What happens when you upload files to ChatGPT?
Uploaded files become copies attached to a conversation, project or custom GPT. OpenAI's help centre lists the limits: documents up to 512 MB per file, text capped at 2 million tokens per file, up to 40 files per project on Business and Enterprise plans, and up to 20 files as knowledge for a GPT. When the original policy changes, the uploaded copy stays as it was. Someone has to remember to replace it in every project and GPT that holds it, or staff keep getting answers from the old version.
How do the SharePoint and Google Drive apps differ?
Connected apps read files where they live instead of copying them in. OpenAI states that the SharePoint app only retrieves content the user's Microsoft account can already access, with SharePoint, OneDrive and Microsoft 365 group permissions still applying. Its SharePoint search supports .txt, .pdf, .docx, .pptx, .xlsx and .csv files up to 100 MiB each. Your Microsoft administrator may need to grant consent, and the ChatGPT workspace administrator decides which actions, such as creating folders or updating files, are allowed. Answers draw on whatever the user can open, whether or not it is the current version.
Is our data used to train OpenAI's models?
OpenAI says it does not train its models on business data from ChatGPT Business, Enterprise, Edu and the API platform by default. For Enterprise and Edu workspaces, administrators control how long data is retained, and OpenAI says deleted conversations are removed from its systems within 30 days unless it is legally required to keep them. Free and personal accounts are a different case. The OAIC recommends, as best practice, that organisations do not enter personal information, particularly sensitive information, into publicly available generative AI tools.
Where does a built knowledge search fit?
A built system indexes an approved document set, such as the current procedure library, and answers only from it. Because it is a shared service, the business can test it against real staff questions, set a threshold below which it says the documents do not cover the question, and require a citation for every answer. Updates flow from the source on a schedule, and documents marked as superseded can be filtered out. The trade-off is a build and an internal owner, which a ChatGPT workspace does not need.
What goes wrong with the do-it-yourself route?
Problems with this route tend to be organisational. Different staff upload different versions of the same document to their own projects. A departing employee's projects hold the only copy of a useful set of files. Nobody can say which answers came from which version. Personal accounts get used for work files because they are quicker to open. None of this shows up in a demonstration. If you take this route, put a few rules in place from the start.
- Keep a register of shared projects and GPTs, with an owner for each.
- Record which document version each one holds, and replace it when the source changes.
- Allow work files only in business workspaces, never in personal accounts.
Which should you choose?
Use uploads for personal research and drafting with a few files. Use a business workspace with connected apps when staff need help across documents they can already open and your permissions are in order. Choose a built knowledge search when one approved set of documents must give consistent, checkable answers to many people, or when you need to measure accuracy before relying on it. Some businesses run both, using ChatGPT for drafting and a knowledge search for policy answers.