Microsoft hits 20M Copilot users

Microsoft’s Copilot story just got more real. Not because the demos changed, but because the seat counts did. Microsoft said Microsoft 365 Copilot now has more than 20 million paid enterprise seats, and that matters because this is the first time the “AI at work” pitch starts to look like ordinary enterprise software adoption instead of a giant experiment. ### Why does 20 million matter? A lot of AI usage numbers are soft. Monthly actives. Trials. People who clicked once. Paid enterprise seats are harder to fake — companies budget for them, procurement signs off, and managers expect employees to actually use the thing. Satya Nadella gave the 20 million figure on Microsoft’s latest earnings call, which makes it one of the clearest signals yet that big companies are buying AI assistants in bulk. ### Is this just lots of small customers? No — and that’s the more interesting part. Microsoft said it has quadrupled the number of customers buying more than 50,000 seats. It also named some very large deployments: Bayer, Johnson & Johnson, Mercedes, and Roche each have more than 90,000 seats, and Accenture agreed to buy more than 740,000. That is not “let’s let one team try it.” That is core workflow software. ### What are companies actually buying? Basically, they are buying convenience with guardrails. Copilot lives inside Microsoft 365 — Word, Excel, Outlook, Teams, PowerPoint — where the permissions, identity systems, and compliance controls already exist. That makes the sales pitch much easier than asking a company to help workers get through routine tasks faster without moving sensitive information into random third-party tools. This is partly inference from Microsoft’s product position and the kinds of customers it highlighted. ### Why mention OpenAI here? Because the same week showed the other half of the enterprise AI market. OpenAI rolled out GPT-5.5-Cyber with restricted access for vetted defenders rather than a wide release, which tells you even frontier labs think some capabilities need tighter distribution. The point is not “bigger model for everyone.” The point is “specific model for a sensitive job, with controls.” ### And what was the Codex giveaway about? OpenAI also emailed more than 8,000 developers who had applied for an invite-only GPT-5.5 event and gave them 10x Codex rate limits through June 5. That looks like a promo, but it also shows where competition is hottest — coding. Labs are trying to get developers to build habits around their tools now, then convert that usage into paid subscriptions later. ### So what’s the pattern? Enterprises seem willing to pay when AI is embedded, auditable, and tied to a clear task. They look less interested in open-ended magic and more interested in software that fits existing budgets, workflows, and risk controls. Microsoft’s 20 million paid seats are the scale story. OpenAI’s cyber restrictions and Codex push are the shape-of-demand story. ### What’s the catch? Paid seats are not the same as deep daily usage, and ROI is still uneven across roles. Some workers save real time. Others get a fancier autocomplete box. But the market has clearly moved past the phase where enterprise AI was mostly pilots and press releases. The buying behavior is now visible. ## Bottom line The important shift is not that AI tools got smarter this week. It’s that buyers got more specific. Microsoft is selling AI where work already happens, and OpenAI is packaging advanced capability in narrower, more controlled ways. That is what a real enterprise software market looks like.