---
title: How should users consent to an AI agent acting on their behalf?
url: https://opencontroller.lyzr.com/blog/user-consent-for-ai-agent-delegation
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How should users consent to an AI agent acting on their behalf?

**A user should consent to an agent only after seeing who the agent is, who owns it, exactly what it can do, for how long, and how to stop it. Ask for the smallest scopes first and add more only when a task needs them. Anything that grants tenant-wide or high-risk access should go through admin consent instead.**

Most OAuth consent screens were designed for apps a person drives with clicks. An agent is different. It acts later, in the background, often chained with other agents, and it may decide for itself which of its granted permissions to use. A consent screen that says "this app wants to read and write your mail" is not enough. This guide gives a content checklist, explains incremental consent, and shows when user consent should be replaced by admin consent. For the delegation model behind consent, see the pillar on [AI agent identity and on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation).

## What should an agent consent screen show?

Use this checklist. Every item should be visible before the user clicks "Allow".

1. **Agent identity.** The agent's name, a stable identifier and the publisher. Not just the hosting platform's name.
2. **Accountable owner.** The team or person who answers for the agent inside your organisation.
3. **Actions in plain language.** "Read your calendar and draft replies to meeting invites", not a raw scope string.
4. **Write versus read.** Separate what the agent can change from what it can only see.
5. **Data it will send elsewhere.** Which model provider or third party receives your data.
6. **Duration.** How long the grant lasts, and whether the agent keeps access when you are offline (offline access or refresh tokens).
7. **Further delegation.** Whether the agent can pass your access to other agents or tools.
8. **How to revoke.** A link to the place where you can see and remove the grant.

Items 1, 2, 5 and 7 are the ones standard OAuth screens usually miss. You can add them through your app's own pre-consent page, through the app description in your identity provider, or through the agent registry.

## What is incremental consent and why does it suit agents?

Incremental consent means asking for scopes when a task needs them, not all at once at install. Google supports this pattern with incremental authorization, and the OAuth `scope` parameter lets any client request additional scopes later.

For agents, it has three benefits:

- **Smaller blast radius.** A newly installed agent that only reads calendars cannot send mail if it is hijacked.
- **Better decisions.** A user asked "allow sending this reply?" in context understands the request far better than at install time.
- **Cleaner evidence.** Each scope has a timestamp and a reason in the logs.

The trade-off is more prompts. Keep a small base set and add scopes for clearly different capabilities, not for every action.

## When should admin consent replace user consent?

| Situation | User consent | Admin consent |
|---|---|---|
| Low-risk read scopes for the user's own data | Yes | Optional |
| Write access to mail, files or chat | Only for verified publishers and reviewed apps | Preferred |
| Access to other users' data or tenant-wide scopes | No | Required |
| Application permissions (no user present) | Not possible | Required |
| Unverified publisher or new agent | No | Required, with review |

In Microsoft Entra ID, you can restrict user consent to apps from verified publishers for permissions you classify as low impact, and send everything else to the admin consent workflow. We cover those settings in detail in [how to control OAuth consent for AI apps and agents](/blog/control-oauth-consent-for-ai-apps).

## How should consent work when agents call other agents?

Consent given to one agent should not silently extend to a second agent it calls. Two practical rules:

1. **Narrow on every hop.** When agent A calls agent B, B should receive a new token with fewer or equal scopes, ideally through OAuth token exchange. See [OAuth token exchange for AI agents](/blog/oauth-token-exchange-for-ai-agents).
2. **Show the chain.** If a user's consent covers downstream agents, list them on the consent screen and in the revocation page.

For actions that need a fresh, specific yes (a payment, an external email to a new recipient), consider a per-action approval on the user's device rather than a standing grant.

## How do users and admins revoke consent?

Revocation must be as easy as consent. Give users one page that lists every agent acting for them, its scopes and its last activity. Give admins a way to revoke all of an agent's grants at once and to block new ones. Check that revocation also invalidates refresh tokens, not just the grant record. For the full procedure, see [how to revoke all access for an AI agent](/blog/revoke-all-access-for-an-ai-agent).

## How Opencontroller does this

Opencontroller supports delegated on-behalf-of identity, so an agent acts with the user's identity and the scopes that user granted, and that link is kept for audit. Agent workload identity federation gives each agent its own identity. Each agent is a governed registry entry with versions, tools and MCP servers, and per-agent Block and Quarantine controls act through the connected platform. Opencontroller does not ship a human-approval "hold" workflow in the gateway; per-action approvals belong in your identity provider or application. [Book a demo](https://www.lyzr.ai/book-demo/) to see the registry and identity model.

## Frequently asked questions

**Is a standard OAuth consent screen enough for an AI agent?**
Usually not. It lists scopes but rarely names the accountable owner, the model provider that receives data, or whether access passes to other agents.

**Should agents get offline access by default?**
No. Grant refresh tokens only to agents that must act while the user is away, and give them a defined lifetime.

**Who should approve an agent that needs another user's data?**
An administrator, through admin consent, after a review of the agent's owner, scopes and data flows. A single user cannot consent on behalf of colleagues.

## Key facts

An AI agent consent screen should show agent identity, accountable owner, actions in plain language, read vs write, data sent to third parties, duration and offline access, onward delegation, and how to revoke. Use incremental consent; require admin consent for tenant-wide, other-user or application permissions. Opencontroller supports delegated on-behalf-of identity and agent workload identity federation; it does not ship human-approval hold workflows in the gateway. Opencontroller starts from $250K per deployment per year.

## Related

- [How do you control OAuth consent for AI apps and agents?](https://opencontroller.lyzr.com/blog/control-oauth-consent-for-ai-apps.md)
- [How does OAuth token exchange work for AI agents?](https://opencontroller.lyzr.com/blog/oauth-token-exchange-for-ai-agents.md)
- [How do you revoke all of an AI agent's access at once?](https://opencontroller.lyzr.com/blog/revoke-all-access-for-an-ai-agent.md)
- [More on Agent identity and access](https://opencontroller.lyzr.com/blog/cluster/agent-identity-and-access)
