Shadow AI on Your Team: The One-Page Staff Policy

Somebody on your team is already pasting client details into a chatbot. Not out of malice. They're trying to get a quote out before five o'clock and the tool is sitting right there in a browser tab. That's shadow AI: real work getting done on tools nobody approved, with data nobody tracked.

Statistics Canada reported that 19.2% of businesses used AI to produce goods or deliver services in the 12 months before its second quarter 2026 survey, up from 12.2% a year earlier and 6.1% in 2024 (Statistics Canada, June 2026). That's the sanctioned use people are willing to report. The quiet use sits on top of it.

So here's a fall project worth an afternoon: one page, ten minutes to read, that says which tools are fine, what never gets typed in, and who to ask. Not a thirty page governance framework. One page.

The default settings are most of the problem

The gap between a personal account and a work account is wider than most owners realize, and it is the single fact that makes a policy worth writing.

OpenAI's policy page, updated March 13, 2026, says it plainly: "When you use our services for individuals such as ChatGPT, Sora, or Operator, we may use your content to train our models." For business plans the default flips: "By default, we do not train on any inputs or outputs from our products for business users, including ChatGPT Team, ChatGPT Enterprise, and the API" (OpenAI).

Anthropic draws the same line. Its privacy centre article, last updated August 18, 2026, states that "By default, we will not use your inputs or outputs from our commercial products ... to train our models," with an exception when a user explicitly submits feedback or bug reports (Anthropic Privacy Center). On consumer Claude accounts the model improvement setting lives in the individual's own privacy settings, and switching it off only affects what comes next: "Your data will still be included in model training that has already started and in models that have already been trained" (Anthropic Privacy Center).

Microsoft's Copilot documentation, dated July 9, 2026, is just as direct: "Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs, including those used by Microsoft Copilot" (Microsoft Learn).

Same sentence, two different accounts, two very different outcomes. Most of what a staff policy does is make sure people are typing into the right window.

What Canadian regulators already expect from you

None of this is a legal grey zone waiting to be settled. Canada's federal, provincial and territorial privacy authorities published joint Principles for responsible, trustworthy and privacy-protective generative AI technologies, and they are blunt that existing law applies: "Organizations developing, providing, or using generative AI are obligated to ensure that their activities comply with applicable privacy laws and regulations in Canada."

Two lines from that document land hardest on a small business. Under Accountability, organizations should "Have a clearly defined internal governance structure for privacy compliance, including defined roles and responsibilities, policies and practices establishing clear expectations with respect to compliance with privacy obligations." And, in case anyone hopes to point at the software when something goes wrong: "Know that accountability for decisions rests with the organization, and not with any kind of automated system used to support the decision-making process" (Canadian privacy commissioners).

British Columbia has teeth here too. On May 6, 2026, BC's Information and Privacy Commissioner and counterparts in Ottawa, Quebec and Alberta released the results of a joint investigation into OpenAI, finding that the initial training and deployment of ChatGPT fell short on overcollection, consent and transparency, accuracy, access and correction rights, and "a general lack of accountability." Commissioner Michael Harvey called it "the fundamental disconnect between how one of the world's most widely used AI applications works and what BC's Personal Information Protection Act demands" (OIPC BC). Private businesses in Burnaby and across BC are subject to that same Act.

The one page policy

Print this, adapt the names, and hand it out. It is deliberately short, because a policy nobody finishes reading is not a policy.

• Approved tools. List the two or three accounts the business pays for, by name, and say that work goes through those and not through personal logins. Name the person who approves anything new.
• Never goes in a prompt. Client names and contact details, health information, banking and payroll data, signed contracts, passwords and API keys, anything covered by an NDA. Describe the problem instead of pasting the file.
• Check the account. Before typing, confirm you are in the company workspace, not a personal free account. If you are not sure, ask.
• A human signs off. Anything going to a customer, a regulator, or the CRA gets read by a person first. AI drafts, people approve.
• Say so when it matters. If a customer is talking to a bot, or an AI tool helped decide something about them, tell them.
• Log it. One shared sheet: tool, what it is used for, who owns it. Five minutes a month to update.
• Report mistakes without penalty. If something sensitive got pasted somewhere it shouldn't, tell the owner the same day. No discipline for reporting, only for hiding it.
• Review quarterly. Defaults, plans and features change. So should the page.

Rolling it out without a fight

The instinct is to lead with the ban. That is what creates shadow AI in the first place, because the work still has to get done.

Lead with the paid account instead. Buy the business plan, walk the team through what it does well, then hand over the page. Statistics Canada found that among businesses using AI, 44.4% made changes to training or staffing practices, but the split by size is stark: 68.1% of businesses with 100 or more employees trained existing staff, compared with just 24.0% of businesses with one to four employees (Statistics Canada, June 2026). Small shops are writing rules without teaching anyone how to follow them. An hour of hands on practice beats the memo.

If you want a structure to grow into later, the NIST AI Risk Management Framework, released as version 1.0 on January 26, 2023, organizes the work into four functions: Govern, Map, Measure and Manage (NIST). Your one pager is the Govern function for a ten person company. That is a fine place to start.

Where this doesn't apply

A policy is not a control, and it is worth being honest about the limits.

It does not stop technical leakage. OWASP's 2025 entry on sensitive information disclosure warns that instructions telling a model not to reveal sensitive data "may not always be honored and could be bypassed via prompt injection," and recommends sanitizing data and educating users rather than relying on the model to behave (OWASP Gen AI Security Project).

It also will not make a bad fit good. In a randomized controlled trial of 16 experienced open source developers working on 246 tasks in repositories they knew well, researchers found that allowing AI tools increased completion time by 19%, even though the same developers had forecast a 24% speedup and still believed afterward that they had been faster (Becker et al., 2025). On deep, familiar, high context work, the tool can slow you down and you will not notice.

And plenty of businesses genuinely do not need this yet. In the same Statistics Canada survey, 40.0% of businesses said AI simply is not relevant to what they produce, while 13.4% cited cybersecurity or privacy concerns as a barrier. If you have four staff, no client personal information on file, and nobody using these tools, a policy is paperwork. Write it when the first person starts using something.

Sources

• Statistics Canada. "Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026." June 11, 2026. Link
• Office of the Privacy Commissioner of Canada and provincial counterparts. "Principles for responsible, trustworthy and privacy-protective generative AI technologies." 2023. Link
• Office of the Information and Privacy Commissioner for British Columbia. "Canadian privacy regulators' joint investigation into ChatGPT leads to improvements in users' privacy." May 6, 2026. Link
• OpenAI. "How your data is used to improve model performance." Updated March 13, 2026. Link
• Anthropic. "Is my data used for model training?" Privacy Center, updated August 18, 2026. Link
• Microsoft. "Data, Privacy, and Security for Microsoft Copilot." Microsoft Learn, July 9, 2026. Link
• OWASP Gen AI Security Project. "LLM02:2025 Sensitive Information Disclosure." 2025. Link
• Becker, J., Rush, N., Barnes, E., Rein, D. "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity." arXiv:2507.09089, 2025. Link
• National Institute of Standards and Technology. "AI Risk Management Framework 1.0." January 26, 2023. Link

Autana Solutions builds AI employees and automations for small and mid-sized businesses around Burnaby, New Westminster and Metro Vancouver, and we usually start by finding out what your team is already doing without telling you. If you want help writing your one pager and getting the right accounts in place before year end, book a free call and we'll walk through it with you.