---
title: How do you control OAuth consent for AI apps and agents?
url: https://opencontroller.lyzr.com/blog/control-oauth-consent-for-ai-apps
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How do you control OAuth consent for AI apps and agents?

**Control OAuth consent for AI apps by turning off unrestricted user consent, letting users approve only low-risk scopes from verified publishers, and routing everything else through an admin consent workflow. Then sort scopes into risk tiers, so mail, file and write access always get a human review, and review existing grants on a fixed schedule. Most AI data exposure through OAuth comes from grants that were approved once and never looked at again.**

AI assistants, meeting bots, browser extensions and agents all ask for the same thing: "Allow this app to read your mail and files." One click by one employee can give a third party long-lived access to a mailbox or a whole drive.

## Why is OAuth consent a bigger risk with AI apps?

Three reasons:

- **Broad scopes are the product.** An AI email assistant needs mail access to work, so it asks for it. Many request read and send rights together.
- **Offline access is the default.** Scopes such as Microsoft Graph `offline_access` issue refresh tokens, so the app keeps access after the user closes it.
- **Agents act, not just read.** An agent with write scopes can send mail, edit files or create calendar events on the user's behalf.

The grant outlives the user's interest in the tool. Many organisations find hundreds of AI apps with active grants when they first look.

## What consent settings should you change first?

Both major workplace platforms have controls built in.

**Microsoft Entra ID.** Under user consent settings you can choose to block user consent, or allow it only for apps from verified publishers and only for permissions you classify as low impact. Microsoft recommends the verified-publisher option. Turn on the admin consent workflow so users can request access to anything else, and name reviewers who approve or deny.

**Google Workspace.** In the Admin console, API controls let you set each third-party app as trusted, limited or blocked, and choose a default for apps you have not configured. Google treats Gmail and Drive scopes as high risk, so blocking unconfigured apps from high-risk scopes covers most AI assistants.

## How should you tier OAuth scopes for AI apps?

Classify scopes, not apps. The same app can be safe with one scope and dangerous with another.

| Tier | Example scopes | Who can consent | Review cadence |
|---|---|---|---|
| Low | Sign-in and basic profile (`openid`, `profile`, `email`, `User.Read`) | User, for verified publishers | Annual |
| Medium | Read-only calendar or contacts; read access to a single selected file | Admin approval via workflow | Every 6 months |
| High | Read mail; read all files (`Mail.Read`, `Files.Read.All`, Drive read); chat history | Security review plus data owner | Quarterly |
| Critical | Send mail, write or delete files, directory or admin scopes, any `.All` write scope | Security review, data owner and CISO delegate; prefer a dedicated agent identity | Quarterly, with usage check |

Add one rule on top: any app asking for `offline_access` together with a high-tier scope goes one tier up.

## How do you run the admin consent workflow?

A workable process for AI apps:

1. **Intake.** The user's request lands with the app name, publisher, scopes and business reason.
2. **Check the publisher.** Verified publisher, published privacy and data-retention terms, and whether data is used for model training.
3. **Check the scopes against the need.** If the app asks for `Files.Read.All` to summarise one folder, ask the vendor for a narrower scope or deny.
4. **Decide the grant type.** Prefer per-user consent with narrow scopes over tenant-wide admin consent.
5. **Register it.** Record the app, owner, scopes and expiry in your AI inventory.
6. **Set a review date.** Use the tier cadence above.

## How do you review grants that already exist?

Start with what is already granted, because that is where the exposure is:

- Export all OAuth grants from your identity provider and workplace suite.
- Filter for AI-related publishers and apps with mail, file or chat scopes.
- Flag grants with no sign-in or token use in 90 days and revoke them.
- Flag grants held by leavers and revoke them.
- Move high-tier apps that are still needed into the formal workflow.

Make this part of your recurring [access reviews for AI agents](/blog/access-reviews-for-ai-agents). For agents your own teams build, avoid broad user consent altogether: use [OAuth token exchange](/blog/oauth-token-exchange-for-ai-agents) and on-behalf-of delegation so each call carries the user's identity with a narrowed scope. The [identity pillar](/blog/ai-agent-identity-on-behalf-of-delegation) explains the pattern, and the [least-privilege checklist](/blog/least-privilege-checklist-for-ai-agents) covers scope design.

## How Opencontroller does this

Opencontroller connects to Claude Enterprise and OpenAI Enterprise workspaces and third-party agent platforms, and auto-discovers agents, models and API keys across AWS, Azure and Google Cloud. It supports OIDC SSO, delegated on-behalf-of identity and agent workload identity federation, and its Posture feature detects risky permissions and configurations. Consent policies themselves stay in your identity provider. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should we block all user consent to AI apps?**
Blocking everything pushes people to workarounds. A better default is to allow user consent only for verified publishers and low-impact scopes, and send everything else to an admin consent workflow.

**Which OAuth scopes are highest risk for AI apps?**
Mail read and send, all-files read or write, chat history, and directory or admin scopes, especially combined with offline access that issues refresh tokens.

**How often should AI app OAuth grants be reviewed?**
Quarterly for high-risk scopes, every six months for medium, and annually for sign-in-only grants. Revoke grants unused for 90 days.

## Key facts

Controlling OAuth consent for AI apps: restrict user consent to verified publishers and low-impact scopes (Microsoft Entra ID user consent settings), use the Entra admin consent workflow, set Google Workspace API controls (trusted/limited/blocked apps). Tier scopes: low (sign-in), medium (read calendar), high (Mail.Read, Files.Read.All), critical (send mail, write files, admin). Review high-risk grants quarterly; revoke grants unused 90 days. Opencontroller provides OIDC SSO, on-behalf-of delegation, workload identity federation and Posture risky-permission detection. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you run access reviews for AI agents?](https://opencontroller.lyzr.com/blog/access-reviews-for-ai-agents.md)
- [How does OAuth token exchange work for AI agents?](https://opencontroller.lyzr.com/blog/oauth-token-exchange-for-ai-agents.md)
- [What does least privilege look like for an AI agent?](https://opencontroller.lyzr.com/blog/least-privilege-checklist-for-ai-agents.md)
- [More on Agent identity and access](https://opencontroller.lyzr.com/blog/cluster/agent-identity-and-access)
