Is Copilot Even in the Game?

Is Copilot even in the game?

I’ve used it. I’ve also run OpenAI’s Codex and Anthropic’s Claude on top of it, and used both straight through their own apps and web versions instead. The difference is not close. Copilot feels stripped down next to the other two: fewer tools, more refusals, less of whatever makes the other models actually useful [17].

Turns out the feeling matches the data. Copilot is losing ground to ChatGPT, Gemini, and Claude by almost every independent measure [1][2][3][4]. But the interesting failure here isn’t Microsoft’s. Most people who use AI, on any tool, skip the one skill that actually decides whether it’s useful: controlling the context, not polishing the prompt.

The Numbers

Copilot’s web market share sits at 1.1%, down from 1.5% a year ago. ChatGPT holds 64.5%. Gemini holds 21.5%. Even Grok, a distant fourth everywhere else, beats Copilot at 3.4% [1]. Windows Latest reported the same stuck-at-1% number independently a few months earlier [2].

The download numbers are worse. About 79 million Copilot downloads against ChatGPT’s 900 million. Only around 1% of Windows’ 1.5 billion users actually open it [4]. Fewer than 4.5% of Microsoft 365’s 450 million customers pay for the Copilot features at all, and GitHub Copilot specifically has been losing developers to Claude Code and Cursor [3]. Microsoft had its worst stock quarter since the 2008 financial crisis over this. Shares down more than 20% in 2026. The Windows AI push earned the nickname “Microslop” [5]. Current and former employees describe confusing branding and interoperability problems to the Wall Street Journal, and some users say Copilot was flat out forced onto them [6].

That’s one direction.

Microsoft’s own disclosed numbers pull the other way. Around 420 million monthly active Copilot users in Q1 2026. 15 million Microsoft 365 Copilot enterprise seats. Seat growth over 160% year over year. Roughly 70% of the Fortune 500 holding a license [7]. A BNP Paribas research note argues the “falling behind” story is already out of date [8]. But independent surveys folded into that same reporting say only 20 to 30% of the seats companies bought actually get used weekly [7].

Both numbers are true at once. Enterprises bought Copilot in bulk. Most of the people holding a seat aren’t opening it.

The Guardrails

Enterprise Copilot is locked down on purpose. It inherits your company’s permissions, sensitivity labels, and retention rules, and it logs what you do with it [9]. That’s not nothing. The same summarizing that makes it useful can also surface PII or confidential data nobody asked for, and a permissions mistake can hand that data to the wrong person [10]. When a tool touches data like that, the restrictions are the point, not an accident.

I still find it frustrating day to day. Claude and Codex, used directly, feel like a different product than the same models sitting inside Copilot [17]. Not because the model changed. Because the wrapper stopped getting in the way. The restriction and the frustration are the same feature, seen from two different sides of the desk.

Just Ask Copilot

Do you use Copilot because you chose it, or because you’re required to?

I’ve lost count of how many times a leader has told me to just ask Copilot. Run it through Copilot. Or the laziest version: just ask AI [17]. Said like AI is a vending machine. Put in a question, get the right answer out, no thinking required.

That framing costs something real. A 2025 study of 666 people found a measurable link between how often someone uses AI tools and how they score on critical thinking tests, strongest in younger, less educated participants [15]. MIT’s Media Lab wired people up to EEG while they wrote essays: one group used an LLM, one used a search engine, one used nothing. The LLM group showed the weakest brain connectivity of the three, felt the least ownership over their own essays, and struggled to quote what they’d supposedly just written. The researchers call it cognitive debt [14].

AI isn’t the problem. Dumping in a mess of text with no context and expecting a clean answer back is the problem. “Just ask AI” as a management strategy makes that worse, not better.

Prompt Engineering Is Dead

Somebody told me once, in these exact words: “you just need to know how to prompt it.”

That’s technically true. It’s also not where the leverage is anymore. Prompt engineering, getting the phrasing of one ask right, is a shrinking part of getting good output. Anthropic’s own engineering team says the bigger lever is context engineering: curating what the model sees, when it sees it, and what it’s allowed to skip [11]. Andrej Karpathy said the same thing in 2025: context engineering, not prompting, is the real skill in any serious LLM application. Plenty of the industry has said it since [11][12][13].

Controlling what the AI gets, and when it gets it, is a different job than writing a clever question. Most “AI failures” are context failures. Somebody pastes a wall of text into a chat window, copies out whatever comes back, and drops it into a document or an email like that’s the finished product. That’s not you using AI as a tool. That’s you working for the AI. Copilot makes this easy to fall into. So does everything else.

Files and Folders

The fix isn’t a smarter prompt. It’s a series of files and folders: instructions, naming rules, file locations. Rules for what gets read and when, routed to the AI only when it actually needs it [16].

