Back to the episode map

Evergreen

Workspace Studio Permissions and Third-Party Data

Learn what Workspace Studio flows can access, how Gemini uses context, what Google says about data, and when a connector sends information to another service.

Aug 4, 20266 min readBy Dalton Anderson

AI Workflow Permissions and Third-Party Data Sharing

Google says Workspace Studio operates with the initiating user's access and respects applicable Workspace controls. A flow does not become an administrator. That still leaves a large possible context because the user may have access to many messages, files, calendars, and connected services.

The most important boundary appears when a flow sends a variable or action to a third-party service. The receiving service has its own permissions, retention, data-use terms, security controls, and administrators. A Workspace-only flow and a connector-enabled flow should not share one security conclusion.

flowchart LR
    subgraph Google_Workspace["Workspace-only path"]
        A["User-accessible Gmail, Drive, Chat, Calendar"] --> B["Workspace Studio flow"]
        B --> C["Gemini step or fixed rule"]
        C --> D["Workspace action and Activity"]
    end
    B --> E["Third-party integration step"]
    E --> F["External service permissions, storage, terms, and logs"]

Start with the initiating user

Google's Workspace Studio product page says Studio can only access data that the initiating user has permission to access. It also says Studio does not override applicable Workspace data-loss-prevention controls for services such as Drive.

That is a useful boundary, but it is not the same as least privilege. A person who can read a shared folder, a large inbox, and several calendars may give a flow broad potential context. The flow designer still has to limit sources, variables, actions, and recipients.

Work and school administrators can control access to Gemini and the Workspace services used by Studio. Google's troubleshooting guide documents failures that occur when an administrator blocks a service or when a user loses access to a file, space, or application.

Permissions also change over time. A flow that worked yesterday can fail after a folder is deleted, a spreadsheet is shared, an application is disabled, or an account connection is revoked.

Gemini context can be wider than the trigger

The starter tells a flow when to run. It does not necessarily define everything an AI step may use.

Google's AI-step guidance says Ask Gemini and Ask a Gem can use web search, Workspace content the user can access, and connected integration data. By default, that can include Gmail and Chat messages, Drive files, Calendar, and linked services such as Asana, Mailchimp, or Salesforce.

For a narrow flow, select only the sources needed to perform the job. If the output will leave a trusted audience, Google advises placing the result in a draft or protected document for manual review before sharing it externally.

A prompt that says "use my Workspace" can reach far beyond the one email that triggered the run. The review question is not only whether the user has permission. It is whether the workflow needs that context for this purpose.

Google's current data commitments

Google states that customer data remains the customer's property, is not used for advertising, and is not used to train Google's general AI models outside the customer's domain. It says existing Gemini and Workspace data-protection commitments apply to Studio.

Those statements should be attributed to Google. They are not an independent evaluation of one organization's configuration, and they do not cover every outside service a flow may call.

The commitments also answer a different question from operational privacy. A vendor may promise not to train a general model on the data while a flow still sends a sensitive summary to the wrong person, saves a file in an inappropriate folder, or places unnecessary content into a log.

A connector creates a separate data boundary

Google's third-party integration documentation labels the feature a limited preview. It warns that variables added to an integration step may contain Google Account data, including Gmail or Chat content and Calendar information. The flow can share that data with the outside service.

Connecting a service may require OAuth permissions, a Workspace Marketplace listing, a helper application, and administrator approval. Some services allow granular permission selection. A missing permission can prevent the step from working.

The review has to continue past Google's connection screen. Identify what the outside service receives, which account stores it, how long it remains, whether it is used for another purpose, which subprocessors handle it, where it is hosted, who can view its logs, and how deletion works.

Google's product page says broader DLP integration for third-party Studio services is not part of its immediate plans. A Drive DLP policy therefore should not be assumed to follow data into an outside connector.

Deleting the account connection can stop dependent flows, but it does not prove that information already sent to the service was deleted. That question belongs to the receiving service and the relevant agreement.

Flow sharing can reveal configuration

Workspace Studio can share a flow by allowing another person in the organization to make a copy. Google's sharing guide says the copy can include the current setup, files, text, and email addresses. It also shows the creator's email address.

Sharing the flow does not grant access to underlying Workspace content. A recipient only sees content they already have permission to access. Even so, the definition may contain sensitive constants, recipients, folder references, prompts, or business logic.

Unsharing prevents new copies. It does not remove copies that people already made. Treat the flow definition as an artifact that may need its own review before distribution.

Trace one data element from beginning to end

The simplest useful review follows one representative data element.

StageQuestion
StarterWhich event introduces the data, and can unrelated events match?
SourceWho owns it, how sensitive is it, and may this purpose use it?
ContextWhich other messages, files, services, or web sources can the decision use?
DecisionIs the rule fixed, model-based, or approved by a person?
ActionWhat changes, who receives it, and can the action be reversed?
ConnectorDoes the data leave Workspace, and under whose terms?
EvidenceWhere do inputs, outputs, errors, and approvals appear?
RetentionWhich systems keep the data, and who can delete it?

Repeat the trace for the most sensitive input and the most consequential action. If the path cannot be described, the permission review is incomplete.

The E097 mistake was a scope warning

During the recorded inbox test, another active flow surfaced a phone number and email address while the screen was being shared. Invoice context also appeared in Activity. The attachment save itself worked, but the full system included more than the one flow being demonstrated.

That is why the public episode article uses a diagram instead of a screenshot from the recording. Redacting one visible value does not establish that the rest of the frame is safe.

A safer test uses synthetic messages, a private destination, self-only recipients, and only the necessary flows. Google's test-run guidance confirms that tests use real data and take real actions.

Use How to Evaluate an AI Workflow Before You Turn It On to record the complete boundary. For the current product model, read What Is Google Workspace Studio?.

E120's [[How to Manage an AI Coworker|AI coworker guide]] addresses the same problem at a broader level of delegation. Permission to read context and permission to make a consequential decision should remain separate, no matter what the product calls the system.

This explainer was developed from current Google documentation and the preserved E097 transcript. AI assistance was used for research organization, drafting, and validation. It does not certify a flow, organization, connector, or data-processing arrangement.

Sources

Follow the evidence.

  1. docs.cloud.google.com: choose design pattern agentic ai systemdocs.cloud.google.com
  2. support.google.com: 16765942support.google.com
  3. NIST AI RMF Measure guidanceairc.nist.gov
  4. support.google.com: 16447677support.google.com
  5. support.google.com: 16431116support.google.com
  6. support.google.com: 16658279support.google.com
  7. servicenow.com: how now platform worksservicenow.com
  8. support.google.com: 16663517support.google.com
  9. support.google.com: 16275487support.google.com
  10. support.google.com: 17176961support.google.com
  11. support.google.com: 16430806support.google.com
  12. support.google.com: 16444479support.google.com
  13. NIST: Artificial Intelligence Risk Management Framework, Generative Artificial Intelligence Profilenist.gov
  14. support.google.com: 16431105support.google.com
  15. zapier.com: developer platformzapier.com
  16. workspace.google.com: studioworkspace.google.com
Workspace Studio Permissions and Third-Party Data