Is ChatGPT for Excel safe for client files? Turn one setting off first

What is ChatGPT for Excel, and how is it different from uploading a file?

ChatGPT for Excel is an add-in, which is a small app that runs inside Excel as a panel on the right hand side. OpenAI put it out in beta on 5 March 2026 and made it generally available on 5 May 2026. That is when it moved to the GPT-5.5 model. It works on every plan, which includes Free, Go, Plus, Pro, Business, Enterprise, Edu and K-12, and this same install covers Word and PowerPoint too.

You install the app by clicking Home, then Add-ins, then searching for ChatGPT, and you sign in with your ChatGPT account. After that, it works on whatever workbook you have open. It reads your tabs and formulas, and it builds new sheets and edits old ones. OpenAI says it points its answers back to the exact cells it used, and it asks for permission before it changes your workbook.

This is different from uploading a file to chatgpt.com, and a lot of the privacy advice online is written for that upload. The add-in keeps its chats separate. OpenAI says spreadsheet chats do not sync with your main ChatGPT history, and they do not have access to ChatGPT memory. So the add-in holds its own data in its own place, and it has its own delete option.

If you read my piece on what Claude in Excel can and cannot touch, you know this is OpenAI’s take on the same idea. The questions you need to ask before you trust it with a client file are exactly the same.

Where does your workbook go when you ask ChatGPT a question in Excel?

To give you an answer, the add-in has to read the workbook. OpenAI’s help page for ChatGPT for Excel says ChatGPT processes your prompt, any attachments and the relevant spreadsheet context. That context means the cells and sheets it needs to answer your question. So if you ask a question about the whole model, it has to read the whole model.

That content leaves Excel. The ChatGPT listing on Microsoft Marketplace says this plainly, explaining that when the app is used, it can read and make changes to your document and send data over the Internet. From there, three things happen, and every one of them is written on a page you can look up yourself.

The first thing is OpenAI’s side. Their help page says some of your data logs might be stored with OpenAI for 30 days for safety and integrity purposes. The page does not list which logs, and it does not tie this to any specific plan. I read this as a rule that applies to everyone.

The second thing is Microsoft. OpenAI’s page says that under Microsoft’s Add In Marketplace Terms of Service, Microsoft might have the ability to read your Excel workbook, attachments and prompts. The Marketplace listing says it slightly differently, noting that data like chats, attachments and workbook content may be shared with Microsoft and OpenAI. This is to let them read and update your files.

The third thing stays with you. OpenAI says your data is stored locally in the product, and you can delete it or download it from the Settings in the add-in. This is helpful when a project ends. It means you can clear out what the add-in holds about a client instead of just leaving it there.

Will OpenAI train its models on your client’s numbers?

This is the part that depends on your plan. The Marketplace listing says ChatGPT for Microsoft apps respects your ChatGPT data training preferences. In normal words, the add-in just follows whatever training setting your ChatGPT account already has. So the real question is what that setting is on your own account.

OpenAI’s page on how your data is used splits its services into two parts. For individual services like ChatGPT, it says it might use your content to train its models, and you have to opt out if you do not want this. For business services like ChatGPT Business and ChatGPT Enterprise, it says by default it does not train on any inputs or outputs.

This setting lives in ChatGPT itself, not in Excel. You open ChatGPT, go to Settings, then Data Controls, and turn off Improve the model for everyone. OpenAI also has the same opt-out in its privacy portal under the name Do not train on my content. OpenAI says you only need to use one of these for it to work.

Two details on that page change how you use this switch. The first is about timing. OpenAI says that after you opt out, it will not train on your new conversations, so the change only works going forward. It does not protect the chats you had about a client file last week, so you must turn it off before you open the file.

The second detail is about feedback. OpenAI says that even after you opt out, giving a thumbs up or thumbs down on an answer might let them use the entire conversation to train its models. One click on a good answer puts the whole conversation back into training. This means your client numbers go with it.

When you do client work, do NOT rate the answers, even the good ones.

What changes if you use it on ChatGPT Business or Enterprise?

