Google Drive Folder Access for AI Agents: Separate Read and Write Scopes
Grant an agent folder-level Google Drive access with separate read and write scopes, plus a review cadence that catches permission drift.

To give an AI agent access to Google Drive, share only the specific folders it needs with the agent's own Google identity, never with a personal login and never at the root of the drive. Use a read-only role on the folder that holds source material and a separate writable folder for drafts or outputs, so the agent cannot change the originals it works from. Then review the sharing list on a fixed cadence, because folder permissions drift as people move files, create links and change roles.
Key takeaways
- Give the agent one folder for reading and a different folder for writing, and never share a parent folder that contains both.
- Use a dedicated agent identity so every action in Drive can be traced to the agent rather than to a person.
- Assign Viewer for source material and Editor only on an outputs folder that a person reviews before anything leaves it.
- Compare the actual sharing state against an approved access matrix every week, and have a named person own each difference.
- Treat any move, copy, link change or ownership transfer on a granted folder as a change that needs an immediate check.
Why folder scope works better than file-by-file sharing
Google Drive permissions flow downward. When a folder is shared, everything inside it is shared, including files added next month by someone who never thought about the agent. Sharing individual files looks safer at first, but it breaks down quickly. The agent needs new files as work arrives, and every new file becomes a manual exception that someone eventually forgets to grant or to remove.
Folder scope matches how teams actually organize work. A reference folder with price lists, procedures and templates is one clear boundary. An outputs folder where drafts land is another. For each folder, the question is simple: what does the agent need to do here, and what happens if it does that wrong?
This is the same logic behind least privilege for any agent: start with the smallest access that completes the task, then widen only with evidence. The principle is covered in more detail in least-privilege permissions for AI agents, and the Google Workspace specifics are in AI agent access to Google Workspace and Microsoft 365.
How do you set up separate read and write folders?
The setup takes less time than the review that follows it, and it is worth doing in this order.
- Step 1: Create a dedicated Google identity for the agent, used only for this role, so its actions in the audit history are not mixed with a person's actions.
- Step 2: Create a reference folder and place copies of the source documents in it. Avoid sharing the company's live working folder, which holds files nobody has reviewed for agent use.
- Step 3: Create a separate outputs folder outside the reference folder, not nested inside it, so inherited permissions cannot blur the two boundaries.
- Step 4: Share the reference folder with the agent identity as Viewer, and share the outputs folder as Editor. Share nothing else.
- Step 5: Turn off link sharing on both folders. Do not use anyone-with-the-link access for an agent.
- Step 6: Record each folder ID, its role, its owner and its purpose in the access matrix before the first run.
The matrix is the document everything else is checked against. Without it, a review can only ask whether the sharing looks reasonable, which is a weak test.
Which Google access method should the agent use?
There are two common patterns. With delegated OAuth, the agent acts through a user account that has granted consent to a specific set of scopes, and it can do what that user can do within those scopes. With a service account, the agent has its own identity and sees only what has been shared with that identity. Google's OAuth 2.0 documentation describes how applications request scopes and how users grant consent, and that consent screen is the moment to read the list carefully.
For Drive, the scope choice matters more than the method. Google offers read-only scopes and broader scopes that allow changes to files. Choose the read-only scope wherever the task allows it, and check Google's current scope list before you finalize, because scope names and their classifications change over time. Do not use an employee's session as a shortcut. When that person changes roles, the agent's access changes with them, and the audit trail blames the wrong person. The trade-offs between the two patterns are covered in service accounts or OAuth tokens for AI agents.
What may the agent write, and what must stay with a person?
Write access is where most of the risk sits, so the rules should be explicit before the first run.
- If the agent only needs to read a file, grant Viewer on the folder that holds it and nothing more.
- If the agent needs to create drafts, let it write only new files into the outputs folder, and never overwrite a file in the reference folder.
- If the agent needs to rename or move files, leave that out of the first version. A person does the filing after review.
- If the agent needs to share a file with someone outside the company, it prepares the request, and a person sends it.
- If the agent needs to delete anything, the answer is no. Editor access generally lets the holder remove files it can edit, so the outputs folder should hold only material the team can regenerate, and deletion stays outside the agent's reach.
