Declarative agent
Custom instructions, knowledge and actions on top of Copilot’s orchestrator and foundation models.
- No extra hosting
- Inherits Microsoft 365 compliance
- Designed for individual use
- Responds when a person asks
Guide · Conversational AI
AI agents in Microsoft Teams: where they run, what they hear, who allows them.
An agent in Teams is an app with a conversation. Microsoft carries the message; your choice of agent decides what answers it — and the Teams admin decides who may use it.
People already spend the day in Teams, which makes it the obvious place to put an agent. Getting it right means knowing three things: how a message travels from a chat to an agent and back, which kind of agent to build, and which controls an administrator holds. This guide covers all three from Microsoft’s own documentation, then the pitfalls teams hit in production.
Updated · 8 sources
How a Teams message reaches an agent and back
01
A person in Teamsa chat, or an @mention in a channel
chat02
Teamschecks the agent is allowed, sends it on
policy03
Messaging endpointone address per agent; JSON in
activity04
The agent’s engineCopilot’s, or one you host
engine05
Knowledge and actionswhat it reads, what it may change
tools06
The replytext or a card, in the same thread
replyHow do AI agents in Microsoft Teams work?
AI agents in Microsoft Teams are apps that hold conversations. A person messages the agent in a one-to-one chat or @mentions it in a group chat or channel; Teams sends the message to the agent’s endpoint; the agent’s engine — Copilot’s, or one you host — reasons over its knowledge and tools and replies in the thread. Administrators decide in the Teams admin center which agents are allowed and for whom.
Declarative or custom engine
Declarative or custom engine: the two ways to build AI agents in Microsoft Teams.
Microsoft separates agents by who runs the brain. A declarative agent configures Copilot with instructions, knowledge and actions. A custom engine agent brings its own orchestration and models. Copilot Studio can build either style with little code.
Custom engine agent
Your own orchestration and choice of models, reached from Teams and Microsoft 365 Copilot.
- Hosted by you, or in Copilot Studio
- You ensure compliance and security
- Works in channels and meetings for groups
- Can message people proactively
MicrosoftWho chooses the modelYou
Within Copilot’s orchestrationMulti-step business logicWhatever your orchestration does
Individual useGroup collaborationIndividuals and groups
NoStarts a conversation itselfYes, with proactive messages
Microsoft 365 appsReach outside Microsoft 365Microsoft 365 and external apps
Start here for focused scenariosMicrosoft’s adviceFor complex workflows and your own models
What an agent hears
Personal chats, group chats and channels: what an agent hears in each.
The same agent behaves differently depending on where it is installed. The biggest surprise for new builders: in group chats and channels, an agent hears only the messages that @mention it.
one to one
Personal chat
Every message the person sends reaches the agent. The natural place for questions about one person’s own work.
- Hears everything in the chat
- Private to the person
several people
Group chat
Installed for the chat; members address it with an @mention. Replies are seen by everyone in the chat.
- Hears @mentions by default
- Shared answers
a team
Channel
Installed for a team; it can see the team’s details and channel list. Answers in the thread it was mentioned in.
- Hears @mentions by default
- Replies in the thread
Details from Microsoft’s documentation
- Hearing everythingWith resource-specific consent declared by the agent and granted by the team or chat owner at install, an agent receives all messages, not only @mentions.
- Other botsTeams does not deliver messages written by other bots or agents to an agent.
- Private channelsAgent apps can be added, but agents can’t post messages or cards in private channel conversations.
- MentionsRead who was mentioned from the message’s entities, not its text — the text can be altered by the sender.
- Welcome messagesPost one in the conversation where the agent was added; never message every member of a large team.
Step by step
What happens when someone @mentions an agent in a channel.
A few seconds pass between the @mention and the reply. In that time the message crosses from Teams to the agent’s own service and back, and the agent decides what to do with it.
- @mention
A person types “@Supplier Desk which invoices are unmatched?” in a thread.
- Delivered
Teams sends a JSON activity to the agent’s single messaging endpoint.
- Cleaned
The agent checks the mention is its own and strips it, leaving the question.
- Reasoned
The engine reads its knowledge, calls its tools, and composes an answer.
- Replied
The answer lands in the same thread — as text or as an Adaptive Card.
In the thread
You
@Supplier Desk which invoices are still unmatched this quarter?
Agent
42 invoices are unmatched for Q3; the largest is $48,920 from Bayside Freight, whose purchase order was closed early. Sources: Invoices table · PO-5521
For administrators
Governing AI agents in Microsoft Teams: who can install, who can use, what they see.
An agent in Teams reads conversations and answers on behalf of the organisation, so it is governed like any app — and then some. These are the controls Microsoft gives a Teams administrator, plus the questions to put to whoever built the agent.
An agent governance checklist for Teams
- Allowed or blockedEach agent is allowed or blocked in the Teams admin center; org-wide settings control third-party and custom apps in bulk.
- Assigned to peopleApp centric management makes an agent available to everyone, or to chosen users and groups; guests can’t use it even if assigned.
- Approved before it is listedAgents built in the organisation are submitted for approval before they appear under “Built for your org”.
- Installed and pinnedSetup policies install and pin an agent for the people who need it.
- Consent understoodResource-specific consent lets an agent read every message in a chat or channel — owners grant it; know which agents ask.
- Data path knownMicrosoft notes that agent content and chat data may flow outside the organisation’s compliance and geographic boundaries — check where.
