An employee wants to finish a useful task: match client account numbers to the corresponding names. An AI assistant could speed up the work, and uploading the spreadsheet feels like the fastest way to give it context.
Consider an illustrative scenario. The two columns needed for the task sit in a workbook that also holds payment details, financial information and internal notes on other sheets. The employee uploads the original file and asks the assistant to use only the account-number and client-name columns.
The instruction narrows the requested answer. It does not narrow the file that was supplied.
The exposure starts with that mismatch: a small task, a much larger input.
Ordinary work, unnecessary disclosure
This is not a story about a malicious employee, and it is not a reason to ban AI. The employee is doing ordinary work efficiently. The input simply carries more than the job requires.
The same pattern can appear when a sales team uploads a full customer export to draft one follow-up, or a manager pastes an entire personnel document to improve one paragraph. These are workflow choices worth examining, not proof that every use causes a breach.
For personal data, the UK ICO's data-minimisation guidance says it should be adequate, relevant and limited to what is necessary for its purpose. That principle is about personal data, but the practical question travels well to other confidential business information: what does this task actually need?
The approved service is only part of the answer
Choosing a service and deciding what to put into it are separate decisions.
An approved enterprise account can bring real contractual and administrative protections. OpenAI's enterprise-privacy commitment says business data from ChatGPT Business, Enterprise and Edu is not used for training by default unless the customer explicitly opts in. That matters. It is not the same as no processing or retention, and it does not settle whether the payment and financial sheets were necessary or authorized for this task.
So check the actual service, the account, the permitted data types and the retention terms. A familiar provider name does not make every account equivalent. Equally, not every upload becomes public or trains a model.
Shrink the input first
A safer workflow starts before the assistant receives anything:
- Choose the right tool. In this scenario, a local spreadsheet lookup may solve the task without sending any file to an external AI service.
- Create a task-specific file. If AI is appropriate and authorized, export only the two required columns, and only the needed records, into a new, clean file. Hiding columns or sheets in the original workbook does not remove them.
- Review what remains. Client names and account numbers can still be sensitive. A smaller file is not automatically suitable for every destination.
- Use the approved service and account. Confirm the organization permits this information in that service, and that the employee is signed in to the approved account, not a personal one. Check the result before relying on it for a business decision.
The aim is not the smallest possible input at any cost. It is enough relevant information to do the job correctly, with nothing unrelated attached.
Make the safe path enforceable
Training helps, but it should not carry the whole burden. Teams also need controls where information is transferred.
Set paste and upload rules for the AI destinations, users and groups that need them, scoped to the sites and browsers the organization manages. Where full-file uploads are unnecessary or prohibited, restrict that transfer and offer an approved alternative. Apply controls to the routes people actually use. An acceptable-use policy alone is not proof that people follow it.
These rules govern the channel. They do not work out which workbook fields a task needs. The organization still has to design the approved workflow and give people a practical way to prepare a narrower input. A blocked upload with no alternative leaves the legitimate task unresolved and invites workarounds.
Then test it with representative work. Confirm that prohibited transfers are stopped, approved tasks can proceed and the available evidence supports a meaningful review of control actions. Coverage varies by browser, device and route, so find where it does not reach.
Draw the boundary, not just the tool list
A useful governance review asks what work people are doing, which tools and accounts they use, what information they supply and which controls apply before it is shared.
The goal is productive AI use with deliberate data boundaries. Nobody should have to choose between finishing the task and following the rules.
Before the next upload, ask one question: does the assistant need this file, or only the part that helps complete the job?
Related: AI security and governance