Writing an AI policy your team can use
List the AI tasks staff can do, what they can enter, who checks the output and how to report a mistake. Give clear examples and keep it current.

An AI policy should help someone decide whether they can use a tool for the task in front of them, what information they can enter and who needs to check the result. A long list of principles won’t answer those questions.
Start with the AI tools already used at work. Build a short policy around real tasks, such as drafting a public event description or summarizing an internal document with no client details in it. Give the policy an owner who can answer questions and sign off on changes.
1. Name the uses and accounts you allow
List the tools, account types and tasks you’ve reviewed. Clearing the business account doesn’t cover a personal account with the same product name, or a new connector someone adds inside it.
Give staff examples of work they can do, work they need to ask about first and work they shouldn’t do with AI. For instance, drafting from public information may be permitted, while processing client records needs a specific review. Decisions affecting employment, credit, health or legal rights need review by someone with the right specialist knowledge before anyone uses AI for them.
Explain how to request a new tool or use case. Assign a business sponsor and someone responsible for checking security, privacy and contractual terms. Until that review is done, staff should keep doing the task the way they do now.
2. State what information can be entered
Connect the policy to your information-handling rules. Address client records, employee information, credentials, confidential contracts and material belonging to someone else. Include uploads, screenshots, meeting recordings and connected data sources as well as typed prompts.
The Canadian privacy commissioners’ generative-AI principles cover accountability, appropriate purposes, transparency and limiting personal information. Paying for the service or turning off training on your data still leaves you responsible for all of them.
Give staff a simple default: use public or made-up examples unless the policy owner has cleared that information for that service. The public-AI data guide covers the supporting data-handling checks. Tell them whom to ask when they aren’t sure.
3. Assign responsibility for the output
Name a reviewer for work that will be published, sent to a client or used in a significant decision. The reviewer needs the knowledge and time to check it, and authority to reject it.
For a factual draft, check names, figures, sources and dates against independent records. For code or a workflow, use the normal review and testing process. For decisions about people, look for unfair assumptions and possible discriminatory effects. Smooth wording can still be wrong.
NIST’s AI risk-management core calls for defined responsibilities and oversight. It’s a useful planning aid, but following it doesn’t mean you’ve met your legal obligations.
Don’t assume a generated image, text or code is cleared for commercial use. Check service terms, source material, client commitments and relevant rights before publication. Get legal advice when ownership or permitted use matters to the deliverable.
4. Keep useful records and provide a reporting route
Record the use cases you’ve allowed, who owns each one, important review decisions and the tool or configuration used where that helps explain the result. Decide what records each workflow needs.
Avoid requiring a permanent copy of every prompt. Prompts and outputs can contain personal or confidential information themselves. Limit access to retained records and set a retention period that matches the purpose and applicable obligations.
Give staff one reporting route for accidental uploads, unexpected actions or incorrect output that was already used. Ask them to report promptly, avoid further sharing and preserve the necessary details securely. Explain who can pause the workflow and assess the incident.
5. Practise the policy and revise it
Walk through a few examples with the team. Ask them to identify which account to use, what they can enter, reviewer and reporting contact. Fix instructions that require guesswork.
Set a review date, and revisit the policy when service terms, connectors, business tasks or incident findings change. Tools that can act on other systems also need enforced permissions and stop procedures; the AI-agent guide covers those controls.
Talk to us if you need help turning your team’s AI use cases into clear rules and workable technical controls.


