*The useful setting is not the same thing as the safe setting. Here is how to keep the Office integration, while cutting off the most avoidable forms of reuse, sharing and exposure.*
The sales pitch for AI in Excel and PowerPoint is frictionless: analyse the workbook, repair the formula, reshape the deck, keep moving. The privacy reality is less frictionless. A system cannot analyse a worksheet it never receives. It must process the prompt and the relevant document context to do the job. That is the price of the feature—not evidence that your work has become public, but not a reason to stop asking hard questions either.
The sensible target is narrower and more honest: **keep the connection, minimise the data, opt out of secondary use, and remove the ways content can travel further than intended.**
For a personal ChatGPT account, switch off **Improve the model for everyone** on the web: **Profile → Settings → Data Controls**. OpenAI says the choice applies to the entire account and that new conversations will not be used to train its models. Your chat history can still remain visible to you; the control is about model improvement, not the operational processing required to answer a request. [OpenAI’s Data Controls FAQ](https://help.openai.com/en/articles/7730893-chatgpt-shared-links-faq) is unusually clear on that point.
For business customers, the baseline is stronger: OpenAI says it does not train on inputs or outputs from ChatGPT Business, Enterprise, and the API by default, unless the organisation explicitly opts in. That is a contractual and product-tier distinction worth understanding before a finance team treats every ChatGPT surface as identical. [OpenAI’s data-use policy](https://openai.com/policies/how-your-data-is-used-to-improve-model-performance/) sets out the difference.
## First, separate the four privacy questions people keep collapsing into one
| Question | What it actually controls | What it does **not** promise |
|---|---|---|
| Will the content train a model? | The account’s model-improvement setting or a business plan’s default policy | That no content is processed to fulfil the request |
| Will it be used for marketing? | Marketing Privacy and browser cookie choices | That operational providers never process minimum necessary data |
| Will it become public? | Shared links, exports, access permissions and who you send it to | That a user-created public link is somehow private |
| How long does it remain? | Temporary Chat, chat/file deletion and product-specific retention policies | Immediate erasure of every safety, legal or de-identified record |
The phrase “don’t share my data” is emotionally understandable and technically useless. Ask which lane you mean.
## What ChatGPT needs to process when it works inside Excel
OpenAI’s current Excel guidance says ChatGPT processes your prompt, attachments and relevant spreadsheet context so it can update the workbook or provide analysis. It also says some data logs may be retained by OpenAI for up to 30 days for safety and integrity. The same guidance notes that Microsoft may be able to read workbook content, attachments and prompts under the Microsoft Add-In Marketplace terms. That is not a footnote; it is part of the data path. Read it before you put a board pack, customer list, payroll model or acquisition model into the workflow. [ChatGPT for Excel and Google Sheets](https://help.openai.com/en/articles/20001063-chatgpt-for-excel/).
PowerPoint follows the same basic logic: it needs the slide content and prompt context you make available to perform the requested work. Training controls do not make that processing vanish. They limit a secondary use of the interaction after the requested work is done.
That distinction should alter behaviour. Do not throw an entire workbook at a question about one chart. Create a clean copy, remove unrelated tabs, redact identifiers, and ask for the smallest useful slice. This is data minimisation, which remains more dependable than any reassuring toggle.
## The practical privacy configuration
### 1. Stop model-improvement use
For individual ChatGPT accounts, use the web setting:
1. Go to **Settings → Data Controls**.
2. Turn **Improve the model for everyone** off.
3. For an extra account-level record, submit the Privacy Portal’s **Do not train on my content** request if desired.
OpenAI says new chats are not used to train when the opt-out is active. Treat it as a forward-looking control. It does not delete earlier chats, and it does not turn Excel or PowerPoint into an offline tool.
### 2. Stop marketing-oriented sharing separately
Model training and advertising are different systems. In the web interface, review **Settings → Data Controls → Marketing Privacy** and opt out of the targeted-advertising sharing control. Then use **Cookie Preferences** on OpenAI sites to reject non-essential cookies where that suits your preferences.
OpenAI says this marketing activity concerns limited information used to promote OpenAI on third-party properties; it also says chat content is not shared for marketing or advertising purposes. The distinction matters: turn the marketing control off anyway, but do not mistake it for a content-training control. [How OpenAI promotes on third-party properties](https://help.openai.com/en/articles/20001156) and the [Consumer Services FAQ](https://help.openai.com/en/articles/7039943-openai-privacy-policy) explain the boundary.
### 3. Shut the public doors
Neither an Excel integration nor a PowerPoint integration automatically puts your source material into the public domain. Public exposure normally needs a deliberate route: a shared conversation link, copied output, an export, a forwarded deck, or access granted through Microsoft or another connected service.
