You can use an AI assistant to improve an email, organize rough notes or create a document outline without giving it the complete customer file. Start with an approved tool, supply only the information the task needs, and check the result before using it. That workflow is more useful than looking for one privacy switch that supposedly makes every upload acceptable.
This guide walks through those decisions using fictional workplace examples. It covers personal-account settings, managed workspaces, practical prompts and the extra questions that arise when an assistant can access files or take actions. The examples are original teaching material, not reports of product testing or real customer incidents.
The product documentation was checked on October 25, 2026. Menus and available controls can vary by plan and organization. Follow your employer’s approved process for the task and data involved; a tool’s technical capability does not establish permission to use a particular company document.
In this article
- 1. Check the workspace before choosing the task
- 2. Prepare a clean input instead of attaching everything
- 3. Understand what privacy settings actually change
- 4. Use this prompt for an email or a service report
- 5. Summarize documents with an evidence trail
- 6. Separate drafting from connecting apps and taking actions
- 7. Review the output and measure the whole job
- 8. Your first practical session
1. Check the workspace before choosing the task
Ask which AI service, account and workspace your organization approves, what information it permits there, and who can answer questions about retention or connected apps. If the answers are unclear, use public or invented material while seeking guidance. A work-related question typed into a personal account remains a submission to that service.
Microsoft’s enterprise-data-protection documentation illustrates why the exact environment matters. For its organizational Copilot offerings, Microsoft describes contractual protections, administrative settings and permissions, with specific controls depending on the subscription. It says prompts, responses and Microsoft Graph data are not used to train foundation models. It also distinguishes web-search queries and agents, which have their own handling considerations.
That is a description of Microsoft’s documented offering, not a blanket approval for every workplace task or a comparison proving one provider best. Record the actual workspace you are using. Then choose one narrow job, such as drafting a generic follow-up email, before considering a whole folder of company records.
Sources: Microsoft Learn — Enterprise data protection in Copilot ↗
2. Prepare a clean input instead of attaching everything
The UK National Cyber Security Centre’s foundational guidance advises against putting sensitive information into public LLM queries. Its broader point remains useful: understand the service’s terms and handling before supplying sensitive material. The guidance dates to 2023; use current provider documentation for present-day product settings.
Here is a practical preparation method. Create a new text note containing only the facts necessary for the task. Replace people, companies and sites with placeholders such as CLIENT_A and SITE_B. Leave out passwords, access tokens, customer contact details, internal access instructions and unrelated history. Keep the mapping to real names outside the AI conversation.
Removing names alone is not an assurance of anonymity. A distinctive location, date or combination of events can still reveal what a passage concerns. If those details matter to the task and you cannot confirm that their use is approved, ask the assistant for a generic structure using invented facts instead.
For documents and screenshots, inspect what you are actually submitting, including visible names, side panels and other exposed details. Where possible, send the clean excerpt you prepared rather than the original file. The goal is to control the input, not to persuade the assistant to forget information after it has already received it.
Sources: UK NCSC — Public LLM queries and sensitive information, March 2023 ↗
3. Understand what privacy settings actually change
For a personal ChatGPT workspace, OpenAI’s current instructions place model-training preferences under Settings, then Data controls. Turning off the setting labeled Improve the model for everyone prevents new conversations from being used to train OpenAI models. It does not delete existing saved chats. OpenAI also distinguishes memory and personalization controls from training preferences.
Temporary Chat is another choice, with a different purpose. While a chat remains temporary, it stays out of history and is not used to improve models; OpenAI may retain a copy for up to 30 days for safety. Current documentation allows a personalization choice before the conversation starts and explains that saving the chat converts it into a regular conversation.
OpenAI also says data sent to third parties through GPT actions follows the recipient’s privacy policy. The temporary-chat retention description therefore cannot be assumed to cover every external destination. Check the actual feature, workspace and recipient involved.
These controls answer different questions: training concerns model improvement, retention concerns how long information is kept, and access concerns who or what can use it. None of these settings, by itself, replaces your organization’s decision about what may be submitted. Check the relevant controls before entering data, rather than treating the absence of a chat from history as proof it was never retained.
Sources: OpenAI Help — Data controls in ChatGPT, checked October 25, 2026 ↗ · OpenAI Help — Temporary Chat, retention and third-party actions ↗
4. Use this prompt for an email or a service report
A useful prompt specifies the task, audience, facts and limits. Here is a fictional example you can adapt with approved material: “Draft a professional customer update in plain English, maximum 120 words. Use only these facts: the inspection is complete; one issue remains unresolved; a follow-up visit must be arranged. Do not invent a cause, appointment date, price or promise. Put unanswered questions after the draft.”
