A Teams AI agent is an AI assistant that works inside Microsoft Teams to summarise meetings, capture actions and answer questions from your organisation’s own data. Built-in agents like Copilot in Teams, Facilitator and Channel Agent turn on with a licence. Custom agents, built in Copilot Studio, are grounded in your policies and systems. This guide covers what they do, what they cost, how to secure them and where they actually pay off.
Adoption is no longer the real question. The harder question is which agents to turn on, what to ground them in and how to keep them governed.
Table of Contents
What is a Teams AI agent?
A Teams AI agent is an AI assistant that operates inside Microsoft Teams and understands context across your chats, meetings, files and Microsoft 365 data. Unlike a scripted bot, it interprets natural language, references earlier discussions and grounds its answers in real organisational information rather than generic web content.
Microsoft groups the built-in options into a few clear types. Copilot in Teams is the personal assistant that summarises chats and meetings, drafts content and answers open questions. Facilitator acts like another participant in meetings, capturing notes and unanswered questions. Channel Agent manages a channel, flagging buried deadlines, writing status reports, assigning tasks and answering questions in natural language. Alongside these sit custom agents built in Copilot Studio and grounded in your own knowledge.
In our work the confusion is almost always between these built-in Copilot features and a purpose-built agent. They are not the same decision and treating them as one is where a lot of wasted effort starts.
Teams AI agent vs Copilot: what is the difference?
Copilot is the assistant. An agent is a scoped worker. Copilot in Teams helps any user across general tasks, while a Teams AI agent is tuned to a specific job and knowledge set, such as answering policy questions or capturing channel actions. Most organisations use both, starting with Copilot and adding agents where a repeatable task justifies one.
The distinction matters for grounding. According to Microsoft, without a Microsoft 365 Copilot licence the Copilot Chat experience is grounded only in public web data, so it cannot see your files at all. Add the licence and Copilot is grounded in your work data. Build a custom agent on top and you decide exactly which policies, sites and systems it reads from. That is the difference between a helpful generalist and an agent that answers like someone who knows your business.
If your priority is a broad Copilot rollout rather than a single agent, our Microsoft Copilot integration work covers adoption and governance and our guide to Copilot Studio agents goes deeper on building custom ones.
How does a Teams AI agent work?
A Teams AI agent works by combining a large language model with Microsoft Graph, which gives it secure access to the meetings, chats, files and people the user can already see. It interprets the request in natural language, retrieves the relevant context, reasons over it and returns an answer grounded in your data.
Microsoft Graph and your organisational context
Microsoft Graph is the layer that lets an agent understand your tenant. Through it the agent can reach recent meetings and attendees, chat and channel conversations, documents in SharePoint and OneDrive, tasks and calendars, plus the roles and relationships in your organisation. This grounding is what separates a useful answer from a confident guess.
Reasoning over context, not just generating text
What makes an agent useful is that it reasons across context rather than only producing fluent text. Ask it what the key risks raised last week were and a capable agent identifies the right meeting, reviews the conversation and summarises the risks that were actually discussed. That is the step from novelty to genuine assistance.
Permissions are inherited, not bypassed
A Teams AI agent inherits your Microsoft 365 security model. Identity runs through Microsoft Entra and the agent can only see what the requesting user is already authorised to see. It does not create a side door into restricted content, which is the first thing most security teams want to confirm.
Built-in vs custom Teams AI agents
Built-in Teams AI agents turn on with a Microsoft 365 Copilot licence and deliver fast, general productivity gains. Custom agents, built in Copilot Studio, are grounded in your own policies and systems and can trigger workflows. Most organisations start built-in and add custom agents only where a specific, repeatable task justifies the effort.
