Skip to main content

Copilot Cowork and Copilot Studio: where Microsoft wants your agents to live

AI article illustration for ai-jarvis.eu
Microsoft has published a walkthrough on its Copilot Studio blog, "Build apps in Copilot Cowork and Copilot Studio". The screenshots are the easy part. The hard part for a European team starts after the demo: which identity the agent runs as, where its data sits, what it consumes per conversation, and who signs off when it starts doing real work.

Two surfaces, one build

The pitch is simple to state. You define an app — its instructions, its knowledge sources, its tools and its data connections — once in Copilot Studio, and it can then surface inside Copilot Cowork, the collaborative space Microsoft positions as the place where people and agents work alongside each other rather than in a separate chatbot tab.

That is a distribution story, not a model story. Nothing in the post points to a new frontier model or a fresh reasoning benchmark. What it describes is the last mile — the part of agent development that has historically eaten most of the budget, because building a competent agent is now comparatively cheap, and getting it in front of the right forty people in an organisation is not.

Copilot Studio, for readers who don't live in the Power Platform

Copilot Studio is Microsoft's low-code builder for agents. You write instructions in natural language, attach knowledge sources (SharePoint sites, uploaded documents, Dataverse tables, external APIs through Power Platform connectors), add actions or tools, and then publish the result to channels — Teams, a website widget, a custom app, or increasingly the Microsoft 365 Copilot surface itself.

Two things make it different from a generic prompt wrapper. First, it runs on Microsoft's own plumbing: authentication is Entra ID, authorisation rides on Microsoft Graph permissions, and an agent answers with the signed-in user's rights rather than a shared service account. Second, administration is centralised — an IT department can see, approve, block and audit agents from the same console where it manages everything else in the tenant.

For anyone who has ever deployed a well-meaning internal bot with a super-user token, that is a real architectural difference, not a marketing line.

Why the "app inside the assistant" move matters

Every major vendor has a place to park a custom agent: OpenAI has custom GPTs inside ChatGPT, Google has Gems inside Gemini, Anthropic has Projects inside Claude. All of them work. All of them share one weakness — they live in a separate application, and separate applications lose to the one people already have open.

Microsoft's bet is that the surface is the moat. If the agent appears inside the same window as the calendar, the inbox, the Teams meeting and the document, the friction of switching context disappears. Whether the agent itself is smarter than a competing one becomes secondary.

There is a catch worth naming: agent quality is still agent quality. A mediocre agent that is always one click away is still mediocre — it is just faster at being mediocre.

The European checklist

Availability. Microsoft 365 Copilot and Copilot Studio are sold across EU markets, and the Copilot surface is localised for the major European languages. Answer quality in smaller languages is the variable to test, not availability.

Data residency. Microsoft states that its EU Data Boundary covers core Microsoft 365 services, which is what most European compliance teams will point to. Before you deploy, verify which of your specific data flows — connector calls, uploaded documents, telemetry — fall inside that boundary and which do not. The Microsoft Trust Center is the place to check, and your data protection officer will want the answer in writing.

GDPR. You remain the controller; Microsoft is the processor. An agent that reads employee records, ticket histories or customer conversations is processing personal data, and internal-only does not exempt you from a data protection impact assessment.

AI Act. Transparency obligations are the practical starting point: users need to know they are interacting with an AI. General-purpose AI obligations have applied since August 2025, and use-case classification matters more than the label on the product. An assistant that summarises meetings is one thing. The same technology pointed at CV screening, credit decisions or exam grading sits in a completely different risk category — and the classification follows the use case, not the vendor.

Price. Microsoft does not quote a single EUR figure in the post. Copilot Studio consumption is metered, capacity is sold in message packs and bundles tied to Microsoft 365 Copilot licensing, and EU list prices are set per market and exclude VAT. Budget the consumption rate before you scale to a thousand users, not after.

What I would test before rolling this out

Four things, in order. First, metered consumption per realistic conversation — a demo answers ten questions, a real user asks forty. Second, behaviour when the agent is shared across tenants or with external partners, because that is where identity models usually break. Third, what happens to the app and its data when a licence lapses or an admin leaves. Fourth, latency under production load — this is exactly the kind of measurement we run in AI Arena, and it is where polished demos and real deployments tend to diverge.

The short version

Microsoft is not selling a smarter model here. It is selling a shorter path from "we built an agent" to "our people actually use it", with enterprise identity and governance bolted on. For European companies already inside the Microsoft stack, that is genuinely useful. It is also a reminder that the deciding factors in enterprise AI in 2026 are distribution, governance and data location — not benchmark tables.

Do I need a Microsoft 365 Copilot licence to use Copilot Studio?

No. Copilot Studio can be licensed separately and agents can be published to standalone channels such as a website or Teams. But some of the deepest integration with the Microsoft 365 Copilot surface — and the most favourable consumption terms — comes with Copilot licensing. Check current terms with your reseller, because the packaging has changed more than once.

Can I build an agent that only answers from our own documents?

Yes. You can restrict an agent to specific knowledge sources and disable general model knowledge, which is usually the configuration compliance teams prefer. The agent will then say it does not know rather than guess — less impressive in a demo, considerably safer in production.

Is Copilot Cowork the same thing as Microsoft 365 Copilot?

Cowork is the collaborative framing — people and agents sharing the same working surface — while Microsoft 365 Copilot is the product family that delivers it. In practice you will meet the brand names before you meet a clear separation of the underlying plumbing.

Discussion

No comments yet — be the first to share your thoughts.
X

Don't miss out!

Subscribe for the latest news and updates.