For drafts, use a naming convention that includes the task identifier and a version number. That makes it possible to tell what the agent produced, when a person reviewed it, and which version was used. The broader question of how a draft becomes a sent document is covered in read, draft and send: three different AI permissions.
What is permission drift, and why does it go unnoticed?
Permission drift is what happens when the real sharing state of a folder slowly stops matching the approved matrix. Nothing fails at the moment it happens, so no one is alerted. The most common causes are ordinary office actions.
- Someone moves the outputs folder into a broader team folder to make it easier to find, and the drafts become visible to people who should not see them.
- A reference folder is copied into a shared drive where other roles hold Editor access.
- Link sharing is switched on during a one-off handoff and never switched back.
- The person who owned a folder leaves, ownership transfers, and nobody revisits the agent's grant.
- A new subfolder appears under the reference folder with files that should not be visible to the agent, or a needed folder falls outside the grant, so the agent fails and someone creates a risky workaround.
Illustrative scenario: a small purchasing team has an agent read supplier price sheets from a reference folder and write comparison drafts to an outputs folder. Later, a coordinator moves the outputs folder into a team folder with a wider audience to make it easier to find. The agent keeps working, so no one notices. The next weekly listing shows an extra group with Viewer access to the outputs folder, which the matrix does not include. The owner moves the folder back, and the matrix gets a new rule: the outputs folder may not be moved without approval.
How does Be Human's managed approach handle folder access?
In a Be Human engagement, the first deliverable names each folder the agent can reach, the role it holds, the person who approves changes and the check that confirms the grant still matches. The agent runs under its own identity, and the credentials stay under the client's control. Managed operations add the recurring work: a scheduled listing of sharing state, a comparison against the approved matrix, incident response when a difference appears, and a controlled update when a new folder is genuinely needed. The approach does not replace human judgment about what the agent should see. It makes the boundary visible and reviewed on a schedule.
What review cadence catches drift in time?
A review cadence works only when each level has a clear owner and a clear output. The NIST AI Risk Management Framework treats monitoring and periodic review as part of managing AI risk, not an optional addition after deployment. A practical cadence for agent folder access looks like this.
- Weekly: an automated listing of every folder the agent can reach, with its role, link setting and owner, compared against the matrix. Each difference goes to a named person as an alert.
- Monthly: a person reviews the flagged items and any folders the agent used that are not in the matrix, then confirms the matrix still reflects real work.
- After every reorganization: any move, rename, copy into a shared drive or ownership transfer involving a granted folder triggers an immediate check, not the next weekly run.
- Quarterly: the owner asks whether each grant is still needed and whether a narrower role would do, using the same least-privilege reasoning used at setup.
- After an incident: if drift exposed content, revoke the grant first, investigate second, and restore access only through a documented matrix change.
Keep a short log of each review with the date, the findings and the person who signed off. The log is what shows, later, that the boundary held and that drift was caught rather than assumed away. Take the free Be Human AI diagnostic to find the first task worth handing to an agent and to decide which folder it should read first.
Frequently asked questions
Can an AI agent edit files in Google Drive without access to the whole drive? Yes. Share only the folders the task needs, one for source material and one for outputs, and leave the root of the drive and unrelated shared drives out of the grant. The agent then sees only what those folders contain, as long as nobody nests them under a wider share. Check the sharing list after any reorganization to confirm that still holds.
Should the agent use my Google account or its own? Use its own dedicated identity. If the agent runs under a personal login, its actions appear as yours in the history, and its access becomes tangled with your own permissions, so revoking one can disrupt the other. A separate identity with its own folder grants is easier to audit, pause and remove.
What is the difference between Viewer and Editor for an agent? Viewer lets the agent read files without changing them, which suits reference material and source documents. Editor lets it modify and often remove files in that folder, so it belongs only on an outputs folder where generated work is expected and reviewed by a person. Commenter sits between the two when the agent only needs to leave notes for people.
How often should we review agent folder permissions? Review the sharing state weekly with an automated listing, monthly with a person who owns the decision, and immediately after any move, copy or ownership change in a granted folder. Add a quarterly question about whether each grant is still needed and whether a narrower role would work. Folders holding customer or financial records deserve a tighter loop than folders with general reference material.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


