Personal accounts are the bigger risk
On business plans, the vendors say your data doesn't train their models. Microsoft says this for business Copilot, and OpenAI for its business plans by default. Anthropic says commercial Claude data isn't used for training unless the customer joins its optional Development Partner Program, so check whether your company has joined.
Those promises don't cover personal or free accounts. On personal ChatGPT accounts, a setting called Improve the model for everyone decides whether new conversations can be used to train OpenAI's models. Anthropic announced on August 28, 2025 that users of its Free, Pro and Max plans choose whether their chats can be used to improve Claude. In both cases the employee makes that choice, and your company can't see or change it.
Give staff an approved business account and say in writing which tools they may use and what stays out. AI policy for Canadian businesses covers what to put on that page.
Where to start
Work through these in order. The first one isn't about AI at all.
- Secure your accounts with your IT provider. Turn on two-step sign-in for email and every business system, use strong passwords, and train staff to spot phishing emails. The security problems we've seen at clients came from weak passwords and phishing emails. An AI connected to a stolen account can reach whatever that account can.
- Know what's sensitive. Decide which client, staff and financial information should never go into an AI tool, and where it lives. Ask IT to tighten folders shared with the whole company that hold that information. AI and sensitive information walks you through it.
- Then choose the tool. Copilot vs ChatGPT vs Claude compares training, access, data location and price.
Reading and acting are different kinds of access
Most of the worry about connecting AI to email and files comes from not knowing what a connection allows. There are two kinds of access.
- // Reading
The AI looks at information
For example, your inbox or a folder of files, so it can prepare a brief or a draft. That information goes to the AI provider to be processed, so the plan and settings matter.
- // Acting
The AI changes something
For example, it saves a draft, updates a record or sends a message. Acting needs tighter rules, and at first a person should approve each result.
Start with reading. Add acting only for specific tasks, once you've seen the results and set the rules. A promise to "only create drafts" doesn't limit what a connection can do. The permission does.
Also check whose permissions a connection uses. Some Claude connectors use each employee's own permissions. Others use one shared login, which can let everyone reach every file that login can see.
What your IT provider will want to see
If your company has an IT provider or an internal IT lead, expect questions. That's a good sign. Before anything is connected, give them:
- which accounts and tools would be connected, and which stay out;
- the exact permissions requested for each connection, in the provider's own names, and what each one allows;
- whether each connection only reads or can also act;
- what information is in bounds and what is excluded;
- who on your side reviews results and approves access; and
- a read-only first step, if they prefer to start there.
They'll decide whether the connections are acceptable, where the file-access boundary sits and who the admin contact is. In our experience, that takes one or two short conversations.
What nobody can promise
No vendor or adviser can promise that an AI tool will never make a mistake, or that your information is safe in every case. What happens to it depends on the service, plan and settings you choose, so check those against how sensitive your work is. This page explains the questions to ask. It isn't a security certificate.
Keep a person checking anything with consequences, such as a message to a client or a change to a record, until you decide otherwise. If your work involves regulated or highly private information, some of it should stay manual.
If you work with Sidekick
- The work runs in your company's Claude, ChatGPT or Copilot business account, connected to Microsoft 365 or Google Workspace.
- You or your IT provider grant access through your company's admin accounts, and can remove it the same way.
- You decide what's in bounds. We write it down before launch, along with who approves access.
- We don't use your information for model training, and we don't sell it.
- We don't keep access after delivery unless you share something for support.
Book a free 30-minute call. You don't need to prepare anything, and you're welcome to bring your IT provider. Many of the questions above are easier to answer with them in the room.
Sources checked October 12 and 16, 2026: Microsoft, Data, privacy and security for Microsoft 365 Copilot (updated August 18, 2026); Anthropic, commercial data and training (March 16, 2026); Anthropic, updates to consumer terms (August 28, 2025); Anthropic, custom connectors; OpenAI, data controls FAQ. Also: OpenAI, Enterprise privacy. Product terms change, so check the current page for your plan.
Free checklist: AI Safety and Confidentiality Check.
