A client services team wants to use OpenAI to summarize project handover notes. It plans to keep decisions and action items in its existing document system. A reviewer needs to know which people can read the source notes and generated summary, which services receive a copy, and who can delete each copy. The team can answer those questions on a synthetic file before uploading confidential work.
Start with one input class, one approved route, and one output destination. Aeon's OpenAI safety guide compares personal ChatGPT, business workspaces and the API. The walkthrough below shows how to turn that comparison into an approval record for a specific workflow. The examples are hypothetical.
1. Define the decision before testing
Write a short statement of the proposed use. For the handover example, six named members of the client services team may summarize internal project notes in an approved workspace or application. A human checks the summary against the source, then puts it in the project folder. The pilot excludes credentials, employee records, and material whose client terms prohibit use by an external service. Give an owner authority to stop the workflow if the scope changes.
Draw the route from the source file to the final folder. Include the upload, any temporary file, the model request, the result, and any logs or exported copies. For each location, record who has access, what evidence supports the retention setting, and who owns deletion. A diagram with five boxes is enough if it includes the actual destinations. Include the support and incident channels if they can receive prompts, files or output.
2. Verify the controls for the route people will use
ChatGPT workspace. OpenAI says it excludes inputs and outputs from eligible business workspaces and the API from training by default. Check the actual product and workspace that staff will use. Confirm the team accounts, who can share material internally and externally, and what controls the particular plan gives the administrator. For example, OpenAI documents an Enterprise workspace-wide switch for disabling shared links and says ChatGPT Business lacks that switch. Ask the workspace administrator to demonstrate the relevant settings and test sharing from a pilot member's account. OpenAI business data privacy · Sharing and privacy in ChatGPT Business
Check chat history, separately saved files, project files, and exports as different copies. OpenAI says a file saved separately in Library or Space needs its own deletion action. For connected apps, confirm which apps are allowed, which provider account is connected, and whether actions can send confidential material to another destination. Capture the approval settings and the downstream provider's handling rules for the actual app. Chat and file retention · Connected apps
API integration. Record the organization, project, endpoints, storage settings, and any external tool destination. OpenAI documents default abuse-monitoring logs that can contain customer content and are generally retained for up to 30 days, with stated exceptions. The company documents separate retention rules for application state and requires prior approval for Zero Data Retention or Modified Abuse Monitoring. Check the current documentation for the precise endpoint and features, then verify the project setting. For the Responses API, check how the application sets storage and whether background mode, file objects or external tools add copies. Include the integrator's own retries, error reports, observability tools, and backups in the retention map. OpenAI API data controls
3. Test with a synthetic handover file
Create a file with invented names and a unique marker such as SAMPLE-HANDOVER-042. Give it three decisions, two action items, and one deliberately ambiguous sentence. Run the proposed workflow and ask the reviewer to trace each extracted decision to the source. Have a pilot user attempt to share the output with a test account outside the project. Check the final project folder and, for the API route, inspect the application's permitted test logs for the marker. Test any approved tool or connected app separately so you can see the destination for each transfer.
Then exercise the cleanup procedure. Delete the test objects where the workflow permits deletion and record what becomes inaccessible to the pilot users. If a test only shows removal from the interface, record that observation and use the applicable terms and settings as evidence for provider-side retention. Where a copy stays in an internal backup, document who can restore it and when normal expiry applies.
4. Put a decision and a review trigger on the record
Keep the relevant plan or project settings, the flow map, the synthetic test result, and the name of each control owner together. Give the approver a pass condition for every material control. In this example, project members must be able to review the summary against the source, an out-of-scope test account must be denied access, and every persistent copy must have an owner and a retention or deletion rule. For the API route, the synthetic marker must stay out of any logs where policy prohibits document content.
If a check fails, restrict the pilot to synthetic files until the team can demonstrate the correction, or narrow the input class to match what the controls support. Reopen the approval when the workspace plan, API endpoint, data class, connected app, output audience or retention setting changes. Aeon's AI Confidentiality Assessment can help a team trace the actual routes, review the available evidence, and prioritize corrections before launch.