Microsoft Copilot Studio

What it is

Microsoft Copilot Studio is the low-code builder for agents and agent workflows on the Power Platform. You describe an agent, give it knowledge (SharePoint, Dataverse, websites, the tenant’s Microsoft Graph), give it tools (more than 1,400 Power Platform connectors, MCP servers, agent flows, other agents), and publish it to Teams, Microsoft 365 Copilot, a website or a phone line. It is the successor to Power Virtual Agents, and it is the thing Microsoft Agent 365 governs: Agent 365 is the registry and control plane, Copilot Studio is where the agents get built.

2026 changed what the product is. Computer-using agents — agents that drive a browser or a Windows desktop app through its UI when there is no API — went generally available on 13 May 2026. Agent-to-agent delegation over the A2A protocol went GA in April. A second runtime, the GitHub Copilot harness, arrived as a production-ready preview in June and went GA in August, with skills, per-user memory and a planning loop that re-plans as it works; the older standard harness runs alongside it. Claude Sonnet 5 and GPT-5.5 Chat are selectable as the primary model. And since July 2026 every new agent gets a Microsoft Entra Agent ID automatically, with no environment-level opt-out.

Why ops teams end up here

The question that brings an ops leader to this page is usually: we are an M365 shop — do we build our ops agents in Copilot Studio, or buy n8n, Relevance AI or Lindy? Three things make Copilot Studio the default answer inside a Microsoft tenant. The agent lands where employees already work (Teams, Outlook, the Copilot app) with no new login. Identity, DLP and Conditional Access apply per agent through Entra Agent ID, which a security review will otherwise ask you to rebuild by hand. And for employees who already hold a Microsoft 365 Copilot licence, conversational usage of standard-harness agents is billed at zero.

The concrete ops cases where it earns its place: a RevOps agent in Teams that answers “what’s the status of the Contoso renewal” by reading Dynamics or Salesforce plus the account’s SharePoint folder; a legal-ops intake agent that classifies inbound requests from a shared mailbox and routes them with a human-approval pause; a recruiting coordinator agent that uses computer use to key offer details into a legacy HR portal with no API.

Pricing reality

Everything is metered in Copilot Credits, pooled across the tenant:

• Capacity pack: $200/month for 25,000 credits ($0.008 per credit), prepaid; unused credits do not roll over.
• Pay-as-you-go: $0.01 per credit, billed to an Azure subscription in arrears.
• Pre-purchase plan: up to 20% off for committing credits upfront.
• Microsoft 365 Copilot ($30/user/month; Copilot Business $21, $18 promotional through 31 December 2026, up to 300 seats) includes the standard harness and zero-rates employee-facing usage when the licensed user is the one talking to the agent.

The rate card is what turns into a budget. Classic (authored) answer: 1 credit. Generative answer: 2. Agent action — a trigger, a tool call, a reasoning step: 5. Tenant Graph grounding: 10. Agent flow actions: 13 credits per 100. Reasoning models add 10 credits per 1,000 tokens on top of the feature rate. A single grounded answer is therefore 12 credits, or about 10 cents at pay-as-you-go.

Two worked bands:

• Internal Q&A agent, 200 employees without Copilot licences, 20 grounded questions each per month: 200 × 20 × 12 = 48,000 credits — two packs, $400/month, about $4,800/year. The same agent used by 200 Copilot-licensed staff costs nothing extra in credits, but the licences themselves are $72,000/year.
• Autonomous pipeline-hygiene agent, 3,000 triggered runs a month at 6 actions plus 1 generative answer: 32 credits × 3,000 = 96,000 credits — $800/month in four packs, or $960 on pay-as-you-go. Autonomous triggers and agent flows started by anything other than a licensed user’s conversation bill at the standard rate whether or not your staff hold Copilot licences.

Computer use bills at the agent-action rate and is never covered by the Microsoft 365 Copilot licence; the hosted machine it runs on (Windows 365 for Agents) is a separate bill. The GitHub Copilot harness is a different meter again: it charges for tokens, tools and the harness itself, and it bills from the moment a maker starts building — previews, tests and evaluation runs included — where the standard harness only bills after publish.