The assistant can help organize that message without knowing the customer’s identity. Add the real recipient, reference number and confirmed arrangements later in your normal work application. Check those details yourself before sending. If the intended task requires protected information, return to the workspace and approval decision first.
For a report, try: “Rewrite these fictional notes into Findings, Work completed and Open issues. Preserve uncertainty. Notes: a camera was intermittently offline; the connection returned after a switch-port check; the root cause was not confirmed; follow-up monitoring is required. Do not change a suspected cause into a confirmed diagnosis.”
The review question is concrete: did the rewrite preserve what is known? “Connection restored” and “fault permanently repaired” make different claims. Reject a polished sentence if it adds evidence you do not have. These prompts are suggested drafting templates, not security controls or guarantees of accurate output.
5. Summarize documents with an evidence trail
Once a document is approved for the chosen workspace, define the summary’s scope. Ask for findings tied to the supplied material and require missing information to remain visibly missing. You still need to inspect the source: asking for citations is a way to organize checking, not proof that every reference is correct.
A reusable prompt is: “Summarize only the supplied document. Separate explicit statements, proposed actions and unresolved questions. For each important claim, give the page or section where I can verify it, using the document’s actual labels. If the evidence is missing or unreadable, say so. Do not supply a deadline or owner that the document does not identify.”
Then open the cited places and compare them with the summary. Verify quantities, dates, responsibilities and exceptions. A general policy statement may have a narrower condition elsewhere in the document. If a scan is unclear or a table is extracted badly, treat the affected detail as unresolved until you inspect it directly.
NIST’s Generative AI Profile identifies confabulation: confidently presented false or erroneous content. Its discussion supports retaining a verification step even when the answer reads fluently. For an operational decision, use the authoritative document and an appropriate reviewer, with the AI summary serving as navigation rather than the final authority.
Sources: NIST — Generative AI Risk Management Profile, July 2024 ↗
6. Separate drafting from connecting apps and taking actions
Asking for an email draft is a different task from granting access to a mailbox and allowing messages to be sent. Before connecting a service, establish which data the assistant can retrieve, what actions it can perform and how that access can be removed. Check the connection’s scope rather than inferring it from the task you have in mind.
The NCSC’s May 2026 discussion of agentic AI recommends bounded pilots, least privilege and meaningful human oversight. It emphasizes that agents can access systems and take actions, so broader autonomy can increase the consequences of a failure. The practical implication is to start with a clearly defined task and the minimum necessary access.
For a first workflow, a sensible boundary is to generate a draft for review and keep sending in the normal work application. If an approved later workflow adds actions, define the destination, permitted change and review point. A request to improve wording should not silently become a request to contact customers or alter records.
For more background, Trevioniq’s reporting on OpenAI’s training-environment incident explains the distinction between access boundaries and monitoring. Its passkey guide covers account sign-in and recovery. Those are related layers: securing the account, limiting the connected tools and checking the output each solve a different part of the problem.
Sources: UK NCSC — Thinking carefully before adopting agentic AI, May 2026 ↗ · Related Trevioniq reporting — Agent containment and monitoring ↗ · Related Trevioniq guide — Passkeys, lost devices and account recovery ↗
7. Review the output and measure the whole job
Before using a draft, compare it with the input. Check names and figures you add later, factual assertions, commitments, unresolved issues and the intended recipient. Inspect attachments as part of the final sending step. A correct draft placed in the wrong thread is still a failed workflow.
Measure usefulness across the whole task: preparing an approved input, generating a response, reviewing it and correcting it. Fast generation can be outweighed by a long verification process. Start with one repeatable, low-stakes writing task and compare the complete workflow with your normal method. This guide does not claim a measured productivity gain.
If confidential material is submitted by mistake, stop adding information and use your organization’s incident-reporting route. Record which service, account and material were involved so the responsible team can assess the situation. Deleting a visible chat should not be treated as proof that every retained or externally shared copy has disappeared.
Sources: OpenAI Help — Data controls in ChatGPT, checked October 25, 2026 ↗ · OpenAI Help — Temporary Chat, retention and third-party actions ↗ · NIST — Generative AI Risk Management Profile, July 2024 ↗
8. Your first practical session
Choose an approved tool and ask for a generic email structure using fictional facts. Inspect the answer for invented details. Adjust the prompt until the output preserves uncertainty and clearly separates a draft from unanswered questions. Save the reusable prompt in your normal approved work notes.
For a later real task, repeat the same decisions: confirm the workspace, prepare the smallest permitted input, generate the draft, verify its claims and review its destination. Expand the workflow only when there is a clear reason to do so. The useful habit is a short, repeatable process that you can explain to a colleague and check yourself.