Can my AI app connect to users' Gmail, Google Drive, or Slack?
The platform rules that decide whether an AI integration ships, including Google's Limited Use and security assessment and Slack's API terms, plus the privacy and security duties that come with users' accounts.
Short answer
Yes, but the platforms set the rules more than the law does, and their rules are strict about AI. Google bars using Workspace data to train AI models beyond a single user's personalized model, and apps that read Gmail or Drive broadly must pass an annual third-party security assessment. Slack bars third-party apps from training large language models on Slack data or bulk-exporting it. Get these wrong and the integration is rejected or shut off. On top of that come the usual privacy and security duties for holding other people's messages and files.
Google: Gmail, Drive, and Calendar
Google's API Services User Data Policy, and its separate policy for Google Workspace APIs, apply to any app that uses Google sign-in to reach users' data:
- Verification and scopes. Broad access to Gmail and Drive falls under Google's sensitive or restricted scopes. Apps need Google's verification, and apps using restricted scopes must pass an annual security assessment by a Google-designated third party (the Cloud Application Security Assessment, or CASA). Ask for the narrowest scopes that work, since they decide how much review you face.
- Limited Use. You may use the data only to provide or improve user-facing features the user can see. Transfers are allowed only in narrow cases, and selling the data or using it for advertising is prohibited.
- No training general models. The Workspace policy prohibits using user data to create, train, or improve an AI or machine learning model, except that specific user's personalized model for the approved feature. Fine-tuning a shared model on users' email is out.
- No human reading. Your staff can't read users' data except in narrow cases, such as with the user's consent for a specific message, for security, or to comply with law.
- Prompt injection. The Workspace policy calls for protecting against prompt injection, for example with Google Cloud's Model Armor. An agent reading email is reading text written by strangers, and some of it will be written to manipulate the agent.
- Disclosure in the product. Telling users how you use their Google data can't live only in your privacy policy or terms. It has to appear in the product too.
Slack
Slack's API Terms of Service set out rules for apps used outside the organization that built them:
- No LLM training. You may not use data from Slack's APIs to train a large language model.
- No bulk export. You may not bulk-export Slack message and file data unless an additional agreement allows it.
- Minimum data. Keep use, processing, and retention of other organizations' data to the minimum necessary, with explicit authorization from the organization that installs your app.
- No indexes from search APIs. Data from Slack's Data Access and Real-Time Search APIs can't be stored as persistent copies, archives, or indexes, or used for product improvements.
- Rate limits. Since 2025, apps that aren't approved for the Slack Marketplace get much lower limits on reading message history. An AI product that reads conversations at scale effectively needs Marketplace approval.
Microsoft 365
Microsoft 365 data comes through Microsoft Graph. Broad permissions usually need an organization administrator's consent, and Microsoft's API terms apply. Expect enterprise customers' security teams to review exactly which permissions you request.
The legal duties that come with it
- Privacy policy and purpose. Your privacy policy has to describe the integration, the data you take, and what you do with it. State privacy laws limit using data beyond the purpose you described, and email and messages are full of other people's information, not just your user's.
- Model providers. Sending email or messages to an LLM provider brings the provider's retention and training settings into the picture. See Can I send user data to an LLM provider?
- Tokens are credentials. OAuth tokens can unlock a user's whole mailbox. Store them encrypted, limit who and what can use them, and plan for revoking them. A leak can trigger breach-notification duties. See How should my agent handle users' passwords and logins?
- Agents acting in the account. If your agent sends email, posts messages, or edits files, that's a new kind of action. It needs the user's clear authorization and a confirmation step for anything consequential.
Checklist an agent can run against the codebase
- Find every OAuth integration (Google, Microsoft, Slack, and others) and list the scopes or permissions each requests.
- Check whether any Google scope is sensitive or restricted, and whether verification and the annual security assessment are in place.
- Find where integration data goes: prompts, embeddings, logs, analytics, or training and evaluation sets. Flag any path that could train a shared model.
- Check that Slack data isn't bulk-exported or stored as a long-term index, and that retention is limited to what features need.
- Check how OAuth tokens are stored, encrypted, scoped, and revoked when a user disconnects.
- Check what happens to stored data and derived data, such as embeddings and summaries, when a user disconnects or deletes their account.
- Check for prompt-injection defenses where the agent reads email, documents, or messages.
- Check that the product, not only the privacy policy, tells users what the integration accesses and why.
Sources
- Google API Services User Data Policy
- Google Workspace API User Data and Developer Policy
- Slack API Terms of Service
- Slack: rate limit changes for non-Marketplace apps (May 2025)
If this applies to you
An agent product review covers an integration end to end: platform rules and scopes, data flows to model providers, token security, the privacy policy, and the terms that allocate responsibility. For one change, such as adding a Gmail or Slack connection, a pre-build feature check covers that change before it ships.