- Usage watchedThe Teams app usage report shows which agents are used and by how many people.
Decision guide
Choosing how to bring an AI agent into Microsoft Teams.
Most of the decision is where the agent’s knowledge and actions live. If they are in Microsoft 365, Copilot’s own engine goes far. If they are in other systems, or the agent needs your own models and rules, bring your own engine.
Where do the agent’s knowledge and actions live?
In Microsoft 365A declarative agent: instructions plus SharePoint, OneDrive and Teams content, on Copilot’s engine and compliance.
In other business systemsA custom engine agent, or a platform that already reaches those systems, connected to Teams as an app.
Both, for whole teamsA custom engine agent installed in channels, with its own permissions model and a record of every answer.
pitfall
Shared chats change what an agent may read
Microsoft documents that Copilot Studio agents in group chats and channels can’t use knowledge that needs each person’s sign-in, such as SharePoint — that works only in one-to-one chats.
pitfall
Installed is not the same as allowed
A person can install an agent from a link and still be unable to chat with it if the agent is not shared with them.
pitfall
Teams rate-limits agents
Keep replies concise and avoid floods of messages; Microsoft throttles agents that send too many.
pitfall
Updates need re-approval
Changing an approved agent’s icon or descriptions means submitting it for approval again; content updates do not.
On Botminds
How Botminds puts AI agents in Microsoft Teams.
With Agent Reach, a Botminds agent answers in Microsoft Teams as itself — the same agent, knowledge, tools and record as in the app, for people who will never open another tool.
Agent Reach for Teamsone Teams app serves every agent in the organisation
Installed onceby a Microsoft 365 admin, for the whole organisation
@mention or one-to-onemention an agent in a channel, or pick it in a chat
Promoted by an adminan admin chooses which agents reach Teams, email or both
Other organisations refusedmessages from outside your tenant get no answer
Paused in one switchan operator pauses a promoted agent in one move
On the recordevery Teams turn kept with the run, under the same spend quotas
You choose what it holdsan agent in Teams searches the sources it was given
Answers as itselfnever as a person, never from a person’s inbox
Read how the same agent reaches every screen, inbox and Teams chat in conversational AI on Botminds.
Questions people ask
AI agents in Microsoft Teams: common questions.
The technology behind itConversational AIAgents in every app, inbox and channel→Can you add AI agents in Microsoft Teams?
Yes. AI agents in Microsoft Teams are installed as apps: people chat with them one to one, or add them to a group chat or a team and @mention them. An administrator must allow the agent in the Teams admin center first.
How do I create an AI agent in Microsoft Teams?
Three common routes: build a declarative agent with Microsoft 365 Copilot or the Microsoft 365 Agents Toolkit; build in Copilot Studio and connect the Teams and Microsoft Copilot channel; or build a custom engine agent with the Teams SDK and host it yourself. Then publish it and have an admin approve it.
How do I add an AI agent to a Teams channel?
The agent must support the team scope. Install it to the team from the Teams app store, then @mention it in any channel of that team. In Copilot Studio, the maker must also allow users to add the agent to a team.
Can an AI agent read all messages in a Teams channel?
Only with permission. By default an agent in a channel or group chat receives only messages that @mention it. With resource-specific consent granted by the conversation’s owner, it can receive every message in that conversation.
What is the difference between a declarative agent and a custom engine agent?
A declarative agent runs on Copilot’s orchestrator and models with your instructions, knowledge and actions, and needs no hosting. A custom engine agent brings its own orchestration and models, is hosted by you or in Copilot Studio, and can work for groups and message people proactively.
What is the Facilitator agent in Teams?
Facilitator is one of Microsoft’s built-in Teams agents, alongside the Channel Agent. Built-in agents are managed on their own page in the Teams admin center, where an admin sets who can use each one.
Who controls which AI agents people can use in Teams?
Teams administrators. They allow or block each agent, assign it to everyone or to chosen users and groups, approve agents built inside the organisation, and can install and pin agents for people with setup policies.
Can an AI agent in Teams show citations?
Yes. Microsoft documents three labels an agent can put on its messages: an “AI generated” label, numbered citations that open the source, and a sensitivity label. All three work in personal chats, group chats and channels.
Sources
- Microsoft Learn: Compare declarative and custom engine agentslearn.microsoft.com
- Microsoft Learn: Channel and group conversation chat for agentslearn.microsoft.com
- Microsoft Learn: Configure an agent to receive all messages in group chats and channelslearn.microsoft.com
- Microsoft Learn: Conversations with an agentlearn.microsoft.com
- Microsoft Learn: AI content labelslearn.microsoft.com
- Microsoft Learn: Manage your apps in the Microsoft Teams admin centerlearn.microsoft.com
- Microsoft Learn: App centric managementlearn.microsoft.com
- Microsoft Learn: Connect and configure an agent for Teams and Microsoft Copilot (Copilot Studio)learn.microsoft.com
Read next
Conversational AI vs agentic AISystems that talk, systems that act, and the ones that do both→Conversational AI vs chatbotScripted replies, matched intents, and models that understand free text→Generative UI vs composable UIInterfaces a model writes at run time, and interfaces assembled from parts→AI governance frameworkRoles, inventory, risk tiers, controls and monitoring — mapped to NIST and ISO→All guidesEvery explainer, by topic→
See it on
your own work.
Bring one process and its documents. We’ll show you the agents, the controls and the record, running on your data.