AI Agents for Microsoft 365: Copilot Agents, the Agent Store, and Agent 365 Licensing

An AI agent for Microsoft 365 is a piece of software that works inside Copilot, Teams, Outlook, Word, Excel or SharePoint, holds its own instructions and knowledge, and can take actions on your behalf rather than just answering a question. Where Copilot summarizes the meeting you were in, an agent watches for a certain kind of request, does the multi-step job it was built for, and reports back.

Microsoft groups them into a few shapes that behave quite differently once you start paying for them:

• Declarative agents are the simple ones. They run on Microsoft's own models with instructions, knowledge sources and a set of allowed actions. Most of what you see in the Agent Store is this.
• Custom engine agents bring their own model and orchestration. These need a Microsoft 365 Copilot licence or a pay-as-you-go arrangement, and they are what you build when the logic is too specific for a declarative agent.
• SharePoint agents are scoped to specific sites, folders or files, and are the easiest way for a non-developer to make something useful. They need pay-as-you-go enabled for SharePoint or a Copilot licence.
• Agent 365 SDK agents are agents you already built elsewhere, on Copilot Studio, Azure AI Foundry, the OpenAI Agents SDK or custom code, wrapped with Entra-backed identity and governed access so they can be treated like any other enterprise agent.

The distinction matters because the marketing talks about all four as "agents in Microsoft 365" while the licensing treats them as four separate products. Almost every surprise bill in this category comes from someone building a declarative agent, pointing it at SharePoint, and discovering that grounding in tenant data is the line that turns a free feature into a metered one.

These names are one character apart and they do opposite jobs. Getting them straight saves a lot of money.

The Agent Store is a catalog. Microsoft describes it as "the central hub in Microsoft 365 Copilot where you can easily discover and install agents", bringing together agents from Microsoft, trusted partners and your own organization. It surfaces inside Microsoft 365 Copilot and across Teams, Outlook, Word, Excel and PowerPoint. This is the supply side: it is where an agent comes from.

Agent 365 is a control plane. It went generally available on May 1, 2026, and Microsoft's own framing is that it exists to "observe, govern, and secure agents and their interactions". It gives every agent an Entra Agent ID, keeps a registry of every agent running in the organization including ones nobody registered, and applies Defender, Purview and Intune to agents the way they already apply to employees. This is the control side: it is what stops an agent from doing something you did not intend.

The practical consequence is simple. Agent 365 supplies zero agents. If you buy it expecting a catalog you have bought a security product. And the reverse: if you deploy agents from the Agent Store at any scale without a governance layer, you end up with what the Agent 365 registry was built to find, which Microsoft calls shadow AI, meaning agents running on managed endpoints that IT never approved.

One detail from the registry is worth knowing before you assume Microsoft-only is simpler. The Agent 365 registry syncs with AWS Bedrock and Google Cloud, so Microsoft itself expects your agents to come from more than one place. Its own product design does not assume a single-vendor estate, and neither should your buying plan.

No, not for every agent. Microsoft 365 Copilot Chat is included at no additional cost for every Entra account user with a Microsoft 365 or Office 365 subscription, and those users can run declarative agents from the Agent Store that are grounded in instructions and public websites. What a Copilot licence buys is grounding in your own data plus the more capable agent types.

Microsoft documents exactly three deployment options, and they are worth reading as a cost ladder:

1. Copilot Chat with no usage billing. Free with your existing subscription. Declarative agents, instructions and public web grounding only. An admin still has to enable the agents. Microsoft's own free examples here include Writing Coach, Idea Coach, Prompt Coach, Career Coach and Learning Coach.
2. Copilot Chat with usage billing. You turn on pay-as-you-go, and agents can now reach tenant data through Microsoft Graph. Microsoft is explicit that agents accessing shared tenant data such as SharePoint or Graph connector content are billed on metered consumption, and that those agents are off by default for Copilot Chat users. Enabling them needs a Copilot Studio subscription.
3. Microsoft 365 Copilot, licensed. The $30 per user per month add-on. Work-grounded chat, embedded Copilot in the Office apps, custom engine agents, Agent Builder, and access to Researcher and Analyst.

