Microsoft Copilot Agents in 2026: How They Work, What They Cost, and What They Can't See

TL;DR

• Microsoft Copilot agents are the best agent option for companies that run on Microsoft 365. They answer from Microsoft Graph, the mail, files, chats and meetings each user can already open, and they inherit Microsoft 365 permissions, sensitivity labels and compliance controls.
• Copilot agents come in two kinds: declarative agents built with Agent Builder or Copilot Studio on Copilot's own models, and custom engine agents that bring their own orchestrator. Copilot Studio adds outside data through 100+ Copilot connectors and MCP servers.
• Pricing is two meters. The Microsoft Copilot licence is $30 per user per month paid yearly. Copilot Studio usage is billed in Copilot Credits, $200 for a pack of 25,000 a month, and employee use by Copilot-licensed users is included within fair use.
• Copilot's memory is personal: saved memories live in each user's mailbox. Sentra adds the shared company memory, joined across ERP, CRM, HRIS and conversations, that Copilot agents can read over MCP.

Microsoft Copilot agents: what they are

Microsoft Copilot agents are AI assistants built on Microsoft Copilot that answer questions and carry out tasks for a specific job, using Microsoft 365 data and any connected systems. Microsoft builds some, such as Researcher and Analyst, and companies build their own with Agent Builder or Copilot Studio.

A naming note first. Microsoft has renamed Microsoft 365 Copilot to Microsoft Copilot, and Microsoft 365 Copilot Chat to Microsoft Copilot Chat. Some licences and screens still use the old names during the transition (learn.microsoft.com). This page uses the new names.

Microsoft documents two types of agent (learn.microsoft.com):

• Declarative agents. They run on Copilot's own orchestrator and models, adding custom instructions, knowledge and actions. They inherit Microsoft 365 security and compliance and need no extra hosting. You build them in Agent Builder (no code), Copilot Studio, or developer tools.
• Custom engine agents. They bring their own orchestrator and models, built with Copilot Studio, the Microsoft 365 Agents SDK or Microsoft Foundry. Hosting and compliance are the builder's responsibility.

Microsoft's Copilot plan page lists Researcher and Analyst as included agents that "employ advanced reasoning" (microsoft.com).

If you are comparing Copilot with ChatGPT, Claude or Gemini as assistants, our ChatGPT vs Copilot vs Claude vs Gemini comparison covers that. This page is about agents.

Copilot Studio and how Copilot agents are built

Copilot Studio is Microsoft's tool for building, publishing and managing Copilot agents, including agents that run on their own triggers rather than waiting for a chat.

What a buyer should know about it:

• Knowledge. Agents can ground answers in your tenant-wide Microsoft Graph through tenant graph grounding, including external data synced in through connectors (learn.microsoft.com).
• Copilot connectors. Formerly Graph connectors. Microsoft and partners offer more than 100 prebuilt connectors, including Box, Confluence, Google services, Salesforce and ServiceNow. Synced connectors index external data; federated connectors query it in real time (learn.microsoft.com).
• MCP. Copilot Studio agents can connect to existing MCP servers and use their tools and resources. Generative orchestration must be turned on (learn.microsoft.com).
• Admin control. Admins choose which agents are allowed in the organization, and a user can only access agents the admin allows (learn.microsoft.com).
• Other vendors' agents. SAP's Joule is available as an agent inside Copilot (help.sap.com). See our SAP Joule guide.

Microsoft Copilot agents pricing

There are two separate meters, and buyers often budget for only one.

The Microsoft Copilot licence also needs a qualifying Microsoft 365 plan. On its Copilot Studio pricing page, Microsoft says the $30 Copilot add-on lets licensed users build internal agents within Microsoft 365 at no additional cost, but publishing to external channels and letting unlicensed users run agents need Copilot Studio capacity (same source).

What a Copilot agent action costs

Microsoft prices agent work in Copilot Credits (learn.microsoft.com):

The "no charge" column applies to employee-facing use when the user is licensed for Microsoft Copilot and the agent runs under that user's identity, within fair use limits. Computer-using agents are not included. Reasoning models add a premium charge of 10 credits per 1,000 tokens on top of the feature rate.

A worked Copilot agent cost example

Microsoft's own example is a sales performance agent grounded in tenant graph data. Each run uses four generative answers and four tenant graph groundings: 4 x 2 + 4 x 10 = 48 credits. With 50 licensed users and 100 unlicensed users, only the unlicensed users count, so the agent consumes 4,800 credits a day (same source).

