AI acceptable use policy: what a Canadian organization should include

An AI acceptable use policy tells staff which AI tools they may use for work, what information they may put into them, what they must check before relying on the result and whom to ask when a case is not covered. It applies to standalone chat assistants and to the AI features inside software the organization already uses: email, documents, meeting platforms, support tools and code editors.

The Canadian Centre for Cyber Security's top 10 AI security actions (May 2026) include "Create an internal business policy for the acceptable use of AI tools." As of October 2026, Canada has no AI statute in force for private businesses, and none of the federal, Ontario, Quebec, Alberta or British Columbia laws we reviewed requires a document by that name. Privacy, employment, human rights, contract and professional obligations still apply to how staff use AI, and a written policy turns them into instructions people can follow. The companion article on AI laws and regulation in Canada explains which of those obligations to check.

Start from the work people already do

Take a short inventory before you write anything. Ask each team which AI tools they use, on which accounts and for which tasks. Include:

• Personal accounts used for work.
• AI features switched on inside software the organization already approved.
• Browser extensions, meeting recorders and transcription tools.
• Coding assistants and any tool connected to email, files or business systems.

The inventory shows which uses to approve, which to move to a managed account and which to stop. Without it, a policy can ban a tool that people rely on every day, and they may keep using it without telling anyone.

What should be in an AI acceptable use policy

The twelve sections below fit most small and mid-size organizations. Keep the part staff read short enough to check in the middle of a task, and put detailed tool decisions in a separate register. Adapt the quoted wording and replace the parts in brackets.

1. Scope

State who the policy covers (employees, contractors, volunteers) and which tools. Define AI broadly enough to include features built into other software and agents that act on a person's behalf.

This policy applies to everyone who works for [Organization], on any device, when they use a tool that generates text, images, audio, code or recommendations with artificial intelligence, including AI features in software we already use.

2. Approved tools and accounts

List each approved tool with the required account type, the tasks it is approved for and the information it may receive. Require accounts managed by the organization, so access ends when employment ends. Approve tools per task: a tool cleared for drafting public text needs a separate decision before it records meetings or searches client files. New AI features in approved software need a decision too. The City of Edmonton's generative AI standard, for example, sends any approved software that gains generative AI features back through the City's technology intake process.

Use only the tools listed in the AI tool register, signed in with your work account. Do not use a personal account for work information. Ask [contact] before you use a new tool or turn on a new AI feature.

3. Information rules

Staff will look at this section in the middle of their work, so make each level concrete. Start from the classification the organization already uses:

• Never: passwords, API keys, access tokens and other credentials, and information a contract forbids you to share with a third party.
• Only with approval for that tool and task: personal information about clients, employees or applicants, health and financial records, client confidential information, unreleased source code, and legal opinions or other privileged communications.
• Approved tools only: internal documents that are not public.
• Any approved tool: information already published.

Removing a name does not make a record anonymous, because context can identify a person. The Canadian privacy commissioners' principles for generative AI (December 2023) ask organizations that use these tools to put de-identified information in prompts where possible and reasonable, to enter personal information only where authorized, and not to keep, reuse or disclose prompts for other purposes unless required. They also treat what a tool infers about an identifiable person as personal information.

Never enter passwords, keys or other credentials into an AI tool. Do not enter personal information or client confidential information unless the tool register approves it for that tool and that task.

4. Checking the output

The person who uses an output is responsible for it. Require a check that fits the task: compare a summary with the original, check calculations against their inputs, open each cited source and confirm that it supports the claim, and send code through the normal review process. In Zhang v. Chen, 2024 BCSC 285, the Supreme Court of British Columbia ordered a lawyer to pay costs personally after a court application cited two cases that ChatGPT had invented.

You are responsible for anything you send, publish or decide using AI output. Check facts against the source, and do not cite a document you have not opened.

5. Decisions about people

Name the decisions where AI may assist but a person must decide: hiring, pay, discipline, performance ratings and eligibility for a service. Two provincial rules apply to this section directly:

• In Ontario, an employer that uses AI to screen, assess or select applicants must say so in each publicly advertised job posting posted on or after January 1, 2026, including when a recruiter does the screening on its behalf. The rule applies when the employer has 25 or more employees in Ontario on the day the posting is published (Employment Standards Act guide).
• In Quebec, a business that makes a decision about a person based exclusively on automated processing of personal information must tell the person and give them a chance to submit observations to a staff member who can review the decision. On request, it must also give the personal information used and the reasons, principal factors and parameters behind the decision (section 12.1 of Quebec's private-sector privacy act).

Do not use AI to make a final decision about a person's employment, pay, discipline or eligibility. Tell [contact] before you use an AI tool to screen or assess job applicants.

