Our article on shadow AI concluded that a small business needs two things to rein in how its staff use ChatGPT: an approved tool, and a one-page policy. We never wrote the page. Here it is. It also underpins any AI agent project, because an agent inherits whatever rules the company has already set for its data.
The templates you find online tend to be long governance frameworks, policies written by public bodies, or generic forms with no tie to Quebec law. This one fits on a page, and every clause is followed by what justifies it: usually a section of Quebec's Act respecting the protection of personal information in the private sector (Law 25), sometimes a term the vendor publishes, or a plain practical reason. A clause you can't explain to an employee won't be followed.
The clauses are written to be copied. Replace whatever is in brackets with your tools and your names.
The short answer, for the busy
An AI use policy for a Quebec small business fits in eight clauses. It names the approved tools and requires the company account, because the business plans of ChatGPT, Claude and Copilot exclude training on your data by default, while personal accounts don't always. It forbids pasting personal information about clients or employees into a tool that hasn't been assessed, since section 17 of Law 25 requires a privacy impact assessment before any personal information is communicated outside Quebec. It forbids letting AI alone make a decision about a person without telling that person, as section 12.1 requires. It makes a human review anything AI produced before it reaches a client, sets rules for meeting note takers, names the person to ask, and ties into the governance policy section 3.2 already requires. Each clause comes with the reason behind it, because a rule nobody can explain gets ignored.
Why one page, and why it fails on its own
Your staff already use AI. A page that only says "don't use ChatGPT with client data" will be ignored by the first busy week.
Its real job is to point to the permitted route: this tool, this account, this data, and who to ask when something fits nowhere. A page gets reread when someone hesitates; a thirty-page framework gets signed at onboarding and never opened again. So buy the tool before you publish the policy. Otherwise what you publish is a ban.
Clauses 1 and 2: the tool and the account
Clause 1. Approved tools. The AI tools approved for work are [ChatGPT Business, Claude Team or Microsoft Copilot Chat, signed in with the company account]. Any other AI tool, including a browser extension or a mobile app, must be approved by [name or role] before it is used with company documents.
Clause 2. Accounts. These tools are used only with the account the company provides. A personal account, free or paid, is never used for work that involves information about our clients, employees or partners. Company administrators can view conversations held in company accounts.
Clause 2 does most of the work, because the gap between the two kinds of account is contractual. On a personal ChatGPT or Claude account, training on your conversations stays allowed until the user opts out; on ChatGPT Business and Claude Team it is excluded by default (our comparison of AI subscriptions cites the vendors' pages). OpenAI's help centre adds a detail: on Free, Plus and Pro, even after opting out, a thumbs up or down on a response means the whole conversation may be used to train its models. On ChatGPT Business, that thumb permits no training at all. So the careful employee who switched the setting off and then rated an answer can still send the client file they had just pasted into training.
At Microsoft, enterprise data protection applies to Copilot Chat at no extra cost, and prompts and responses aren't used to train foundation models. But the Copilot app also accepts a personal Microsoft account, and the protections follow whichever account is signed in. Two cues tell them apart: a "Work" label under the profile and a green shield at the top of the window. Show them to your staff; it's the simplest check in the whole policy. Administrators can also block personal accounts.
The last sentence of clause 2 matters too: on ChatGPT Business, workspace admins can view, export and delete members' conversations. Staff who know that choose more carefully what they type.
Clause 3: what never gets pasted in
Clause 3. Personal information. We don't paste personal information about our clients or employees into an AI tool (a name with contact details, a social insurance number, health information, financial records, a performance review), except in a tool [person responsible] has assessed for that use. When in doubt, remove whatever identifies the person, or ask first.
Section 3.1 makes the business responsible for the personal information it holds, whatever tool an employee picked. Section 17 requires a privacy impact assessment before personal information is communicated "outside Québec", followed by a written agreement. The same applies when a business entrusts someone outside Quebec with using or keeping that information on its behalf, which is a fair description of an AI subscription. We walk through the mechanics in our article on data leaving Quebec, and our privacy impact assessment generator produces the document.