The reason Microsoft built option two is that it lets you measure before you commit. Run agents metered for a quarter, look at what each active user actually consumes, and buy seats only for the people whose credit burn clears $30 a month. For a lot of organizations the long tail of occasional users never gets close, and buying them seats is money spent on availability nobody used.

Four routes put an agent into the store, and knowing which one a listing came from tells you most of what you need to know about it.

Prebuilt agents from Microsoft and trusted partners. An admin browses the agent catalog in the Microsoft 365 admin center under Agents, then All agents, and reviews each one before enabling it. The detail pane is genuinely useful for evaluation: Overview gives publisher, channel and last updated; Data and tools lists capabilities, knowledge and actions; Security and compliance shows agent risks and protections; Certification shows the certification or publisher attestation. Assignment can be scoped to just yourself, the entire organization, or specific users and groups.

Copilot Studio agents built by your own makers. A maker builds the agent, selects the Microsoft 365 Copilot and Microsoft Teams channel, and submits it. It is not visible to anyone until an admin approves the request under Agents, All agents, Requests. Once published it appears in the "Built by your org" collection.

Agents from external platforms. Anything built on Azure AI Foundry, custom code or a third-party AI platform can be packaged with the Microsoft 365 Agents Toolkit and submitted to the organizational catalog, again subject to admin approval.

Agents wrapped with the Agent 365 SDK. For agents already built on Copilot Studio, Azure AI Foundry, the OpenAI Agents SDK or custom code, the SDK adds Entra-backed identity, governed access to Microsoft 365 data through managed MCP servers, observability, and the ability to receive and respond to notifications across Teams, Outlook and Word like a human participant. These land in the "Agents for your team" collection.

Two things follow from this. First, every route except the first passes through admin approval, so the Agent Store is not a self-service app store in the consumer sense and rollout speed depends on your IT queue. Second, agents in the store are heterogeneous: a certified partner agent and something a colleague built in Copilot Studio last week sit in the same catalog with different assurance behind them. The Certification tab is where you tell them apart.

Budget for four layers, not one. A Microsoft 365 Copilot seat is $30 per user per month on an annual commitment. Agent consumption is metered separately in Copilot Credits, which work out at roughly $0.008 per credit on the prepaid pack of 25,000 credits for $200 a month, or about a cent each pay-as-you-go. Agent 365 governance is $15 per user per month on top, or $99 per user per month if you take it as part of Microsoft 365 E7. And agents that need a real desktop run on Windows 365 for Agents at $0.40 per Cloud PC per hour in the US, plus $5 per Cloud PC per month if you want one instantly available rather than provisioned per task.

The credit rates are where estimates go wrong, because the spread between operations is wide. A classic answer is 1 credit. A generative answer is 2. An agent action is 5. Tenant graph grounding is 10. Content processing is 8 per page. Agent flows are 13 per 100 runs, which makes them dramatically cheaper per unit of work than agent actions. Voice is 10, 35 or 75 credits per minute depending on the tier, so a voice minute can cost more than fifty text conversations.

Two rules of the credit system decide whether your bill is predictable. There is no rollover, so unused prepaid capacity is gone at the end of the month. And there is a hard stop: cross 125% of your allocation and custom agents are disabled, with agent flows cut at 100%. Environments on pure pay-as-you-go are exempt from that cliff, which is the main argument for keeping some PAYG capacity alongside prepaid packs rather than running entirely on packs sized to the average month.

Put together, an organization giving 200 people licensed Copilot with Agent 365 governance is looking at roughly $6,000 a month in Copilot seats, $3,000 in Agent 365, and whatever the agents consume, before any Cloud PC runtime. That is the number worth comparing against, not the $30 headline.

Agent 365 is licensed per user, not per agent. Microsoft is explicit that a single Agent 365 licence covers all agents associated with that user, and that blueprints, agents and agent instances are not licensed separately. That is unusually generous for a product priced against a proliferating resource, and it is the strongest argument in its favor: an organization running fifty agents per employee pays the same as one running two.

