Use the exact AI tool, account and configuration approved for the task and its data. Then send only what the task needs and check where the result will go. Approval for one product feature does not automatically cover a new connector, repository index or destination.
Start with your workplace policy. This guide turns that policy into a check on an actual input, including files and background connections.
Trace the whole path
Write one line describing the proposed flow: source, account, AI service, connected recipients and output destination. Identify who can access prompts, files, outputs and retained records along that path.
Check the applicable contract and configuration for training use, feedback, retention, deletion and administrative access. ChatGPT’s Data Controls FAQ and Temporary Chat FAQ explain separate controls. Turning training off or selecting a temporary mode is not permission to process workplace material, and neither name establishes zero retention.
The ICO’s internal policy is a useful real example of exact-tool and data permission. Ireland’s National Cyber Security Centre 2026 AI deployment guidance also treats acceptable use, source handling and incident response as parts of the workflow. These are references, not a replacement for your employer’s rules.
Classify this input, not the app’s reputation
Use your organization’s categories. As a starting check, distinguish public material, internal working information, confidential business or personal information, and restricted material requiring a designated specialist decision. Those labels are not universal legal classifications.
Ask what is actually present: customer identifiers, staff details, private figures, credentials, internal architecture, contract conditions or protected records. A public document can still contain personal data or content you lack rights to reuse. A confidential attachment does not become public because its filename looks ordinary.
If approval is unresolved, keep that material out and ask the responsible owner. Avoid using a personal account as a workaround for a blocked work feature.
A minimal coding example
This is a fictional teaching example.
An engineer’s real debugging bundle contains source code, a staging address, a token and a customer record. The question is merely how to explain an out-of-range number. A separate toy example can express that issue without the bundle:
Fictional rule: count must be an integer from 1 to 5.
Test input: count = 7.
Task: Suggest a clear validation message using only that rule.
Checked teaching output: “Enter a whole number from 1 to 5.” No repository, hostname, token or customer is needed to answer that wording question.
If the real bug depends on private code or configuration, the toy example may not reproduce it. Use the authorized development path for that work. Do not call a changed problem solved because the simplified example worked.
Inspect attachments and reduce only what is unnecessary
Open the file you intend to send. Check hidden sheets, comments, tracked changes, attached pages and metadata where relevant. Prepare a permitted working copy without destroying records that must be retained.
The ICO’s minimisation guidance explains keeping personal data relevant and limited to the purpose. Changing names or rounding values does not prove anonymity. Rounding can also destroy the relationship you wanted to analyze. If real values are necessary, use the approved path; otherwise create synthetic practice data from scratch.
A “local AI” label needs the same path check: where do inference, indexing, telemetry, plugins, logs and backups actually operate? Local execution does not decide workplace permission.
Review the output and handle incidents
Verify material facts and commitments before the result enters another system. A customer-facing message, policy summary or task assignment can have effects beyond the draft. Follow required review or the explicitly authorized automated workflow and its controls.
If data may have crossed a boundary, stop the affected use and promptly use the designated incident route. Record the tool, account, time and data category without repeating the disclosure. Preserve relevant evidence and use the authorized route to revoke or rotate exposed credentials. Let the responsible responders determine notification and deletion obligations.
Use the data-hygiene checklist for recurring work. Recheck when a material part of the path changes, not merely when someone remembers to revisit a generic privacy setting.