The table below sets out how the two compare across the decisions that usually matter.
| Area | Built-in Teams AI agents | Custom Teams AI agents |
|---|---|---|
| Setup | Enabled with a Microsoft 365 Copilot licence and minimal configuration | Designed and built in Copilot Studio to match a specific process |
| Time to value | Immediate, general productivity gains | Higher upfront effort, longer term operational impact |
| Grounding | General Microsoft 365 context | Your policies, sites and line-of-business data |
| Language | Generic understanding | Your internal terminology and rules |
| Integration | Microsoft 365 services | CRM, ERP, finance and other systems via connectors |
| Automation | Basic follow-up suggestions | Can trigger Power Automate workflows |
| Best suited for | Fast adoption and broad productivity | Specific, repeatable tasks with real payoff |
Our usual advice is to start with built-in agents to build confidence, then commission a custom agent only where the payoff is clear. Jumping straight to a bespoke build before anyone trusts the basics is how agents end up unused.
What does a Teams AI agent cost to run?
A Teams AI agent’s cost depends on the licensing path, not on a single sticker price. Built-in agents and Copilot in Teams require a Microsoft 365 Copilot licence, charged per user each month. Custom agents built in Copilot Studio are included for licensed users inside Teams, or billed on consumption through Copilot Credits where no licence applies.
Microsoft describes the Microsoft 365 Copilot licence as the foundation for the built-in tool set, including Copilot in Teams, Facilitator and Channel Agent. Microsoft’s licensing documentation then confirms that for licensed users, running Copilot Studio agents in Teams for generative answers and Microsoft Graph grounding is zero-rated, meaning it does not draw down a separate pack. Where there is no Copilot licence, custom agents bill on consumption through Copilot Credits, available pay-as-you-go through an Azure subscription with no upfront commitment.
We keep the dollar figures out of a public post because Microsoft revises them regularly and the model matters more than the number anyway. In practice the licence is rarely the real cost. The work is the groundwork, which is the next section.
What actually makes a Teams AI agent work
The thing that makes a Teams AI agent useful is not the model, it is what the agent is grounded in. In our experience the agents that disappoint are reading the wrong information or have no clean knowledge base behind them. The ones that succeed sit on well-structured, permissioned content the agent can actually retrieve.
We saw this clearly with an Australian healthcare provider. We built a custom Copilot agent grounded in their policies and SharePoint content. Staff who used to spend a long time filing through documents to answer a question now get a grounded answer in seconds. The win came from organising and permissioning the source content first. The model was the easy part.
The pattern repeats on the failure side too. Another client came to us with an agent that kept returning poor answers and had lost the team’s trust. The model was fine. It was reading the wrong information and had no proper knowledge base behind it. We fixed the foundation, structured the content, set the permissions and scoped what the agent was allowed to see, and the answers improved.
The data backs the pattern. IBM notes that poor data quality is one of the most common reasons AI initiatives fail and reports that 68% of AI-first organisations have mature, well-established data and governance frameworks compared with just 32% of other organisations. Our own view is simpler. Treat the knowledge base and permissions as the project and the agent as the front end. If you want the longer version of why this is where projects stall, we wrote about why most SMB AI pilots fail and the fix usually starts with the AI agents groundwork rather than the agent itself.
Is a Teams AI agent secure for an Australian organisation?
A Teams AI agent is as secure as the Microsoft 365 tenant it runs in. It inherits Microsoft Entra identity and your existing permissions, so it only sees what the user can already see. Microsoft also commits to Australian data residency for Microsoft 365 Copilot and content of interactions stays within your service boundary.
Microsoft states that prompts, files and responses never leave your organisation’s Microsoft 365 service boundary and that Microsoft 365 Copilot is ISO/IEC 42001:2023 certified for responsible AI. On location, Microsoft’s data residency commitments list Australia among the countries where the content of interactions is stored at rest in the local region geography. That covers the questions most Australian boards ask first.
The part that is on you, not Microsoft, is the Australian Privacy Principles and your own permission hygiene. Most of the governance work we do before pointing an agent at SharePoint is making sure the permissions underneath are actually right, because an agent will faithfully surface whatever a user already has access to.