The split between free and paid matters more than the price. Every Microsoft Cloud subscriber already gets the foundational layer at no additional cost: agent identity, registry visibility, basic usage insights, the core admin governance actions such as publish, deploy, block, delete and approve, plus audit logs and eDiscovery. That is enough to answer "what agents are running here and who owns them", which is the first question most IT teams have.

The paid licence adds advanced usage analytics, agent governance policies, security posture management, threat detection and data security controls. In other words you are buying detection and enforcement, not inventory. If your current problem is that nobody knows what agents exist, the free tier may already solve it.

The prerequisite is the part that catches people out. Agent 365 is not included in E3 or E5. Enterprise customers need Microsoft 365 E5, or both the Defender Suite and the Purview Suite as standalones. Frontline workers need Microsoft Defender and the Purview Suite for frontline. Small business customers need Microsoft 365 Business Premium plus the Defender and Purview suites for SMB. On an E3 estate, the honest cost of Agent 365 is the E5 uplift plus $15, which is a different conversation entirely.

One piece of good news on predictability: there are no consumption-based costs for Agent 365 itself. All interactions are covered by the per-user licence. The metered spend in this stack lives in Copilot Credits and Cloud PC runtime, not in the governance layer.

The agents that earn their keep tend to sit on a repeated, bounded process with a clear definition of done. The ones that disappoint are usually broad assistants competing with Copilot itself.

• Knowledge retrieval scoped to one place. A SharePoint agent pointed at a single policy library answers "what is our travel policy for a two-night domestic trip" without anyone opening a document. Cheap to build, easy to trust, easy to measure.
• Intake and triage in Teams. An agent sitting in a shared channel that catches requests, asks the two or three qualifying questions a human would ask, and routes the completed request to the right queue. This is where the time actually goes in most support and operations teams.
• Document assembly in Word. Agents that draft a first version from a template plus structured inputs, then hand it to a person to finish, rather than trying to produce something final.
• Meeting follow-through. Turning the decisions from a recurring meeting into tracked items in the system that owns them, which is the step that quietly fails in most teams.
• Onboarding and enablement. Microsoft's own free coaching agents, Writing Coach, Idea Coach, Prompt Coach, Career Coach and Learning Coach, cost nothing on Copilot Chat and are a reasonable way to see whether people in your organization will use an agent at all before you buy seats.

What consistently does not work is an agent whose job is defined as "help with marketing" or "assist with finance". Agents need a trigger, a bounded set of actions and a stopping condition. If you cannot write down what would make a run a failure, the agent is not ready to be built, and no amount of licensing fixes that.

If the process lives entirely inside Microsoft, a Microsoft agent is the right answer and this is not a close call. The identity is already there, the permissions are already scoped, Purview already sees the data, and if you hold Copilot seats the marginal cost of an Agent Store agent is often zero. Fighting that with an external tool is work you do not need to do.

The calculation changes when the work crosses a boundary Microsoft does not own. A process that starts in a customer email, touches a billing system, updates a CRM that is not Dynamics, and ends in a spreadsheet is not a Microsoft process with some extra steps. It is a cross-system process, and running it through Copilot means either connector work for every hop or a person moving data between them.

The second consideration is what you keep if you change your mind. An agent bought from a vendor catalog runs in that vendor runtime under that vendor identity model. The configuration work does not travel. That is a perfectly reasonable trade when the platform is one you are committed to for a decade, and a bad one when you are still deciding. Agents that connect through scoped, revocable API permissions and run on whatever model you pick cost you a reconnection to move, not a rebuild.

The third is the entry ticket. Every layer above assumes you already pay for Microsoft 365, and most of it assumes E5. If you want to put an agent on one process to see whether the category works for you at all, a neutral agent at $29 a month tells you that for a fraction of what the evaluation costs inside the Microsoft stack. Most teams that get this right end up with both: native agents for the work that already lives in Microsoft, and neutral agents for everything that spans systems no single vendor owns.