One move by the current US administration that has gained a lot of media space this month was the firing of the leader of the Bureau of Labor Statistics, the main agency responsible for producing figures on the job market. The BLS commissioner was accused of “faking” the latest employment figures for “political purposes”.

It feels dramatic when governments do something like this. Just deciding to reject the reality and substitute bad figures with ones that better suit their narrative. We tend to rightfully expect more from democratic nations. Yet here we are now in 2025.

Companies, on the other hand, are almost expected to be non-transparent with where their figures come from. Aside from financial reporting regulations, there aren’t many guarantees that what is reported to the media, to customers, partners, and other interest groups, would be true. This means that we should approach big numbers used by tech vendors to be specifically designed to convey a particular message, as opposed to being something objective and neutral.

Outright lies may not be too popular, given how there’s a chance that some nasty person within the organization might leak the information and cause a PR crisis while exposing the truth. Instead, a safer approach is to distort the reality just enough that the numbers themselves have a source to point to. How the numbers are then interpreted, that’s the creative playground for corporate marketing personnel.

In this post I will explore some of the ways Microsoft has been using numbers to build stories that fit their agenda. I don’t intend to act in the way how certain leaders of nation states do and claim they’d be “fake news”. I simply want to draw some attention to how the numbers don’t necessarily mean what people think they do.

This popular meme template tells all you need to know about why Microsoft renamed “copilots” to “agents” in September 2024:

Let’s not go into who it was that decided to rename the Power Virtual Agents product to Copilot Studio in the end of 2023. I bet that person’s family is also rich. But anyway…

Unlike OpenAI, which held back from calling their custom GPTs “agents” and waited until it was ready to launch ChatGPT Agents last month, Microsoft decided to go all-in on stretching the agent concept. In the previous FY25 Q3 earnings call, MS said: “this quarter alone, customers created over 1 million custom agents across SharePoint and Copilot Studio, up 130% quarter-over-quarter.”

Now, in the final quarter of their fiscal year, the updated stats were that “this year, customers had created 3 million agents using SharePoint and Copilot studio”. Sounds like the kind of exciting growth that invites people to share this big number on their social media posts. Like Jeff Teper, the president of M365 collaborative apps and platforms:

Notice how the order that Jeff chose is “3 million agents in 1) Copilot Studio and 2) SharePoint”? That’s reversed compared to what Microsoft’s FY25 Q4 earnings call used. I guess in that forum MS wasn’t comfortable in suggesting that the number of Copilot Studio agents would be higher. Whereas on social media there’s a recurring theme already from a couple months back that folks assume these numbers primarily reflect Copilot Studio adoption.

For anyone that hasn’t worked hands-on with these tools, and instead is just reading the news about it, the concept of an “agent” may still sound like something complex and fancy. Whereas if you’ve ever looked at what SharePoint agents technically are, you’d realize nothing could be further from the truth. To save you the trouble, I recorded a short video of how it takes 10 seconds to create a new SharePoint agent. And that it literally takes one click to get to “your agent is ready”.

So, out of the 300+ million global users of SharePoint Online, enough of them clicked on a prominent menu to trigger the creation of a new .agent file on a SharePoint site ~2M times in 2 months. Since there are little governance capabilities available for controlling agent creation, let alone allowing users to delete accidentally created agents they no longer want - we’ll have many, many millions more of these agents to celebrate in the coming quarterly earnings calls.

Note that I’m not criticizing the proper Copilot Studio agent creation tools and capabilities here. You can surely build very advanced Copilot agents that talk with business data via Power Platform connectors and apply intricate logic specific to the unique business processes of an organization. Many BizApps pros are working on trying to build those for customers who are excited about the opportunities of GenAI.

It is, however, an outright travesty to put those solutions under the same category as the glorified intranet search UIs of what SharePoint agents onboarding experience invites users to create. It’s like saying that a demo canvas app built on static Excel data and a D365 FinOps ERP system are just “apps”. Sure, you could do it. No, the combined figure will not tell anything about the truth behind technology adoption and use cases.

Publicly listed Big Tech companies are giving detailed financial figures in a structure regulated by the authorities. There must be a million ways to massage these quarterly financial numbers to make them look pleasing to the current and potential investors. Still, at least they are regularly disclosed.

When it comes to the products and services that deliver the billions behind these results, there’s no need to publish any stats that don’t support the company’s marketing narrative. Such is the case now with Microsoft deciding to stop reporting its AI revenue. No new figures have been disclosed after January 2025.

We therefore have to rely on estimates distributed in the media to understand what the growth rate might be. The projections published by The Information put Microsoft’s AI revenue for 2025 at $13B. Now, the really important bit about this is how it was also split into revenue from OpenAI vs. other AI sales. Like this:

$13B is a fair chunk of money, sure. But how it is used in communicating the growth and business potential of AI is misleading. That’s because when you spend one dollar on a service like ChatGPT Plus, it is also booked as Microsoft AI revenue. Because they provide the computing resources for a company that they happen to be a significant owner of (and being entitled to 49% of OpenAI’s profits - if they ever make a profit).

It’s not fraud. There is, however, “a structural similarity to the Enron-style illusion of inflated financial health”. At least that’s what ChatGPT o3 said when I asked its take on this arrangement. Together, MS and OpenAI can show bigger figures and convince their investors that the AI market is growing as they’ve been promising. Furthermore, it allows two parties to attach themselves to share the perceived success of both. There’s less pressure to talk about people choosing ChatGPT over Copilot when you’re “winning together”.

