Microsoft has sold more than 30 million paid Copilot seats and will not tell you how many people use Copilot every month. Those two facts, held together, are the whole story of Microsoft Copilot statistics in 2026.
The seat number is disclosed on earnings calls because it is a revenue metric and investors want it. The usage number is given only as a growth multiple, never as an absolute, and Microsoft has declined to provide one when asked. That asymmetry is not an accident and it is the first thing to understand before quoting any figure below.
Every number here is attributed to a dated Microsoft disclosure or to reporting on one. Where Microsoft gives a multiple instead of a count, that is stated rather than converted into an estimate.
Paid Seats Doubled Past 30 Million in One Year
On 29 July 2026, reporting Q4 FY2026, Microsoft said Microsoft 365 Copilot paid seats “exceeded 30 million,” with net additions more than doubling sequentially.
Six months earlier, on the Q2 FY2026 call of 28 January 2026, the figure was 15 million, described as up more than 160% year over year. Doubling a paid seat base in two quarters is a genuinely strong commercial result and it is the figure Microsoft most wants quoted.
The enterprise shape underneath it is more interesting than the headline. The number of customers with over 50,000 seats rose more than sevenfold year over year. Microsoft said hundreds of enterprise customers bought millions of seats of its new E7 suite within two months of launch, and that EY deployed E7 to 400,000 employees. Large deployments are concentrating, which means the seat count is being driven by a relatively small number of very large decisions rather than by broad self-serve adoption.
Microsoft 365 Copilot launched in late 2023 as a $30 per user per month add-on.
Be careful what a seat means before you repeat the number. It is a licence a company has purchased and assigned, not a person who used anything. The sequential doubling of net additions is a real acceleration in purchasing, and it is also what happens when a vendor bundles an assistant into a suite upgrade that enterprises were buying anyway. E7 is exactly that mechanism: hundreds of customers buying millions of seats inside two months is a suite decision, and the Copilot entitlement travels with it.
That is not a criticism. Bundling is how Microsoft has won most of its markets. It does mean the seat figure measures contract value more faithfully than it measures adoption, and the two are separate questions a board should ask separately.
The 3.3% Number Is the Most Important One
In February 2026, The Register reported that only 3.3% of Microsoft 365 and Office 365 users who touch Copilot Chat actually pay for it.
Read the denominator carefully, because almost everybody who repeats this figure gets it wrong. It is not 3.3% of all Microsoft 365 users. It is 3.3% of those who touch Copilot Chat, many of whom get Copilot Chat at no additional cost as part of their existing subscription. The article gives no date for the figure itself.
Set it against the base. Analyst Mary Jo Foley, cited in the same piece, put the Microsoft 365 commercial user population at some 450 million. Thirty million paid Copilot seats against roughly 450 million commercial seats is under 7% penetration of the installed base, two and a half years after launch, from a vendor with unmatched distribution advantages and an aggressive enterprise sales motion.
That is the honest framing of Copilot adoption in 2026: a large and fast-growing absolute business, and a small fraction of the available population. Both halves are true and most coverage picks one.
The distinction between Copilot Chat and paid Copilot is worth getting straight, because it is the reason the two sides of the story can both be accurate. Copilot Chat is the no-additional-cost tier available to a large share of Microsoft 365 and Office 365 users. Microsoft 365 Copilot is the paid add-on that puts the assistant inside Word, Excel, Outlook and Teams with access to a company’s own content. The 3.3% figure measures the conversion between those two things.
Read that way, it is a product funnel statistic rather than an indictment. A 3.3% free-to-paid conversion on a product priced at $30 per user per month is not obviously bad for a business of this scale. It is only damning if you had been told that everyone who touched Copilot loved it.
Why Does Microsoft Not Publish a User Count?
Because the growth multiples read better than the absolute numbers would.
On the January 2026 call, Satya Nadella said daily active users of Microsoft 365 Copilot were up tenfold year over year, and average conversations per user had roughly doubled. For consumer Copilot he said daily users grew nearly threefold year over year across AI chats, the news feed, search, browsing, shopping and operating system integrations. No absolute figure accompanied any of those, and Microsoft declined to provide one.
The only absolute user figures in circulation are older and broader. The prior annual report said Copilot had passed 100 million monthly active users, counting commercial and consumer together, and a Microsoft spokesperson later said the total had grown to 150 million, again combining both groups. Note what that means: a figure spanning every Copilot surface, including ones bundled into Windows and Bing, is 150 million, while Microsoft 365 alone has around 450 million commercial seats.
TechCrunch’s coverage of the January call called some of these numbers squishy, which is fair. A tenfold increase in daily users from an undisclosed base is unfalsifiable. It could represent a move from a very small number to a small number.
GitHub Copilot Is the Clearest Signal in the Set
The developer product is the one place where Microsoft publishes both a paid count and a free count, which makes it the only Copilot surface where you can actually compute a conversion rate.
As of 28 January 2026, GitHub Copilot had 4.7 million paid subscribers, up 75% year over year, against 20 million total users including the free tier per the prior annual report. That is roughly 23% paid conversion, and it is a healthy software business by any standard.
In June 2026 GitHub moved Copilot to usage-based pricing, which Microsoft said pushed quarterly revenue acceleration above 60%. It did not restate a subscriber count after the change, which is what happens when a metric stops being the one that flatters you.
The reason this matters as a reference point: 23% of developers who try a coding assistant pay for it, against 3.3% of office workers who touch a chat assistant. Developers have a measurable output problem that the tool measurably helps with. That gap is the real finding in the Copilot numbers, and it is a lesson about where AI assistance currently earns money rather than a point about Microsoft.