Apart from the name with contact details, the examples in parentheses are sensitive information under section 12: medical, biometric "or otherwise intimate" in nature, or sensitive because of their context. The AI vendor is also a mandatary, and its contract has to contain what our article on vendor contracts describes.
Clause 4: decisions about people
Clause 4. Decisions. No decision about a person (hiring, credit, pricing, refusal of service, discipline) is made by an AI tool alone. A member of the team reviews the file and owns the decision. If a process has to decide automatically, [person responsible] approves it first and the person concerned is told.
Section 12.1 covers a decision "based exclusively on an automated processing" of personal information. The business must tell the person no later than when it informs them of the decision, then, on request, explain what information was used and why, and give them the chance to submit observations to a staff member who can review it. In our reading, a human who approves without looking doesn't take the decision outside the section. Section 11 also requires information to be accurate when it's used to decide, and an AI-written summary is a new source of error (our article on section 12.1 goes further).
Clause 5: a human reads it before it leaves
Clause 5. Review. Any text, figure or document produced with AI is reviewed by the person sending it before it goes to a client, a supplier or a public body. That person answers for it as if they had written it alone.
No statute requires this clause. It exists because language models produce plausible mistakes: an invented figure, a reference that doesn't exist, another client's name left in a template. Your client won't tell your error from the tool's, and they'll be right. Reviewing means checking every name and every number. The clause matters even more for ChatGPT Work and Claude Cowork, which can act in a browser or in your files: approval comes before the action.
Clause 6: meeting note takers
Clause 6. Meetings. We don't record a meeting with an AI tool without telling the participants. Only [approved tool] is allowed, with automatic joining turned off. We don't record disciplinary or health-related meetings, or meetings covered by professional secrecy.
This is the short version of the rule in our article on AI note takers: the host business answers for what the tool collects about participants who never signed anything.
Clauses 7 and 8: who answers, and where the policy attaches
Clause 7. Questions and mistakes. Questions about this policy go to [name, role, email]. If you pasted personal information into an unapproved tool by mistake, tell them the same day. Reporting a mistake is never punished.
Clause 8. Framework. This policy is part of our governance policies on personal information. It is approved by [person in charge of the protection of personal information] and reviewed every year, or whenever a tool changes.
Under section 3.1, the person with the highest authority in the business is the person in charge of the protection of personal information unless that role is delegated in writing. In a small business, the name in clause 7 is often the owner's. Not punishing reports is a practical choice: a hidden mistake can't be fixed.
Section 3.2 requires governance policies that, among other things, "define the roles and responsibilities of the members of its personnel", approved by that same person. An AI policy is one piece of that. Attach it to your governance policy rather than leaving it orphaned, and cover it in your staff training.
What this page doesn't cover
The policy governs what your staff do with the tool. It doesn't replace assessing the tool itself. Section 3.3 requires a privacy impact assessment for any project to acquire an information system involving personal information, proportionate to, among other things, its sensitivity and quantity. In our reading, buying an AI subscription is one. For a team drafting emails the exercise stays short, but it has to exist.
Nor does it fix the copy-paste repeated every morning between the inbox and the invoicing software. When a task repeats like that, the better answer is often an agent connected to your systems that only receives the data it needs. For a look at the whole picture (tools, contracts, assessments), see our website security and compliance service.
Where to start
Choose the tool and buy the accounts. Fill in the brackets, have the page approved by your person in charge of privacy, then take twenty minutes at a team meeting to show two things: how to check you're in the company account, and what clause 3 says.
If a specific task looks like it deserves more than a subscription, let's talk. We'll look at what your team already does with AI and tell you plainly whether an agent is worth it.
→ A first conversation, no commitment
Vendor terms were read on the vendors' official pages on October 28, 2026; they change often. This article explains Law 25 obligations in plain language to help a business ask the right questions. It is not legal advice. The policy template is a starting point to adapt to your business. The statutes published on LégisQuébec are authoritative.
Written byXavier Peich