An AI box is still an outside service
Typing information into an online tool may send it to a service provider. A familiar logo or a useful answer is not enough to show that the company has approved that use. Start by agreeing which account and settings employees may use for each type of work.
Set a simple information boundary
List what employees may use for practice and what they must keep out. Examples to check include customer details, staff information, prices, contracts and account credentials. Use fictional or approved non-sensitive examples until the company has confirmed the right controls.
Check the actual product and account
Ask the person responsible for company systems to confirm how the selected product handles entered information, access, retention and any use for model training. Settings and terms can differ between products and account types. Do not assume a personal account has the same controls as a company-managed service.
Make the exercise useful without real records
An employee can practise drafting a customer update with invented names and dates. They can learn the structure and checking process before using approved company information. Removing a name alone may not remove every detail that identifies a person or business.
Check what comes back
Read the output against the approved source. Check facts, dates and commitments. Decide who must approve the result before it reaches a customer or changes a company record. If information was entered somewhere it should not have been, follow the company's incident process promptly.
A practical starting point, not a guarantee
The Canadian Centre for Cyber Security describes how employee prompts can expose organizational information to third parties. Read its guidance on organizational risks. Your company's privacy, security and contractual requirements still govern the tools you use; training alone does not establish compliance.