Microsoft Copilot vs Copilot Studio: one seat, one meter, one open-source option

Which of the two a company is actually buying

Copilot and Copilot Studio are two different products. The question is which one a team needs. Copilot is the assistant a person opens inside Word, Outlook, Teams or the Copilot app. Copilot Studio is where a maker assembles agents and publishes them. Microsoft sells the first by the person and the second by the credit. Two contracts, two admin surfaces, two ways to run out.

GitHub Copilot and Copilot in Windows carry the name too, one for code in an editor and one for the desktop. Neither belongs in a comparison of Copilot against Copilot Studio. The two get confused because Microsoft sells both as company-wide AI in the same Microsoft 365 conversation, and because a subscription to one leads into the other.

A seat buys an assistant that reads the tenant

Microsoft 365 Copilot is a per-person add-on stacked on a Microsoft 365 plan. Microsoft's Copilot overview says every tier runs on three things: "Large language models (LLMs) for natural language understanding and generation", "Grounding in web and/or organizational data (Microsoft Graph and Work IQ)", and "Access scoped by user permissions (security and compliance enforced)".

The assistant reads a tenant's mail, files, meetings and chats through Graph. It answers inside the boundary each person already has. That is the reason to buy the seat. It is also the reason the seat ends where Microsoft 365 ends. For a company that lives in Word and Outlook, that fit is the whole argument.

The tiers differ in how much organizational data arrives without anyone pasting it in. Copilot Chat (Basic) grounds on web data, and on whatever a person supplies with the prompt. The paid tier grounds "automatic via Microsoft Graph and Work IQ". Copilot Cowork and Copilot Code access bills by consumption on top of the add-on, a third meter on the same invoice.

A meter buys agents, counted in Copilot Credits

Copilot Studio builds agents. A maker gives one instructions, knowledge and connectors. Then the maker publishes it into Teams, SharePoint, Copilot Chat or a website. It is bought by consumption too. Microsoft's billing and licensing page states the change: "Starting on September 1, 2025, the common currency for agents changed from messages to Copilot Credits."

The entry point looks free. Microsoft calls the Copilot Studio builder arrangement "free of charge". The same page names the condition: a company needs "a Copilot Studio tenant prepaid Copilot Credit pack subscription" before the arrangement can be issued. Credits then arrive as a prepaid pack, a one-year prepurchase plan, or an Azure pay-as-you-go meter tied to a Power Platform environment.

Capacity behaves like a meter. "Copilot Studio enforces purchased capacity monthly, and unused Copilot Credit don't carry over to the next month" (the same billing and licensing page). Usage past what was bought meets technical enforcement that can end in service denial. A trial gets a maker the test chat, and no way to publish an agent.

Where the two rents come from

A Microsoft 365 Copilot subscription is one of the listed ways into Copilot Studio. Microsoft describes it as a way to "extend Microsoft Copilot with agents". Agents used in Copilot Chat, Teams and SharePoint for classic answers, generative answers and tenant grounding come out "zero-rated". That path rides inside the seat.

Now the other path. An agent that answers a customer on a website, works a Slack thread or runs on a nightly trigger sits outside those surfaces. It draws on the credit meter. The seat covers the assistant and the meter covers the agent. A company that wants both capabilities buys both products. The bill then carries two lines priced in two units, a person and a credit.

Ownership is the other half. The assistant's behaviour is Microsoft's to change, and its grounding lives in the tenant. An agent's definition lives in a Power Platform environment, governed through an admin center and reachable through the studio. A team can edit instructions and connectors. It cannot read the software that executes them, fork it, or carry the environment to another host when the price or the roadmap moves.

One system instead of two invoices

A team that wants the assistant and the agent to be one owned thing writes both into a repo. In Kortix, the open-source AI Operating System, an agent is an agents/<name>.md file whose frontmatter carries the mode and the permission, plus an agents.<name> block in kortix.yaml that carries the governance. Skills, memory, connectors and secrets live in the same repo. The agents run on cloud computers, landing finished work as a change request a person approves.

The starter project ships that file. It takes a minute to read: a description, a mode, a permission line, then the prompt itself. A change to how the agent behaves arrives as a diff a reviewer reads. A company that leaves takes the repo with it.

Everything a buyer would otherwise negotiate with a vendor is a setting in that repo. Where it runs: "Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free." Who may touch what: "Per-resource permissions for people and agents. Roles, groups, and an audit trail." Which model answers: "Any model provider with your own keys." And the code itself: "Open source. Read it, fork it, audit it."

The case for buying Microsoft's pair

A team already deep in Microsoft 365 has a reason to stay. Answers grounded in the tenant's own mail and files, inside permissions people already hold, are hard to rebuild elsewhere, and the meter has a published rate card before anyone signs. Nothing needs hosting, patching or upgrading, and the GCC plan covers public-sector buyers.

The argument is narrower than a verdict on Microsoft. Two products, two meters, two sets of behaviour a company cannot read, bought to cover one job that each covers halfway. A team that notices the second invoice is a team that wants both halves in one system it owns.