There is a pattern that repeats every time someone brings up Copilot in an IT conversation: the snicker. "Clippy on steroids." "The one that copies ChatGPT, late and badly." On Capterra and G2, 2026 reviews repeat variations of the same jab: "I view it as a copycat of ChatGPT," "it doesn't quite match Claude or Cursor for code," "sometimes the answers feel off from what I asked." According to Microsoft's own numbers, only 4% of licensed users actually use it actively.
And yet, while that reputation was hardening in forums and reviews, something completely different was happening inside Redmond. Copilot stopped depending on a single AI provider. It learned to turn an Excel table into a clickable app inside PowerPoint itself. Its agents started getting their own identity in the company directory, as if they were one more colleague. None of it matched the reputation the product already carried.
This is an article about that gap. About a product that still looks like the ugly duckling on the surface while, underneath, the swan feathers are already coming in.
The duckling everyone points at
Copilot's bad reputation did not come from nowhere. It was born fragmented: one assistant for Word, another for Teams, another for code, one for consumers and another with an almost identical name for businesses. Nobody quite knew which one to open, and the confusion between "Copilot" and "Microsoft 365 Copilot" was, for years, a recurring complaint among IT administrators.
On top of that came the free tier problem. Through 2025 and much of 2026, the free version worked as a permanent demo: enough to show the concept, not enough to replace anything. In April 2026, Microsoft directly limited free Copilot Chat inside Word and Excel. In August, it pulled Deep Research for consumers. The feeling of "this is bait toward the paid tier" got burned into the experience of millions of users who tried the free version and drew their own conclusions right there.
The first impression was formed by the 2023-2024 version of the product. The 2026 version is, functionally, something else. But reputation doesn't update at the same pace as software.
It is the classic problem of any ugly duckling: judged by how it looked at birth, not by what it is becoming.
The day it stopped pretending only OpenAI existed
For years, using Copilot meant using GPT. There was no other option, and it wasn't up for debate. That changed in 2026, and it is probably the most underrated shift in the whole story.
Today, within the same Microsoft 365 Copilot license, any user can choose between GPT (OpenAI) and Claude (Anthropic) in Copilot Chat, Excel, Word, PowerPoint, Copilot Studio and the Researcher agent. No extra fee, nothing else to install. The choice sits with the user, depending on the task.
This is not a cosmetic detail. It is a company that for years sold "everything in one closed ecosystem" admitting, in practice, that the best tool for the user isn't always its own. There's an important detail few people read all the way to the end: Claude models are processed outside infrastructure directly managed by Microsoft, with data transferred to Anthropic, and in the European Union, EFTA and the United Kingdom the admin has to turn it on manually. It's not just a menu option, it carries real data governance implications any compliance team should review before enabling it.
But the underlying message matters more than the technical detail: the ugly duckling stopped pretending it could do everything alone.
When PowerPoint became an app
In August 2026, Copilot in PowerPoint gained a skill that turns tables, Excel spreadsheets and even non-numeric content into interactive slides: with filters, buttons and clickable navigation, working in normal editing mode, no need to start the presentation. A headcount table becomes a panel filterable by year and department. A timeline becomes click-through navigation. All inside the same .pptx file as always, no code, no new platforms.
(We already covered this in detail here, with concrete examples and a comparison against Power BI.)
This kind of leap doesn't fit the caricature of "the one that copies ChatGPT, late." It's a feature no competitor had, solving a problem anyone who builds corporate decks recognizes instantly: the report that's born old the moment it becomes a PDF.
Someone new in the system
The hardest part to explain to someone who still sees Copilot as a chatbot with good marketing is this: it stopped being just an assistant that suggests text. It started having agents with their own identity.
In May 2026, Microsoft launched Agent 365, a platform for registering and governing AI agents inside the tenant. An agent built there gets its own identity in Entra ID (not a delegated copy of whoever created it), shows up in the directory alongside human colleagues, can be mentioned in Teams, receive emails in Outlook, and be referenced on SharePoint pages. The tenant admin approves, audits and revokes it like any other registered application.
In parallel, the Model Context Protocol (the standard Anthropic donated to a neutral Linux Foundation body) reached general availability in Copilot Studio and in M365 Copilot's declarative agents, allowing Copilot to connect to external systems without custom connectors. And the new AutoPilot tier introduced Microsoft Scout, an agent that manages calendars and tracks deadlines in the background, without anyone having to ask each time.
| Copilot 2023-2024 (the reputation) | Copilot 2026 (the facts) | |
|---|---|---|
| AI model | OpenAI only, no choice | GPT or Claude, user's choice |
| Agent role | Suggests text, waits for instructions | Has its own identity in Entra ID, acts on its own |
| PowerPoint | Generates text and images | Turns data into interactive apps |
| External integration | Custom, bespoke connectors | MCP: open standard, no code |
| Free tier | Generous entry point | Deliberately limited demo |
None of this is what people picture when they hear the word "Copilot" in a meeting. And that's exactly the point.
The point
The ugly duckling of the story doesn't change overnight. It changes while nobody's watching, until one day someone looks up and no longer recognizes what's in front of them. Copilot is exactly at that midpoint: it still carries the reputation of 2023, but it already runs on the substance of 2026.
That doesn't mean it's worth buying for any company tomorrow without thinking it through (we already covered that here), nor that the criticism about answer quality has fully disappeared. It means judging the 2026 product by an experience someone had in 2023 is, at minimum, a dating error.
The question worth asking isn't whether Copilot is going to become a swan. The question is how many companies will keep staring at the duckling long after the feathers already changed.
If you want to keep up with this kind of shift before it becomes obvious, macareno.net's newsletter brings exactly this kind of analysis every week.
Sources
- Microsoft 365 Copilot Reviews 2026 — Capterra
- Microsoft Copilot Reviews 2026 — G2
- Expanding model choice in Microsoft 365 Copilot — Microsoft 365 Blog
- Your First Teams-Native n8n Agent on Microsoft Agent 365 — Automation Labs