Business data and AI tools: what to check before you connect anything

What "the tool uses your data" can mean

First distinction: personal account or business account. A personal account runs on consumer terms, and the company has no admin control over it. Business offerings usually come with separate commercial terms and an admin console.

Next, "the tool keeps or uses your data" covers several different things. Check each one separately:

• Retention: how long your prompts and the responses are stored, and who can delete them.
• Human review: whether the provider's staff can read content, for example to detect abuse.
• Model training: whether your content is used to improve models, by default or only if you agree.
• Sub-processors: which other companies take part in processing, such as a hosting or model provider.
• Location: which countries the data is processed and stored in.

Finally, connectors change the scale. A tool linked to a mailbox, a CRM or a shared drive can generally see everything the connected account can see. If that account can reach everything, so can the tool, including folders nobody thought to lock down.

The Canadian frame, in plain words

According to the Office of the Privacy Commissioner of Canada, the federal law, PIPEDA, generally applies to personal information held by private-sector organizations that are not federally regulated and do business in New Brunswick, Nova Scotia, Prince Edward Island and Newfoundland and Labrador. The same page notes provincial health information laws in some of these provinces. If you handle health information, get confirmation of which law applies to you.

Second, handing data to a provider does not hand over responsibility. The OPC's guidelines on processing across borders say an organization remains responsible for personal information transferred to a third party for processing, including outside Canada. It must use contractual or other means to provide a comparable level of protection, and tell people clearly that their information may be processed in another country and be accessible to authorities there.

In 2023, federal, provincial and territorial privacy commissioners published principles for generative AI. For organizations using these tools, they say to enter personal information into a prompt, especially sensitive or confidential information, only where authorized. They also point to assessments such as privacy impact assessments.

Sort your data first

Before comparing tools, decide what is allowed in. The OPC notes that health and financial information are generally considered sensitive, that any information can become sensitive depending on context, and that safeguards should match the level of sensitivity. Four categories are enough to start.

Check a specific tool

For each tool, look for five documents or screens. Most large providers have them, under various names.

• The commercial terms and data processing agreement.
• The data usage page: training, retention, human review, sub-processors.
• The admin console: who has access, and which features and connectors are allowed.
• Region or data residency options, and who is eligible for them.
• Audit logs: what gets recorded, for how long and who can view it.

Here is what a few official pages said at the time of checking, on September 15, 2026. These are summaries of public sources, not recommendations, and they don't cover every offering from each provider.

• Microsoft: its documentation for Microsoft Copilot (formerly Microsoft 365 Copilot) states that prompts, responses and data accessed through Microsoft Graph aren't used to train foundation models. It says Copilot only surfaces data a user has at least view permission to, and that queries from customers outside the European Union may be processed in the US, the EU or other regions.
• Google: its privacy hub for generative AI in Google Workspace states that Workspace does not use customer data to train models without the customer's prior permission or instruction, and that Gemini only retrieves content the user has access to.
• Anthropic: its privacy centre states that, by default, inputs and outputs from its commercial products are not used to train its models, unless you submit feedback or otherwise choose to allow it.
• OpenAI: its API platform documentation states that data sent to the API is not used to train its models unless you explicitly opt in, that abuse monitoring logs are kept for up to 30 days by default, and it lists Canada among the data residency regions available to eligible customers.

Notice what Microsoft and Google have in common: the assistant follows your existing permissions. That's reassuring and demanding at the same time. If a payroll folder is shared with the whole company, the assistant can find it for anyone.

Connecting the tool to your systems

Linking an AI tool to email or a CRM means handing it keys. The good practices are the same as for any new access, just applied more carefully.

• Least privilege: one folder, mailbox or list, not the whole account.
• Dedicated service account: never the owner's or an admin's.
• Read-only first: during the trial, the tool reads and suggests, without writing or sending.
• Separate test environment: fictional data or a limited copy before the real thing.
• Logs: record what the tool reads and does, and actually look at them in the first weeks.
• Revoking access: know where to cut access and who can do it, and test it once.
• Human approval: any external action, such as sending an email or changing a client record, waits for a person's go-ahead.

An example from start to finish

Your internal rule doesn't need to be long. One clear page, walked through at a team meeting, beats a twenty-page policy nobody reads. Adapt this template to your business.

AI TOOL USE RULE
[Company name] · Effective [date] · Contact: [name, email]

1. Approved tools and accounts
- Only [approved tool], with your company account.
- No personal accounts for work involving clients or internal data.

2. Allowed
- Drafting or rewording text that contains no personal information.
- Summarizing public or non-sensitive internal documents.
- Preparing a draft client reply after removing names, addresses and numbers where possible.

3. Not allowed
- Entering health information, financial data, employee files, ID documents, passwords or confidential contracts.
- Sending AI-generated text without reading it first.
- Connecting an AI tool to email, the CRM or shared drives without written approval from [contact].

4. Always
- Check facts, amounts and names before sending.
- When in doubt, don't paste it. Ask [contact].

5. If something goes wrong
- Entered sensitive data by mistake? Tell [contact] the same day. Nobody gets blamed for reporting it.

Next review: [date, at least once a year]