How to roll out a Teams AI agent
Rolling out a Teams AI agent works best as a staged process where each step earns the right to the next. Understand how Teams is used, enable built-in agents with governance, fix the knowledge base, then build custom agents where they pay off. That sequence is how you avoid another underused feature.
- Map how Teams is actually used. Find where meetings, decisions and follow-up work really happen, not how the process is meant to run.
- Enable built-in agents with governance sorted. Get licensing, security and a little user education in place so Copilot in Teams, Facilitator and Channel Agent land with clarity.
- Fix the knowledge base and permissions first. Structure and permission the content before you ground anything on it, as the healthcare example above shows.
- Build custom agents where a repeatable task justifies it. Use Copilot Studio to design agents tied to specific roles, systems and rules.
- Measure, refine and scale. Track usage, refine prompts, expand what works and keep governance current as adoption grows.
If you want a partner for that sequence, our AI and intelligent solutions team does exactly this, and the broader picture sits in our guide to AI for Australian business.
Teams AI agents in the Australian business context
Australian adoption is uneven but moving in one direction. The National AI Centre reports 43% of SMEs using AI in early 2026, with health among the leading sectors, while agentic AI remains largely untapped. For distributed teams across metro and regional Australia, a Teams AI agent is a practical first step inside a tool people already use.
The same research found that 19% of SMEs simply do not know where to start and that more than half of non-adopters cannot see how AI is relevant to a business like theirs. That matches what we see. The blocker is rarely capability, it is a clear first use case. A meeting recap agent or a policy question agent is concrete enough to get a team over the line, which is why we usually start there.
Frequently asked questions
Do I need a Microsoft 365 Copilot licence for a Teams AI agent?
For the built-in agents, yes. Microsoft describes the Microsoft 365 Copilot licence as the foundation for Copilot in Teams, Facilitator and Channel Agent. Custom agents built in Copilot Studio can run on consumption-based Copilot Credits without a per-user Copilot licence, billed pay-as-you-go through an Azure subscription, so the right path depends on how many people need the agent.
What is the difference between Copilot and a Teams AI agent?
Copilot is the general assistant that helps any user across everyday tasks. A Teams AI agent is scoped to a specific job and knowledge set, such as answering policy questions or managing a channel. Most organisations run both, using Copilot broadly and adding agents where a repeatable task justifies a purpose-built one.
Can a Teams AI agent see data I am not allowed to see?
No. A Teams AI agent inherits your Microsoft 365 permissions through Microsoft Entra and can only surface content the requesting user is already authorised to access. It does not bypass existing permission boundaries, so the security question is really whether your underlying SharePoint and file permissions are correct.
Where is the data stored for Australian organisations?
Microsoft lists Australia among the countries with data residency commitments for Microsoft 365 Copilot, with the content of interactions stored at rest in the local region geography. Microsoft also states that prompts, files and responses stay within your organisation’s Microsoft 365 service boundary.
Why do Teams AI agents give poor answers?
Almost always grounding, not the model. In our experience poor answers come from an agent reading the wrong information or having no clean knowledge base behind it. Fixing the source content and permissions is what turns a frustrating agent into a reliable one, which is consistent with research linking AI failure to data quality.
Should we start with built-in or custom agents?
Start built-in. The built-in agents deliver fast, general productivity gains and build confidence with minimal setup. Move to a custom Copilot Studio agent once a specific, repeatable task has a clear payoff, your knowledge base is in order and people trust the basics.
How long does it take to build a custom Teams AI agent?
The agent itself is quick to stand up. The real timeline is the groundwork, organising and permissioning the content the agent will read from. A clean knowledge base can mean a working agent in weeks, while a messy one is the thing that stretches a project out, so the content is where to focus.
Final thoughts
A Teams AI agent can change how people find information, capture decisions and follow through, whether through built-in capabilities or a custom build. The successful ones are not the ones with the cleverest model. They are the ones grounded in well-organised, properly permissioned content, rolled out in a sequence people can trust.
If you want to work out where a Teams AI agent would pay off in your organisation, we offer a free initial conversation to map the use cases and the groundwork involved.