That’s the idea behind Interpretable Context Methodology, ICM. Instead of retyping the same instructions into a chat window over and over and getting a slightly different, occasionally hallucinated answer each time, you put the workflow on disk. Folders route the request. Reference files hold the standing rules. Each stage does one job and reads only what that job needs [16].

It works underneath Copilot the same as anywhere else. ICM doesn’t fix Copilot’s download numbers, and it doesn’t unblock a restriction your IT department put there for a real reason. It fixes the part that was always yours to control: what the model sees, in what order, before you hit send.

Before You Blame the Tool

Before you blame whatever AI your employer mandates, figure out which failure you’re actually looking at. Is the model genuinely not up to the task, or did it get a pile of unstructured text and no instructions about what mattered? The first one might need a different tool. The second one doesn’t. Swap Copilot for Claude and you’ll carry the same failure into a nicer interface.

Copilot’s numbers are real. The frustration with its restrictions is fair, and so is the reason those restrictions exist. None of that was ever really about which model answered. It was about whether anyone did the work of deciding what the model saw before asking it anything.

Spend more time planning what goes in. Let AI do less of the actual thinking for you. Stop treating chat history as a workspace, and start building one you can actually see.

One More Thing

I wrote this using the same setup I’m describing. Input files, ICM. A stage that only researches. A stage that only organizes sources. A stage that only drafts. Each one read just what I told it to. The Copilot numbers, the cognitive debt research, and the context engineering sources all got filed before I wrote a single sentence of this. The entire process was controlled by me, reviewed by me, seeded by my thoughts, and comments. AI being used as a tool to do all the typing for me.

Sources

[1] “Pretty much no one is using Microsoft’s Copilot AI, report suggests,” PCWorld. https://www.pcworld.com/article/3029760/pretty-much-no-one-is-using-microsofts-copilot-ai-report-suggests.html

[2] “Is Microsoft losing the AI race? Copilot (web) is still stuck at 1% market share,” Windows Latest, 2026. https://www.windowslatest.com/2026/01/09/is-microsoft-losing-the-ai-race-copilot-web-is-still-stuck-at-1-market-share-we-dont-know-how-popular-it-is-on-windows-11/

[3] “Is Microsoft Copilot Falling Behind in the AI Race?” techcheck. https://techcheck.online/is-microsoft-copilot-falling-behind/

[4] “Microsoft Copilot vs. AI Rivals: Why It’s Falling Behind in the Global Adoption Race,” Windows Forum. https://windowsforum.com/windows-news.4/microsoft-copilot-vs-ai-rivals-why-its-falling-behind-in-the-global-adoption-race.373824/

[5] “AI Is Killing Microsoft,” Futurism. https://futurism.com/artificial-intelligence/ai-killing-microsoft

[6] “Inside Microsoft’s Copilot Crisis: How the Tech Giant’s AI Flagship Lost Its Way,” WebProNews. https://www.webpronews.com/inside-microsofts-copilot-crisis-how-the-tech-giants-ai-flagship-lost-its-way/

[7] “Microsoft Copilot Enterprise Adoption in 2026: What the Data Shows,” Stackmatix. https://www.stackmatix.com/blog/microsoft-copilot-enterprise-adoption-2026

[8] “From Lagging to Leading? Microsoft Copilot’s Commercialization May Exceed Market Expectations,” NAI 500, 2026. https://nai500.com/blog/2026/06/from-lagging-to-leading-microsoft-copilots-commercialization-may-exceed-market-expectations/

[9] “Enterprise data protection in Microsoft 365 Copilot and Microsoft 365 Copilot Chat,” Microsoft Learn. https://learn.microsoft.com/en-us/microsoft-365/copilot/enterprise-data-protection

[10] “Microsoft Copilot Security: 2026 Guide to Risks, Oversharing & Safe Enterprise Rollout,” Strac. https://www.strac.io/blog/microsoft-copilot-security

[11] “Effective context engineering for AI agents,” Anthropic. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents

[12] “Why AI teams are moving from prompt engineering to context engineering,” Neo4j. https://neo4j.com/blog/agentic-ai/context-engineering-vs-prompt-engineering/

[13] “Context Engineering: The Next Frontier Beyond Prompt Engineering,” deepset. https://www.deepset.ai/blog/context-engineering-the-next-frontier-beyond-prompt-engineering

[14] “Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task,” MIT Media Lab. https://www.media.mit.edu/publications/your-brain-on-chatgpt/

[15] “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking,” MDPI, 2025. https://www.mdpi.com/2075-4698/15/1/6

[16] Jake Van Clief and David McDermott, “Interpretable Context Methodology: Folder Structure as Agent Architecture,” arXiv, 2026. https://arxiv.org/html/2603.16021v2

[17] My own experience using Copilot, Claude, and Codex, and workplace conversations about AI tools.