Answer a question from your own files with ChatGPT apps
The answer to your question is in a document somewhere in the shared drive, written by someone who has since left. Connecting ChatGPT to that drive lets you ask directly — as long as you make it name the files it used, and you go and look at them.
What you'll make
one checked answer to a real question, drawn from your own files, with each source document named, quoted and opened by you.
What you'll need
a ChatGPT work account, a file service such as Google Drive or SharePoint that your organisation lets you connect, and a question with an actual answer. About 20 minutes. In a managed workspace, an administrator may need to enable the app first.
Example result.
Finance Approvals Policy v4.docx (modified 12 March 2026): “Purchases with a total value above £5,000 require written approval from the Finance Director…” Procurement Handbook 2024.pdf (modified 8 November 2024): “Approval is required for purchases above £2,500.” Approvals — DRAFT update.docx (modified 29 September 2026): “Threshold to be raised to £10,000 (pending sign-off).” These disagree. I haven’t chosen between them.
A naming note: OpenAI now calls the things that connect ChatGPT to a service such as Google Drive apps, and the packages that bundle apps with reusable instructions plugins. People still say “connectors”. If a colleague’s menu path doesn’t match yours, this is usually why.
1. Connect the account that can already see the files
Open Settings › Plugins, select the app or its plugin, choose Install plugin if shown, then Connect. Sign in to the account that has the files, review the permissions and complete the provider’s authorisation.
If you have more than one account with the provider (a personal Google account and a work one, say), connect the one with the files. If authorisation is blocked, the cause is often on the provider’s side: for Google Drive, a Google Workspace administrator separately approves the ChatGPT application. That may not be the same person as your ChatGPT admin, which is worth knowing before you file a ticket.
Then set the permission level. Options include Always ask, Allow read actions and Allow low-risk actions. This task needs only search and reading. App permissions never grant an app new access: they only change when ChatGPT asks before using access it already has. Allow read actions doesn’t remove write permissions the service granted, so ask an administrator to restrict them if your policy needs a read-only connection. To remove access, disconnect the app.
Why this works: the wrong account or the wrong administrator are the two most common reasons the files “aren’t there”.
What you should see: the app connected to the right account, with read access only in use.
2. Ask one bounded question, and demand the file names
The usual problem isn’t a wild invention. It’s a fluent answer assembled from two relevant documents and one that was superseded long ago. So the prompt’s job is to make the sources visible.
Mention the app with @, or select + and choose it. Then:
Search my [file service] and answer this question: [your question].
Name every file you used, with its title as it appears in the drive and a source link. Include the last-modified date if available; otherwise write "date not available".
Quote the sentence or figure each part of your answer rests on.
If two documents disagree, show both and say which is more recent. Don't resolve the disagreement for me.
If you can't find a document that answers it, say so and list what you searched for. Don't answer from general knowledge.
Flag anything that looks superseded, in draft or undated.[file service] is “Google Drive”, “SharePoint” or similar; [your question] is the one question you need answered.
Why this works: quoting the supporting sentence is what makes checking possible at all. Refusing to resolve a disagreement matters because the resolution is the decision, and the decision is yours. “Don’t answer from general knowledge” stops a plausible generic answer to a question about your own contract.
What you should see: a list of named files, each with a quote, and any disagreement shown rather than settled. If ChatGPT asks for approval before reading, read the request rather than clicking through.
3. Open the files
Open the source behind every claim that could change your decision. Find the quoted passage, check its scope and effective date, and find out whether it’s current. A recent edit to an old draft doesn’t make it policy, and a newer date alone doesn’t make a document the one that counts.
Why this works: OpenAI notes that finding a file name doesn’t confirm ChatGPT retrieved the contents. An answer that names a file isn’t, by itself, an answer drawn from that file.
What you should see: each quote, word for word, in the file it was credited to.
Check the result
- Every claim has a named file and a quote.
- Each quote appears in that file, in context.
- Any disagreement is shown, with drafts and old versions labelled.
- You know where the answer is allowed to go: your own next step is fine, but “the policy says” in a meeting needs the document, not the summary.
The usual failure: a quote credited to the wrong file, or a file named with no quote at all. Spot it by searching the opened file for the quoted words.
That quote doesn't appear in [file name]. Search again and give me only claims you can support with an exact quote from a named file. Where you can't find support, write "no source found" for that point.Worked example: the current sign-off threshold
Fictional example. A new team lead needs to know the spend above which a purchase needs finance sign-off. Three colleagues have told him three different numbers, which is how he knows the real answer is in a document. He connects his work Google Drive, keeps it to reading, and asks the step 2 question.
The first answer said: “The threshold is £10,000. Finance Approvals Policy v4.docx: ‘Threshold to be raised to £10,000 (pending sign-off).’” He opened v4. The sentence isn’t in it; v4 says £5,000.
He used the repair prompt with “Finance Approvals Policy v4.docx”. The second answer is the example result at the top of this page: three files, three figures, one marked as a draft, and no number chosen for him.
He records that the 2024 handbook conflicts with v4, without assuming its age means it was replaced, ignores the draft’s £10,000 because it’s pending, and sends the policy owner a two-line message with links to the conflicting files. That took a message instead of a meeting.
- The quote was real; the file was wrong. It came from the draft. Only opening v4 showed it.
- A single number would have hidden the disagreement. Had he asked “what’s our approval threshold?”, the likely answer was one confident figure. It might even have been right. He’d have had no way to tell, and the version that picked the draft figure reads exactly as convincingly as the one that picked v4.
- The search is only as complete as the drive he can see. Nothing here proves there’s no fourth document elsewhere, which is why the owner confirms.
For the same check on files you upload rather than connect, see use business files and check the answer.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Live access and administrator-managed sync
A personal connection to Google Drive is live access: ChatGPT searches when you ask, and nothing is indexed in advance. Eligible managed workspaces can set up administrator-managed sync, which indexes selected Drive content for the whole workspace using a Google Workspace service account; personal accounts can’t configure it. Two consequences. The first index can take hours, and changes to files, sharing or group membership take time to appear, so a document shared with you this morning may not be findable this afternoon. And sync never grants access: it doesn’t let anyone see files they can’t already open in Drive.
What the app sees, and what happens afterwards
An enabled app may receive context from ChatGPT for your request, including relevant memories if Memory is on, plus the basic information a website usually sees, such as approximate location and browser type. Data shared with an app falls under that app’s own terms. If Memory is on, ChatGPT may remember information it read from the app. Disconnecting stops future access; it doesn’t delete past conversations that used the app’s content. Business, Enterprise and Edu content is excluded from model training by default, and OpenAI applies further restrictions to Google data; see its Google app data controls FAQ. None of this replaces your organisation’s permission to use its files.
Questions this is bad at
Completeness questions. “How many contracts have notice periods under 30 days?” needs an exhaustive sweep, and what you get is a search result. Use it to find candidates, not to produce a count.
Negative questions. “Is there anything in the drive about X?” can’t be answered. A search returning nothing means the search returned nothing. You can narrow it — “search for these five terms and tell me what you searched” — but you can’t turn it into proof of absence.
“Most recent” questions. The index and permissions may lag. If recency is the point, check the date yourself in the source system.
What it’s genuinely good at is the middle ground: finding the three documents that bear on a question and showing where they disagree.