Audit these routinely:
- **Shared links:** in ChatGPT Data Controls, review and revoke links you no longer need.
- **Conversation history:** delete chats containing material that has outlived its business purpose.
- **Files/library:** delete files separately where your product interface stores them outside the chat.
- **Microsoft sharing:** check workbook and deck permissions, SharePoint/OneDrive links, and add-in access. The AI setting cannot rescue an over-shared Office file.
### 4. Reduce retention and cross-chat recall
For a one-off sensitive task, use **Temporary Chat** where the feature is available. OpenAI says Temporary Chats do not appear in history, do not create memories and are not used for training; they may be retained for up to 30 days for safety review. That is shorter-lived, not magic disappearance. [Data Controls FAQ](https://help.openai.com/en/articles/7730893-chatgpt-shared-links-faq).
For ordinary chats, review **Settings → Personalization**. Turn off Memory and chat-history reference if you do not want information from one discussion resurfacing in another. Then delete the relevant conversation and associated files when they are no longer needed. OpenAI says cleared chats are deleted within 30 days, except where content was already de-identified and disassociated or needs to be retained for security or legal reasons. [Consumer Services FAQ](https://help.openai.com/en/articles/7039943-openai-privacy-policy).
## A harder truth: privacy controls do not abolish operational access
OpenAI says a limited number of authorised personnel and trusted service providers may access content when necessary for abuse/security investigation, support, legal matters, or model improvement when the user has not opted out. It says those providers are bound by confidentiality and security obligations. This is not the same as giving advertisers your spreadsheet. It is also not a licence to upload secrets casually. [Consumer Services FAQ](https://help.openai.com/en/articles/7039943-openai-privacy-policy).
The professional standard is simple: if a document is so sensitive that necessary processing by the service would be unacceptable, do not use the integration with that document. Use an approved enterprise arrangement, isolate/redact the material, or keep the work in a controlled environment.
## The five-minute pre-flight check
Before connecting a live workbook or deck:
- Confirm **Improve the model for everyone** is off (or confirm the organisation’s Business/Enterprise policy).
- Confirm **Marketing Privacy** is opted out and browser cookie choices are set appropriately.
- Remove personal data, account numbers, credentials, unreleased financials, and irrelevant tabs/slides.
- Use the minimum necessary range, worksheet, or slide selection.
- Check that the Office file has no uncontrolled sharing link.
- Avoid leaving sensitive work in a permanent chat when a Temporary Chat is appropriate.
- Do not publish a shared ChatGPT link containing the analysis or source material.
- Review and delete the chat and stored file when the task is complete.
This will not make the integration zero-risk. Nothing useful does. It does make the data path narrower, more legible and less dependent on wishful thinking.
## FAQ (schema-ready)
### Can I keep Excel and PowerPoint connected to ChatGPT and stop model training?
Yes. On a personal account, turn off **Improve the model for everyone** in web Data Controls. OpenAI says new conversations will not be used to train its models. Excel or PowerPoint content still has to be processed to complete the task.
### Does turning off training mean ChatGPT never sees my workbook or deck?
No. The service must process your prompt and relevant document context to answer or make changes. The training opt-out limits model-improvement use; it does not make the add-in work without seeing the relevant content.
### Does Marketing Privacy stop OpenAI from using my workbook in advertising?
Marketing Privacy controls targeted-advertising sharing. OpenAI says it does not share chat content for marketing or advertising purposes. Use the training control separately, because advertising and model improvement are different issues.
### Does connecting Excel or PowerPoint make my file public?
No. A connection is not publication. Public exposure is more likely through shared chat links, copied output, file-sharing permissions, exports or forwarding. Review and revoke those routes.
### How can I minimise retention for a sensitive task?
Use Temporary Chat when suitable, minimise the content supplied, turn off Memory/chat-history reference if desired, and delete the conversation and separately stored file when the task ends. Some limited safety, legal, or de-identified retention can still apply.
## Sources and editorial note
This post was checked against OpenAI’s current public documentation on 28 July 2026. Product interfaces and retention language can change; link to the primary sources rather than treating a blog post as a substitute for the applicable plan terms or an organisation’s security review.
- [OpenAI: Data Controls FAQ](https://help.openai.com/en/articles/7730893-chatgpt-shared-links-faq)
- [OpenAI: ChatGPT for Excel and Google Sheets](https://help.openai.com/en/articles/20001063-chatgpt-for-excel/)
- [OpenAI: How your data is used to improve model performance](https://openai.com/policies/how-your-data-is-used-to-improve-model-performance/)
- [OpenAI: Data Usage for Consumer Services FAQ](https://help.openai.com/en/articles/7039943-openai-privacy-policy)
- [OpenAI: How we promote OpenAI on third-party properties](https://help.openai.com/en/articles/20001156)
Comments