Its own bot, scoped permissions
How an agent joins Microsoft Teams without access to your whole organization
To post in Teams, the agent needs an identity of its own: a bot in Azure AI Bot Service with a Microsoft Entra ID identity, published as an app your admin approves and assigns to everyone, to specific groups or to no one.
Permissions are requested through resource-specific consent (RSC), which applies to one team or one chat. By default the bot only receives messages where it's @mentioned; reading a whole channel requires a separate RSC permission that the team owner accepts at install time, and we request it only when the use case needs it.
Alerts are proactive messages, which Teams allows only where the app is installed, within limits of seven messages per second per conversation and 50 requests per second per app in each tenant (Microsoft Learn). On a 429 the agent retries with an increasing wait, and if someone blocks the bot, Teams returns a 403 and the agent stops messaging that person.
A bot is a non-human identity, which is where governance tends to lag: 78% of organizations cite controlling non-human identity access and permissions as a top concern, and only 10% have a strategy for governing them (Okta). Admin approval, per-team consent and a log of every action keep the agent inside your governance strategy.
Microsoft also publishes official MCP servers, including one with Teams tools hosted in Agent 365, which Copilot Studio documents as a preview kept for backward compatibility. For new projects Microsoft points to Work IQ, whose access is delegated: every action runs on behalf of a person and sees what that person can see.
MCP and company-wide agents: a server with delegated permissions is built for each employee's assistant. An agent that posts alerts on the company's behalf works with its own identity and scoped permissions, so we evaluate MCP case by case.