6. Telling people when AI is involved

Decide when clients, customers or the public must be told that AI produced or helped produce content, and require a chatbot to identify itself as one. An organization answers for what its chatbot says. In Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia's Civil Resolution Tribunal rejected the argument that a website chatbot was responsible for its own statements, and held the airline liable for wrong fare information the chatbot gave a customer.

Label AI-generated content when [cases]. A customer-facing assistant must state that it is automated and offer a way to reach a person.

7. Meetings and recordings

AI note-takers record and transcribe other people. Require participants to be told before recording starts, follow the client's rules in client meetings, and store and delete transcripts like other records.

8. Connectors, agents and actions

A tool that can read a mailbox, search a shared drive or call another system needs its own approval, separate from the chat tool it belongs to. List the approved connectors. For agents that can send messages, change records or spend money, state which actions need a person's approval each time, and who can give it. Then set the same limits in the agent's permissions, because the written policy by itself does not restrict what an agent can do.

9. AI-enabled fraud

Voice cloning and video deepfakes make impersonation cheap. The Cyber Centre recommends that staff "verify unusual requests across channels" and that organizations use out-of-band verification for sensitive actions. Put the rule in the policy and name the channel.

If you receive an unusual request for a payment, a credential or sensitive information, confirm it through a separate channel you already know, even if the voice or video seems genuine.

10. Records

Prompts, uploaded files and outputs can be business records. Edmonton's standard, for example, treats all generative AI inputs and outputs as City records. They can fall under retention schedules and litigation holds and, when they contain personal information, access requests. When AI is part of a decision about a person, the privacy commissioners' principles ask organizations to keep enough records to answer an access request about that decision. State where AI work products are kept and for how long, and check what the provider keeps on its side.

11. Questions, mistakes and exceptions

Give one contact or channel for questions, new-tool requests and incidents. Tell staff to stop and report if they entered restricted information, without copying that information into another message. Record each exception with its purpose, approver, safeguards and expiry date.

If you entered information this policy does not allow, stop using the tool for that task and tell [contact] the same day. Report the tool, the account, the time and the type of information, not the information itself.

12. Ownership, training and review

Name the person who owns the policy and the tool register. Brief staff before they get access to a tool. Review the policy on a fixed schedule and also when a provider changes its terms, a tool adds connectors or agents, or the organization starts to handle a new kind of information.

What official Canadian guidance recommends

Four Canadian sources describe what a policy like this should cover:

• The Cyber Centre's top 10 AI security actions (ITSAP.10.049) list an acceptable use policy, allow and deny lists for AI tools, less personal information in prompts, control of unsanctioned "shadow AI" and vendor terms that prohibit unauthorized training on organizational data.
• The privacy commissioners' generative AI principles cover legal authority and consent, appropriate purposes, limits on collection, accuracy, safeguards, openness and individual access.
• The Treasury Board's Guide on the use of generative AI applies to federal institutions only. Its FASTER principles (fair, accountable, secure, transparent, educated, relevant) and its lists of what staff should and should not do still make a useful model. The City of Windsor's AI policy, adopted in April 2026, uses the same six principles.
• CAN/DGSI 101:2025, a Canadian standard on the ethical design and use of AI by small and medium organizations (generally fewer than 500 employees), covers third-party tools used inside the organization as well as tools it builds.

ISO/IEC 42001 and the NIST AI Risk Management Framework describe complete management systems. An acceptable use policy is one part of each, and a small organization can write a useful policy without adopting either.

Make the rules hold in the tools

Most rules need a setting behind them. "Use your work account" holds only if personal sign-ins are blocked, and a rule about client files holds only if the tool cannot reach those files. For each rule, write down what enforces it:

• Managed accounts with single sign-on for each approved tool.
• Administrator settings for model training, data retention and sharing, checked against the provider's current terms.
• Allow and deny lists in browser, network or device management.
• Data loss prevention rules for the information levels in section 3.
• Connector and agent permissions limited to the task, with approval steps for actions.
• Usage logs that staff are told about. An employer with 25 or more employees in Ontario on January 1 must describe how it electronically monitors employees in a written policy under the Employment Standards Act.

The same rules apply when the organization connects AI to its own systems, such as an assistant that answers from the help centre or an agent that drafts replies from client records. The policy then has to be built into the system: which data it can reach, which actions need approval and what it logs. ai.lua.haus builds those systems inside the client's environment, as an AI integration for one workflow or an AI platform setup with model access, permissions, logging and cost controls under the client's own accounts.

This is general information, not legal advice. Confirm the obligations that apply to your organization with a qualified professional.