Over 22 working days that is 105,600 credits a month, which needs five 25,000-credit packs, or $1,000 a month, or the same usage billed pay-as-you-go. License all 150 users for Copilot instead and the agent costs nothing extra, but the licences cost 150 x $30 = $4,500 a month. The cheaper path depends on how many people use the agent and what else they use Copilot for.

Watch the cap. When a tenant reaches 125% of prepaid capacity, custom agents are disabled until capacity is added, unless pay-as-you-go is linked (same source).

What Copilot agents can see and what they can't

• Microsoft Graph. Copilot answers from user documents, emails, calendar, chats, meetings and contacts, and only surfaces data the user has at least view permission for (learn.microsoft.com).
• Sensitivity labels. When content is encrypted with Microsoft Purview Information Protection, Copilot honours the usage rights granted to the user (same source).
• Connected systems. Data from Copilot connectors appears in answers when the user has permission to see it, and MCP servers add live tools.
• Training. Microsoft states that prompts, responses and data accessed through Microsoft Graph aren't used to train foundation models (same source).

• Systems you haven't connected. An ERP, HRIS or finance system is invisible until you add a connector, an MCP server or a custom action. Each is a project with its own owner.
• A shared company memory. Copilot memory is personal. Microsoft documents that memories "are stored in the user's Exchange mailbox in a hidden folder" (learn.microsoft.com). What one person's Copilot learned about a customer does not inform a colleague's agent.
• Governance over that memory. On the same page, Microsoft states that Purview retention policies don't apply to Copilot memory, memory actions don't generate audit log entries, and admins can't restrict what information is added to memory. Copilot memory is also still in preview.
• The join between systems and conversations. A Copilot agent can read an email and, if connected, an ERP record. Nothing in Microsoft's documentation describes a maintained record of which commitment superseded which, who owns it now, and when the old fact stopped being true.

None of this is a flaw for personal productivity. It is the gap that shows up when a company wants every agent, Microsoft's and others', to give the same answer about the same customer, order or project.

Works with Sentra: company memory for Copilot agents

Copilot knows each user. Sentra knows the company. The two fit together.

The mechanism:

1. Connect. Sentra connects to email, chat, meetings, documents, ERP, CRM, HRIS and BI through 200+ integrations, including the Microsoft 365 data Copilot uses. Nothing is entered by hand.

2. Understand. It builds one organization-wide record of what is true now, who owns it and what was committed. It knows when a fact became true and when it stopped being true, so an agent does not repeat a superseded price or date as current.

3. Serve Copilot over MCP. Sentra exposes MCP and REST. Copilot Studio agents connect to MCP servers and use their tools and resources (learn.microsoft.com), so a Copilot agent can ask Sentra before it answers or acts. The same memory serves ChatGPT, Claude, Gemini, Agentforce and internal agents.

4. Respect permissions. Answers respect each source's existing permissions, checked at query time. A failed check counts as no access.

5. Act first, with approval. Sentra's own agent finds what is going wrong, works out who should act and drafts the alert. A named person approves before anything reaches anyone else.

An illustrative case: a finance lead asks a Copilot agent why a customer's invoice is overdue. Copilot finds the invoice email in the lead's mailbox. Sentra adds what the lead's mailbox doesn't hold: the ERP shows a partial shipment, and the account manager agreed in a Teams call last week to hold billing until the remainder ships. The agent answers with both sources, and the follow-up goes to the account manager, who owns it.

Better context also cuts cost. On Terminal-Bench 2.1, agents using Sentra scored 88.31% against an 83.37% baseline, with 72.6% lower model cost and 41.2% fewer tokens. On a credit meter, fewer tokens is fewer credits.

Sentra is SOC 2 Type II and ISO 27001 certified and deploys in the cloud, in an isolated VPC, or air-gapped on-premises.

To see Sentra working inside your Copilot agents, book a demo.

Alternatives and complements to Copilot agents

How to choose and how to start

1. Start with what you already pay for. If your users hold Microsoft Copilot licences, build a first declarative agent in Agent Builder before buying anything else.

2. Count unlicensed users. They are what turns Copilot Studio into a credit bill. Run Microsoft's agent usage estimator with real traffic.

3. List the data the agent needs outside Microsoft 365. For each source, choose a Copilot connector, an MCP server, or a shared company memory that already joins it.

4. Decide where memory should live. Personal memory in each mailbox suits individual productivity. Company facts that every agent must agree on need an organization-wide record with audit and retention.

5. Set approval rules. Any agent that sends, changes records or commits the company should have a named approver.

For more on agent strategy, see enterprise AI agents, AI agents for business, agentic AI vs AI agents and the best enterprise AI platforms.

FAQ