The Dirt
This month’s dig starts here
Copilot isn’t one product — it’s a name Microsoft uses across several distinct tools that work in different ways depending on where you open them. It’s worth breaking down what’s actually happening under the hood in each version, and what that means for what you can do with it.
The Dig
Let’s tunnel in
The base layer: what “Copilot” means depending on where you open it
- Web and mobile — a standalone chatbot at copilot.microsoft.com and in iOS/Android apps, running on GPT-4o for free users.
- Windows 11 — lives in the taskbar, triggerable from a dedicated key on newer keyboards. This version can see and act on local files on your PC, not just answer general questions.
- Microsoft 365 (Word, Excel, PowerPoint, Outlook, Teams) — grounded in your organization’s data: your emails, documents, Teams messages, SharePoint files.
- Copilot Pro/Microsoft 365 Copilot (paid tiers) — faster processing, higher-quality image generation, and the ability to build custom agents.
- GitHub Copilot — a separate product for code, built by GitHub with Microsoft and OpenAI, with its own release cycle entirely independent from the others.
You’re not flying solo with 365’s Organizational Integration
When Microsoft 365 Copilot answers a question or drafts something, it isn’t generating purely from training data the way the free chatbot does. It’s retrieving specific content you already have access to (an email thread, a document, a meeting transcript) and using that as the basis for its answer, depending on your organization’s existing permissions.
Two direct consequences follow from that. First, it can’t see what you can’t see. If the information lives in a document you don’t have access to, Copilot won’t surface it, same as a coworker without access wouldn’t. Second, when the grounding data is thin or ambiguous, the model still has to produce an answer — that’s typically where inaccuracies show up, because retrieval-based answers are only as good as what got retrieved.
Model choice is now a real variable, not a fixed setting
As of recently, Microsoft 365 Copilot Chat lets you select Claude as the model instead of GPT for a given task. That means the “Copilot” brand and the model doing the actual reasoning are now separate layers — the interface, the grounding, and the permissions stay the same regardless of which model runs underneath.
Practically, that turns model choice into a per-task decision (document analysis vs. quick drafting vs. structured outlines) rather than something fixed by which product you bought.
Copilot Studio: building an agent, not just asking questions
Copilot Studio is a low-code tool for building a custom agent rather than using the built-in one. Mechanically, you’re doing three things: connecting the agent to a knowledge source (SharePoint, Dataverse, Microsoft Fabric, or an external site), defining what it should handle, and setting what actions it’s allowed to take — look something up versus actually trigger a workflow elsewhere.
A well-scoped agent grounded entirely in Microsoft data can be built this way in an afternoon, no engineering background required. The moment the agent needs to authenticate into a non-Microsoft system, or coordinate with other agents rather than act alone, you’re generally into custom development — the authentication and orchestration Copilot Studio handles natively for Microsoft systems doesn’t extend the same way to everything else.
GitHub Copilot: a different animal
Built by GitHub with Microsoft and OpenAI, and its own product entirely. It’s evolved from autocomplete-style suggestions into an orchestration layer: agent mode lets it pick files to edit, propose and run terminal commands, and iterate until a task is done, with a human approving anything sensitive. The cloud agent version runs asynchronously — assign it a GitHub issue, it works in a sandboxed environment, and it opens a pull request when finished, no one needing to be at a keyboard. As of late 2025, it can also route tasks to agents from other providers (Anthropic, OpenAI, Google) through the same interface.
So, who is Copilot best for?
Copilot makes the most sense for teams already working deeply within the Microsoft ecosystem. For marketers and small businesses, its biggest advantage is not simply generating content. It is also helping people find, understand, and act on information already spread across their emails, meetings, documents, and workflows. The more organized and accessible that information is, the more useful Copilot becomes.
The Shiny Stuff We Found
“Grounding” explains almost everything about how Microsoft 365 Copilot behaves.
It answers from what you have access to, not from general knowledge — which explains both its strongest moments and its gaps.
The model and the product are no longer the same decision.
With Claude selectable inside Copilot Chat, “which Copilot” and “which model” are now two different questions.
Copilot Studio’s real ceiling is authentication and orchestration, not intelligence.
Single agent, Microsoft-native data: low-code. Multiple agents or outside systems: an engineering project — good to know going in.
GitHub Copilot spans a spectrum, not a single feature.
From inline suggestions to a fully autonomous cloud agent opening its own pull requests.
The name Copilot may cover a lot of ground, but the basic flight plan is fairly simple: start with the work you need help with, identify where the relevant information lives, and choose the version built to access it. The right Copilot is less about having the most features and more about fitting the way your team already works.
Till next time, stay scrappy and we’ll keep digging.
Team Baby Badger