The training question has a different default on a business plan. OpenAI does not train on business inputs and outputs unless the organisation chooses to opt in. Their launch post for ChatGPT for Excel repeats this for Enterprise, stating that by default, data shared with ChatGPT Enterprise is not used to train or improve its models.

Enterprise also adds the controls that a client’s IT team will usually ask about. The help page lists data and inference residency where available, which keeps the data processing in a chosen region. It also lists Enterprise Key Management for the encryption keys, and role based access controls so an admin can decide who gets the add-in. Your prompts and responses are also put in the Compliance API, which is a data feed a company uses to check what its people did.

This last point is worth a comparison. When I looked at Claude for Excel, Anthropic said its add-in activity is not included in Enterprise audit logs, and its Compliance API coverage is still in public beta. OpenAI says plainly that the Excel add-in’s prompts and responses are in its Compliance API. If a client ever asks you to prove what you did with their file, that difference matters a lot.

There are two things that do not change with your plan. The Microsoft line on OpenAI’s page is tied to the Marketplace add-in, and the rule about 30 day logs does not name any plan either. A business plan settles the question about training. But the page does not say it changes the other two rules.

What will ChatGPT for Excel not do well, even on the right account?

The rules about data are only half the picture. The other half is whether the tool can actually do the job on your file. OpenAI lists its own limits for this on the same help page.

Macros and VBA come first, as OpenAI says these features may not be fully supported. VBA is the code behind a workbook that runs when you click a button or open the file. If a client’s workbook relies on a macro to do its real work, do not expect the sidebar to handle it. Claude hits this exact same limit, and Anthropic lists macros and VBA as completely unsupported.

Size is the second limit. OpenAI says usage depends on the size and complexity of your task, meaning larger workbooks, multi-step edits and complex analysis may use more of your agentic usage limit. Free and Go include limited usage, and Plus and Pro fall under the plan’s agentic usage limit, which OpenAI says is shared with its other agentic features once their pricing starts. So a large client model is exactly where the size counts against you.

Memory and history are the third limit, and they matter most when you share work. Spreadsheet chats do not have access to ChatGPT memory, and the Marketplace listing says that Skills from ChatGPT do not automatically carry over to the add-in. This means the preferences and rules you taught ChatGPT on the web are missing from the sidebar. You have to set them up there all over again.

The last limit is the one every vendor prints today. OpenAI’s page says outputs may be incomplete or incorrect, and it tells you to check formulas, calculations, citations and changed cells before you rely on them. It also says you should duplicate the file first for important work so you can revert to the old version. The quick way to test a number it gives you is in how to check if ChatGPT or Copilot got your Excel numbers right.

What should you do before you open a client’s workbook with it?

When you put this all together, this is the order I suggest. None of it takes very long. Also, every single step comes straight from one of the vendor lines I mentioned above.

1. 01
   Sign in with the account you mean to use

   The add-in follows the training setting of the ChatGPT account it is signed into. A client file opened under your personal account follows your personal setting.
2. 02
   Switch training off on a personal plan

   In ChatGPT, open Settings, then Data Controls, and turn off Improve the model for everyone. Do it before the first question, because it only covers new conversations.
3. 03
   Ask the client if the contract limits where data goes

   On every plan the add-in sends what it reads to OpenAI, and OpenAI says Microsoft may be able to read it. If the data has to stay with the client, that decision is theirs.
4. 04
   Work on a duplicate

   OpenAI tells you to duplicate the file first for important work. Keep the original untouched until you have checked every changed cell.
5. 05
   Do not rate the answers

   A thumbs up or down can send the whole conversation for training, even after you opted out.
6. 06
   Clear the add-in when the job ends

   Its Settings let you download or delete the data it stored locally. Clear it once the client work is finished.

If your client’s data is the kind that should never leave their company, the safest route has not changed at all. You can give the model a small made-up sample with the same columns, get the method back, and then run that method on the real file yourself. This is how I use the AI tools I keep for data analyst work. It takes a little more time, but it is much easier to explain to a client.