A lot of teams roll out Microsoft Copilot with real enthusiasm, then quietly keep a ChatGPT tab open for anything that actually matters. If Copilot feels a step behind on hard tasks, you're not imagining it, and the gap isn't really a knock on the idea of AI in Office; it comes from a handful of structural choices in how Copilot is built and deployed.
You don't control which model runs
With a general-purpose chat tool, you can choose the newest, strongest available model for a hard problem, and switch to a different one if the first attempt disappoints you. Copilot runs whatever model and version Microsoft has wired into that particular surface at that particular time, and there is no way to swap in a different provider or a stronger model when the built-in one struggles with your specific task. For routine requests this rarely matters, but for your hardest prompts, that fixed ceiling is often the entire problem: the same request that Copilot handles unevenly might be handled well by a different model entirely, and Copilot simply doesn't offer that choice.
Context and grounding are doing a lot of the work
Copilot's answers depend heavily on a retrieval step that pulls context from your documents, emails, and tenant data before the model ever generates a response, a process usually called grounding. When that retrieval step finds the right material, Copilot can feel remarkably sharp. When it retrieves the wrong document, an outdated version, or simply too little relevant context, the resulting answer feels vague, generic, or subtly wrong, even though the underlying model might have done fine with better input. In a plain chat interface, you paste exactly the context you want the model to use, which removes an entire failure mode and is a quiet reason the same request often "just works" there when it doesn't in Copilot.
The context window and history handling differ too
How much prior conversation and document content a tool keeps available to the model, and how it decides what to trim when a task runs long, shapes output quality just as much as the model itself. Different Office surfaces and chat tools make different tradeoffs about how much context to carry forward, and those tradeoffs aren't always visible to the person typing the prompt. A tool that silently drops older context to stay within budget can produce answers that seem to "forget" earlier instructions, which reads as the model being weaker when the real cause is context management.
One assistant, one set of tradeoffs
Copilot is, by design, a single assistant with a single set of behaviors tuned to work reasonably well across many tasks at once. But different kinds of work, drafting long-form prose, financial analysis, code generation, summarizing a dense document, genuinely perform better on different models; providers optimize their models differently, and no single model is the best choice for every task. A one-size-fits-all surface can't specialize the way a person choosing their own tool for each job can, which is a structural limit rather than a bug that a future update quietly fixes.
Deployment and version lag
Even when a stronger model exists, enterprise software has to qualify, test, and roll it out across a huge and varied customer base before it reaches your tenant, which introduces a lag between what's available on the open market and what Copilot is actually running for you on a given day. A general chat product iterating on its own timeline can adopt a new model faster than an enterprise suite can validate and ship the same change across every customer and every Office surface.
How Switchboard helps
Switchboard's native Excel, Word, and PowerPoint add-ins put frontier models from multiple providers directly inside the documents your team already works in, so you aren't stuck with one assistant's model choice or its grounding limitations. You pick the right model for the task at hand, the add-in keeps the work inside Office rather than forcing a copy-paste round trip to a separate chat tab, and every request stays governed and cost-attributed the same way it would be anywhere else in Switchboard. You get Office-native convenience without giving up the frontier-model quality that sent people back to ChatGPT in the first place.