Best for

An IT or RevOps leader in a Microsoft 365 E3/E5 organisation of 200-5,000 employees who needs employee-facing agents in Teams that read SharePoint and the CRM, take actions through connectors, and pass a security review without a separate identity layer. The use case where it wins outright: an internal agent whose users already hold Microsoft 365 Copilot licences, because the per-conversation cost goes to zero and governance comes with the tenant.

Not for

A 10-50 person ops team on Google Workspace, or a team whose agents mainly write to HubSpot, Salesforce and Slack rather than to Microsoft systems. You would be buying into Power Platform environments, capacity allocation and Entra administration to get what n8n gives you from €20/month or Lindy from $50/month. It is also the wrong buy for high-volume, customer-facing automation where you need a fixed cost per resolution: a credit meter that charges per step gets expensive when agents reason in long loops.

Versus the alternatives

• Salesforce Agentforce — the other market-share leader in enterprise agent builders. Pick it when the records the agent acts on live in Salesforce. Agents that sit next to the CRM record beat agents that reach into it over a connector, and Salesforce-centric revenue teams get further with the native trust layer than with a Power Platform connector.
• Google Gemini Enterprise — the direct counterpart for Google Workspace shops, priced per seat from $21/month with no-code agents built in. Pick it when email, docs and identity live at Google; Copilot Studio’s advantages mostly disappear outside a Microsoft tenant.
• n8n — the fastest-growing entrant in this segment, and the one to pick when the workflows are deterministic, the systems are mostly non-Microsoft, or you need to self-host. Pricing is per execution, not per reasoning step, which is far more predictable for high-volume flows. n8n agents also register in Agent 365 with an Entra Agent ID, so choosing it does not cost you tenant governance.
• Relevance AI / Lindy — pick these for a GTM team that wants working SDR or research agents this month without a Power Platform admin in the loop.
• UiPath — pick it over Copilot Studio’s computer use when UI automation is the core workload (hundreds of runs a day, SAP or mainframe screens, attended robots) and you need mature exception handling and orchestration rather than an agent that occasionally clicks.

Watch-outs

• The 125% cliff disables agents. When a tenant on prepaid capacity reaches 125% of its credits, custom agents are switched off — users see “This agent is currently unavailable” — and new agent flow runs are blocked once capacity is exhausted. Guard: link a pay-as-you-go Azure billing plan to every production environment so overage bills rather than blocks, and set per-agent monthly credit caps in Power Platform admin center → Licensing → Copilot Studio → Manage Agents.
• “Included with Copilot” is narrower than it sounds. Zero-rating covers conversations where a Copilot-licensed employee is the authenticated user. Autonomous triggers, agent flows with other triggers, computer use, and anyone without the licence all bill at the standard rate. Guard: before rollout, run the Copilot Studio agent usage estimator per agent, splitting licensed conversational traffic from autonomous and unlicensed traffic, and budget the second bucket at $0.01 per credit.
• The GitHub Copilot harness bills during build. Makers iterating in preview spend real credits before anything ships. Guard: give maker environments their own small credit allocation, separate from production, so experimentation cannot drain the capacity that production agents depend on.
• Two agent identities in one tenant. Agents created before July 2026 still run on classic app registrations and are waiting for a Microsoft-scheduled migration; new ones get Entra Agent IDs. Conditional Access written for one does not cover the other. Guard: export the agent inventory (admin center, API or Azure Resource Graph) and tag every agent by identity type, then write policy for both until the older agents are migrated or recreated.
• Premium connectors and cloud flows are a separate licence. Power Automate cloud flows use Power Automate licensing, not Copilot Credits; only agent flows run on credits. Guard: build agent-invoked automation as agent flows inside Copilot Studio, and price any standalone cloud flow against Power Automate Premium before you promise it.

Related: the enterprise AI rollout stack and the AI agent ops stack show where a builder like this sits next to the governance and observability layers.