Copilot Agents vs Custom AI Agents: Which Should You Build?

Build a Copilot agent when the work lives inside Microsoft 365: answering questions over SharePoint and Outlook, drafting in Word and Teams, and light workflows Copilot Studio's connectors already cover, used by employees with Microsoft 365 Copilot licenses or Copilot Chat. Build a custom agent on a model platform such as Anthropic Claude when the job crosses systems Microsoft does not own, needs approval logic you define, runs unattended on a schedule or an event, serves customers rather than staff, or would otherwise need a Copilot license for every person who touches it. Copilot agents sit at the desk; custom agents run the back office.

What is the difference between a Copilot agent and a custom AI agent?

A Copilot agent is an assistant Microsoft hosts, running inside Microsoft 365 Copilot and Teams. Agent Builder lets a non-technical user describe an agent and point it at SharePoint sites, files and Copilot connectors; Microsoft's Agent Builder documentation notes that these agents are included in the Microsoft 365 Copilot license and that anything needing actions against external services belongs in Copilot Studio. Copilot Studio is the fuller product: knowledge sources, connectors, agent flows, publishing to more channels, and its own consumption billing.

A custom agent is software you own. It runs where you put it, calls a model through an API, and uses tools you define to read and write your systems. On the Claude platform that can mean the API with your own tool loop, the Agent SDK, which according to Anthropic's overview runs the loop in your own process with built-in tools, hooks, permissions and Model Context Protocol support, or a hosted option where Anthropic runs the loop. The agent's memory, tools, guardrails and logs are yours to design: more control, more responsibility.

How does each one get access to your data?

Grounding is the polite word for what the agent is allowed to read. Copilot agents ground on the Microsoft Graph: SharePoint, OneDrive, Outlook, Teams and whatever you have indexed through Copilot connectors. The strength is that the index already exists and already respects who can see what. If you want line-of-business data in that index, Microsoft Graph connectors are the route.

Custom agents ground on whatever you connect: Microsoft 365 through the Graph API, the accounting system, the vendor portal with no API, and the ten-year-old SQL Server that runs the business. The Model Context Protocol, an open standard for connecting AI applications to external systems, gives you a reusable way to expose each of those as tools, and Anthropic's platform supports MCP servers directly. The strength is reach; the cost is that you build and maintain each connection and decide what the agent may retrieve.

Who controls what the agent is allowed to do?

A Copilot agent acts as the signed-in user, so it cannot read a document its user cannot read, and administrators approve, block or scope agents from the Microsoft 365 and Power Platform admin centers. It also means the agent inherits every over-share in your tenant; a Copilot rollout is when most companies discover that "Everyone" has had access to the HR drive for years. Cleaning that up first is the opening job in any Copilot engagement, and our Microsoft 365 Copilot FAQ covers the readiness questions.

A custom agent has no permissions until you give it some. It runs under a service identity you create, with the narrowest access that does the job: read invoices and post to one ledger. Every tool it can call is on a list you wrote, and the Agent SDK exposes permission settings and hooks that decide which tools run automatically and which stop for a human. Nothing is inherited, so nothing is over-shared by accident, but every capability must be justified and logged. For unattended agents the explicit model is safer; for a chat assistant used by a hundred employees, the inherited model is far less work.

Where do approval steps fit?

Every agent that can spend money, change a record or message a customer needs a point where a person says yes. In the Copilot world that point is usually an agent flow with an approval action: an agent gathers a purchase request, routes it to a manager in Teams and continues only once it is approved. The approval lives in Microsoft's tooling, is visible to admins and needs no code.

In a custom agent the approval is wherever you want it. A common pattern is a hold queue: the agent drafts the action (a vendor payment or a customer email), writes it to a queue with its reasoning, and a person approves it in Teams before the tool executes. The obligation is to test the gates as carefully as the happy path, because an approval step that can be talked around by a cleverly worded email is not an approval step. Our AI agents FAQ covers how we think about guardrails.

