express.

Do I have to tell customers I use an AI tool?

When you collect information, the act requires you to tell customers why, and where applicable, which third parties or categories of third parties it's shared with and that it may be communicated outside Quebec (s. 8). If your AI vendor processes their data outside Quebec, your privacy policy should say so. Fully automated decisions come with their own notice requirement (s. 12.1).

What should I do if an employee pasted sensitive data into ChatGPT?

Move quickly: delete the chat and note what was pasted, when and on which account. A communication the law doesn't authorize can be a confidentiality incident (s. 3.6), and the business must then take reasonable steps to reduce the risk of harm. If that risk is serious, it must promptly notify the Commission d'accès à l'information, Quebec's privacy regulator, and the people concerned (s. 3.5): call a lawyer the same day.

In Acteurly

Read next