A practical data-handling check before using an AI tool

Classify the input before opening the tool

The relevant question is not simply whether an assistant is popular. It is whether this account, workspace and configuration are approved for this information and task. A personal account and an employer-managed workspace should not be assumed to have identical controls.

Use your organization's actual classification scheme. If one does not exist, start a review with four working labels: public, internal, confidential and restricted. These labels are a discussion aid, not a newly invented company policy. Never paste passwords, API keys, payment-card details or other secrets into a prompt.

Example: summarize a support issue

A synthetic ticket contains a customer's name, address, invoice number, phone number and a description of a device failing after a software update. To prepare a troubleshooting summary, the assistant may need the device family, software version, observed error and attempted remedies. It usually does not need the address or payment reference.

Create a minimized copy: "Customer A; device family D; version V; error E; restarted twice; failure recurs after ten minutes." Keep the mapping to the real customer in the authorized support system. Removing a name alone does not guarantee anonymity: a unique combination of details can still identify someone.

Complete this tool approval note

Task and business owner:
Data classification and reason:
Minimum fields needed:
Approved workspace and account:
Retention / training / sharing settings checked by:
Connector permissions and intended recipients:
Output storage location:
Deletion or incident escalation process:

Ask the responsible administrator to verify the current provider settings and contractual terms. Do not copy an old screenshot as evidence that today's account has the same controls. Connected drives, shared conversations and exported outputs can create additional disclosure paths even when the original prompt was acceptable.

Review the output too

An output can reintroduce personal details, quote confidential passages or combine information in an unexpected way. Check it before posting to a public channel or pasting into a shared document. Keep the final artifact in the approved system and apply the organization's retention rules to both inputs and outputs.

Success is a useful summary with no unnecessary identifying details, an approved destination and a named reviewer. If the task cannot be completed without information the tool is not approved to receive, stop and choose another process rather than weakening the classification.

Sources and review

Reviewed October 16, 2026. Ticket and approval note are original EAIRI examples. Context for assigning and reviewing risk: NIST AI RMF. This is an operational checklist, not a legal compliance determination.