What does each cost, in shape rather than numbers?

Copilot agents are priced around people. A Microsoft 365 Copilot license is per user, and Microsoft's cost considerations page states that licensed users incur no extra charges for using agents, connectors or plugins. Users without it can reach agents through Copilot Chat, included with eligible Microsoft 365 subscriptions, but agents that touch shared tenant data are then metered in Copilot Credits. Credits come as prepaid packs pooled across the tenant or as pay-as-you-go on an Azure subscription, and Microsoft's billing rates page shows features consuming credits at different rates. Cost rises with headcount if you license everyone, and with usage if you do not.

Custom agents are priced around work. The platform bills by tokens processed, in and out, per Anthropic's pricing documentation, which also describes prompt caching and batch processing as ways to lower per-task cost, and the same usage can be bought through cloud marketplaces such as Microsoft Foundry. Add a small hosting bill and the real line item: building, testing and maintaining the agent as your systems change. No per-seat license is involved, so an agent that processes every invoice costs the same whether four people or forty read the results. Our guide to AI automation ROI explains how we estimate that.

Which should you build?

Build a Copilot agent when the users are employees who already have Copilot, the knowledge lives in SharePoint, Outlook and Teams, and the outcome is a better answer or a faster draft. Onboarding assistants, policy questions, proposal drafting from past proposals and meeting follow-up are Copilot jobs. Our overview of Copilot Studio for small business covers the maker experience.

Build a custom agent when the job is a process rather than a conversation. Invoice intake across email, a portal and a scanner. An overnight reconciliation that runs whether or not anyone is at a keyboard. Anything where the approval logic is yours, the data lives outside Microsoft 365, or licensing every participant for Copilot would cost more than the process saves. That is the work our AI agents service and custom AI applications practice exist for; an AI readiness assessment sorts candidate processes into the two buckets first.

Frequently asked questions

Do we need Microsoft 365 Copilot licenses to use Copilot agents?

Not for every user. Microsoft's documentation states that Copilot Chat is included with eligible Microsoft 365 subscriptions and can run agents, with usage-based billing in Copilot Credits when an agent reads shared tenant data such as SharePoint or Copilot connectors. A Microsoft 365 Copilot add-on license removes those metered charges and adds Copilot inside the Office apps. Building agents in Agent Builder requires the Copilot license; Copilot Studio has its own subscription.

Can a custom AI agent read our SharePoint and Outlook data?

Yes, through the Microsoft Graph API with an application registration in Entra ID scoped to the sites, mailboxes and permissions the agent needs. Unlike a Copilot agent it does not inherit each user's permissions; it acts as its own identity, which is why the scope must be narrow and reviewed. Many custom agents combine both: they read a SharePoint list through Graph, pull matching records from an accounting system, and post the result into Teams.

Are custom AI agents safe to give access to company systems?

They can be, and the safety comes from design rather than from the model. Give the agent a dedicated identity with least-privilege access, restrict its tools to an explicit list, require human approval for any action that spends money or changes a record above a threshold, and log every tool call. Our AI usage policy guide covers the governance side, and our AI security assessment reviews an agent design before it touches production data.

Can we start with Copilot agents and add custom agents later?

Yes, and it is the order we usually recommend. Copilot agents show where people want help and expose the permission cleanup you need anyway. Once processes appear that Copilot cannot reach, because the data sits outside the tenant or the approval logic is yours, those become the first custom agents. A custom agent can even be surfaced inside Teams or Copilot as a custom engine agent, so people keep one interface.

Since 1998 we have built the reliable systems businesses run on, and agents are the newest layer of that stack. Our Microsoft Copilot consulting handles licensing, permission cleanup and the Copilot agent build, and our AI agents service designs, builds and hosts custom agents on Anthropic Claude with the approval steps and logging your auditors will ask about, on a month-to-month agreement. Talk to us and we will tell you which bucket your first few processes belong in.