Field Notes from the Frontier
Which AI provider powers each Microsoft 365 Copilot surface, which governance tier that puts your customer's data under, and what changes the moment their tenant sits in the EU or a government cloud. Every cell cites Microsoft. Where Microsoft doesn't publish, this says so instead of guessing.
Availability is not universal. Set this to your customer's scope and the matrix re-renders to what they can actually turn on.
Each cell shows how the model gets chosen on that surface. The colour of the left bar is the governance tier — that's the data-residency answer. Click any cell for the talk track and the citation. Click a provider header to isolate its column.
Elsewhere Microsoft names models but routes between them dynamically. Studio is the only place you pin one and know that's what runs — and the only place with a published per-region availability table. Filter by provider, or read the government column; it's the shortest column on the page for a reason.
Microsoft classifies every AI provider as hosted, subprocessor, or independent processor. That classification — not the model name — determines whether your customer's DPA, data residency, SLA, and Customer Copyright Commitment apply. It is the whole answer to "where does my data go."
Microsoft names models per surface in blog posts and the Learn release notes — GPT-5.6 is the stated preferred model for Word, Excel, PowerPoint, Chat, and Cowork. But it publishes no canonical per-surface table outside Copilot Studio, and routing between named models stays dynamic. So "named" is not "pinned," and any map showing one fixed version per feature is inventing precision.
Learn and the product pages drift apart. A prerequisite documented at launch can stay published months after it stops being true. Where that happens, this tool shows both sources with their dates rather than silently picking one — those cells carry a ⚠ in the matrix.
External models in Copilot Studio need enabling twice: the provider in the Microsoft 365 admin center, then external LLMs in the Power Platform admin center for that environment. Missing the second gate is the most common reason a maker can't see a model.
Models tagged cross-geo may process and store data outside the tenant's region. Anthropic models in Microsoft offerings are currently excluded from the EU Data Boundary outright. If your customer has in-country processing commitments, start here.