The pricing change deserves one more line, because it signals something about where Microsoft thinks the value sits. Moving from a flat subscription to usage-based billing is a bet that heavy users will pay more than $19 or $39 a month, and that the metered model will capture value from agentic workflows that run without a person watching. The reported revenue acceleration above 60% suggests the bet is paying. It also means future GitHub Copilot figures will be revenue figures rather than subscriber figures, so the one Copilot surface where you could compute a conversion rate has just stopped publishing the inputs.
Dragon Copilot Shows What Real Adoption Looks Like
The healthcare product is the quiet proof case and it almost never gets quoted.
Dragon Copilot was available to 100,000 medical providers and documented 21 million patient encounters in the January 2026 quarter, up threefold year over year. That is not a seat count or a growth multiple. It is a count of completed units of work, in a setting where the task being automated, clinical documentation, is universally disliked and consumes hours a day.
Twenty-one million encounters in a quarter from 100,000 providers works out to roughly 70 encounters per provider per quarter. These are people using it as the way they do their job, not people who tried it once.
The structural reason it works is worth extracting, because it predicts where the next real adoption will show up. Clinical documentation is a task with a defined output, a captive audience who hate doing it, a measurable time cost, and a quality bar that a machine can clear. Drafting an email has none of those properties except the last one. When you see an enterprise AI number that counts completed units of work rather than licences, go and look at whether the underlying task has that shape. If it does, the number is probably real.
The Seat-to-Habit Ratio
Here is the lens worth keeping for any enterprise AI number, Microsoft’s or anyone’s. Call it the Seat-to-Habit Ratio.
A seat is a procurement decision made by somebody who is not the user. A habit is a behaviour of the user. The two are measured by different departments and they diverge, and vendors disclose whichever one is larger.
Microsoft’s disclosures map cleanly onto this. Seats: 30 million, doubling, over 50,000-seat customers up sevenfold, EY at 400,000. Habits: a tenfold daily active multiple with no base, a doubled conversations-per-user figure with no base, and a leaked 3.3% payment rate among people who touched the product. Where Microsoft has a genuine habit to report, as with Dragon Copilot’s 21 million encounters, it reports the absolute number.
Run the Microsoft Copilot statistics above through that lens and the disclosure pattern becomes legible. So the test, applied to any AI adoption claim: does the figure count a purchase or a use? If it counts a purchase, it tells you about a sales team. If it counts a use, it tells you about a product. If the vendor gives you a multiple with no denominator, it is telling you the denominator is unflattering.
Capital Spending Is the Number Nobody Quotes
Scale the adoption numbers against what they cost to produce and the picture sharpens again.
The Register described a $37.5 billion quarterly AI spend for the quarter reported in January 2026. Microsoft’s capital expenditures ran $88.2 billion in the prior fiscal year and $72.4 billion in the first part of the current one. Azure grew 43% in constant currency in Q4 FY2026 with guidance of roughly 45% for the following quarter, and Azure passed $100 billion in annual revenue.
Nadella and finance chief Amy Hood have both said AI demand exceeds data centre supply, so new equipment is effectively booked to capacity for its usable life. Capacity constraint is a demand signal, and it is also the reason seat growth and usage growth can both be real while margins stay under pressure.
Analyst reaction to that quarter was mixed rather than triumphal. UBS’s Karl Keirstead said both Azure and Microsoft 365 fell a bit short, calling it the key negative investors were hearing, while still recommending the stock.
For anyone quoting these figures, the ratio that matters is not spend to revenue, which nobody outside Microsoft can compute per product, but spend to disclosure. A company spending tens of billions a quarter on AI infrastructure and declining to publish an absolute user count for its flagship AI product has made a choice about what it wants discussed. That choice is itself information, and it is more useful than any of the multiples. When the absolute number becomes flattering, it will be disclosed.
What Any of This Means for Being Recommended
If your interest in Microsoft Copilot statistics is whether Copilot matters as a surface where your brand can be named, the seat count is the wrong number and the 3.3% is closer to the right one.
A Copilot seat is a licence attached to an employee. It becomes a recommendation surface only when that employee asks it a question whose answer involves choosing a vendor, a product or a service. The disclosed habit metrics suggest that is happening at a rate Microsoft prefers to express as a multiple. The 150 million combined monthly figure across every Copilot surface is the realistic outer bound of the audience, and a meaningful share of that is Windows and Bing integration rather than deliberate use.
Compare that with Google’s disclosures for the same period, where AI Mode inside Search passed a billion monthly active users and the Gemini app reached 950 million. If you are allocating effort across AI visibility surfaces, the orders of magnitude are not close, and the allocation should not be either.
That does not make Copilot irrelevant. Thirty million paid enterprise seats are 30 million knowledge workers whose default research tool now sits inside the document they are writing, and the queries that happen there are high-intent business queries rather than idle ones. The material it draws on is substantially the same indexed web that every other assistant reads, which is why the practical work of being citable does not fork per assistant.
That last point is the one worth holding onto, because the market is full of services priced per assistant. There is no separate Copilot index to get into. Microsoft’s assistant grounds answers in web search results and in the customer’s own tenant content, which means the external half of the job is the same work that makes you visible in AI Mode, in ChatGPT’s browsing, and in classic search: be described accurately, on indexed pages, by sources other than yourself. Of the thousand-plus outlets in the Instant Press inventory, about 96% are Google-indexed, and indexation is the precondition for being readable by any of these systems. Buying a separate programme per assistant is paying several times for one piece of work.
So before you quote any of these numbers in a board deck, the question worth answering is the one Microsoft has spent three quarters not answering: of the seats your company has bought, how many people opened it yesterday, and do you actually know?