Only 3.3 percent of Microsoft 365 and Office 365 users who touch Copilot Chat actually pay for it, an awkward figure that landed alongside Microsoft’s $37.5 billion quarterly AI splurge and its insistence that the payoff is coming.
That single percentage stat undermines the company’s carefully polished Copilot success story. On its Q2 FY26 earnings call, Microsoft repeatedly cited “record” AI momentum, telling investors it now has 15 million paid Microsoft 365 Copilot seats, with seat growth up more than 160 percent year-over-year. Satya Nadella described Copilot as “becoming a true daily habit,” claiming daily active users are up tenfold year-over-year and that average conversations per user have doubled.
What Microsoft did not articulate is how small that paid footprint looks against the vast base of Microsoft 365 users experimenting with Copilot Chat for free, as highlighted by Directions on Microsoft analyst Mary Jo Foley.
OK, lets set the stage.
I have access to ‘free’ co-pilot for home
I have access to the basic co-pilot through O365 at work
I have had access to basic Gemini through work
I have access to full-fat private ChatGPT through work
I have access to private Claude Pro through work
I have been handed edicts to find useful places for AI, prove time saving and identify issues where it’s not a good idea.
Today I ask co-pilot, in my county, what are the regulations about shoveling sidewalks on your property.
Co-Pilot: There are no regulations about that, but if you go to a nearby city there are regs and they are this.
Gemini: There are no regulations about that in your county
GPT: yes there are regs, they are here
Claude: yes there are regs, they are here, historically, they have also said this and that and here are some exclusions
Today was a good day, at least the more expensive ones had an answer.
If I ask the same question to GPT or Claude the same thing I ask to Co-Pilot, right or wrong, I’ll get a well (for a lack of better term) “thought-out” answer, It’ll give me concepts of why it answered that way and generally some citations where I can see how it came to those conclusions. Yes, in the end it’s just autocomplete, but it’s capable of giving me the context it used to complete the response. If it’s wrong, it’s easy to explain where it went off the rails and get it to change tracks and usually get a good answer.
I use Copilot for personal use because it’s free, relatively feature rich, and generally gives me faster results than whatever in the hell search engines have become. But it’s far worse than any of the other options. I think behind the scenes, they’re running models that have been pruned to hell. Where as the other guys (admittedly expensive plans) are not only giving me higher quality responses, but they run many additional queries while looking things up, is some cases you can watch the “thinking model” find out it was wrong and then backtrack and look it all up again.
There are heaps of things that it’s bad it, there are loads of things that should never be asked. But if Microsoft is intending to have people get even remotely interested in copilot, that have to make it not objectively worthless to use.