Some might say this is just the regular supply chain mechanism at play. After all, MS is providing infrastructure to OpenAI and they are then running their models and software on top of it. But the issue is in how this is communicated and lumped into the same AI bucket. If treated separately, it is way too easy for analysts to mistakenly assume that an AI dollar reported by OpenAI and an AI dollar reported by Microsoft equal two dollars. But that’s not how it’s supposed to work.

If Ford is buying steel to manufacture cars, then the car market revenue figure should only contain what Ford gets from the end customer who buys the car. The steel company’s operations do not participate in expanding the car market. People aren’t spending more in car purchases, just because suppliers further in the chain are also getting money. Yet if we’d talk with vague terms like “AI revenue” and not specify what exactly counts as AI, mistakes could easily happen.

So, today Microsoft is no longer publishing such figures. This time around, as part of FY25 Q4 earnings report, they decided to provide an actual Azure revenue number - for the first time ever. From a total FY25 revenue of $282B across MS, Azure brought in $168B. Nearly half of it. Looking at the quarterly cloud revenue growth, it’s easy to understand why this felt like the right moment to disclose it:

One can speculate whether the previously used AI revenue metrics would have produced as attractive charts to present. In all fairness, for a company like MS that operates so broadly across the spectrum of computing, defining the category for what is & isn’t AI these days would be challenging. Unless the specific individual workload types were disclosed, it would be yet another potential source of confusion.

From a narrative perspective, it’s now easy to point at the total Azure revenue and imply that “yeah, we’re seeing pretty amazing demand for AI services (wink, wink)”. The media will likely jump on the numbers given and project that into a news story that tells exactly what MS would hope for.

Luckily, not all reporting is done using only vendor provided information as the source material. There are also less exciting reasons for why the cloud computing demand is growing. Such as the massive usage of Azure resources for facilitating the surveillance operations carried out by Unit 8200 of Israel’s military:

Yeah, Azure revenue is coming in alright. It’s the (human) cost of that revenue that should be paid more attention to…

For smaller units within Microsoft, it may not be feasible to bundle revenue sources together to keep the numbers growing. In such cases we may see strategies that are common to other SaaS vendors across the industry.

GitHub CEO recently posted about GitHub Copilot having reached 20 million users, up more than 5 million from last quarter. Again, on a graph that looks like amazing AI momentum:

If you were to be presented with only that figure and graphics, what would your first interpretation of it be? Especially when followed by the statement “we’re also seeing explosive AI growth on GitHub, with AI projects more than doubling over the past year.” When leading with a big number like 20M, the audience may well be conditioned to assume that whatever follows is related to that number.

It startled me that even software engineers who are really, truly deep in analyzing the practice of software development may fall victim to the numbers game. “I’m assuming this number means the number of accounts that are paid for this plugin.” This comment made me realize how people will jump to conclusions about numbers that are presented in support of the marketing narrative.

It just so happens that at the end of December 2024, GitHub announced a free version of Copilot for all VS Code users. Gee, might that have something to do with the jump we see for 2025 figures? Even I, a person who can’t write a line of code on his own, have activated and tested GitHub Copilot as part of my VS Code adventures when working with static website generators and such.

Launching a free tier is a common growth hack. Lowering the barrier for users to try out the product is not cheating, of course. It’s a logical way to broaden the possibilities for upselling the premium tier. What that does generate for marketing purposes, though, is social proof that can be drawn from a big spike in total user numbers. That’s a highly valuable currency in convincing that you are succeeding and that they should join the game to enjoy this aura of success.

We’ve seen the free tier strategy applied in mainstream MS products, too. Introducing the concept of Microsoft 365 Copilot Chat to replace Bing Chat Enterprise (and whatever it was called before) made it instantly possible to refer to much higher numbers when talking about M365 Copilot usage.

As a growing number of Microsoft customers are facing the moment for deciding whether to renew their M365 Copilot licenses based on initial experiments, they’ll be looking at whatever stats available on the results so far. MS is of course doing its best to make sure that a “data driven” decision can be made, and that the decision is to buy more seats for M365 Copilot.

The recent announcement of new prompt categories in the Microsoft Copilot Dashboard made me stop and think “but why”? As a part of Viva Insights updates, Copilot Chat will now show prompt categories Ask and find, Catch up, Draft and brainstorm, and Other:

Vendors like to see active usage go up, but this is a dashboard for the customer organizations. Is a higher prompt count better? Let's say the user is smart at prompting and M365 Copilot is great at responding. That means less chatting is needed with AI. If, on the other hand, they spend a lot of time going back & forth, is that a positive sign of adoption, let alone business value?

If the number goes up, people will naturally tend to frame it as a good thing. Well, would the number of emails grouped in such categories tell something meaningful? "Yay, our employees are sending more emails about Ask & Find now compared to the previous quarter!" Okay, great, but conclusions or actions should be made on this? Depending on whether you’re arguing for or against the tool, the same figure can justify opposite conclusions.

It doesn’t even matter if the underlying mechanism used for categorizing these prompts is an LLM black box that no customer will be able to drill deeper into. The numbers that are generated into Viva Insights can now be used for “data driven insights” that serve as the justification for various claims. Such as, A) how M365 Copilot is already delivering value, and B) why it is essential that more effort is put into upskilling users to better leverage M365 Copilot in their work. Oh, and of course upselling M365 Copilot Chat “free” users to take a more powerful M365 Copilot premium subscription into use to improve value per prompt.

This whole playbook has been around for a while. The numbers game was not invented for the AI era; it has simply been pushed to overdrive due to the immense market pressure. Looking at previous hype technologies, like low-code, I can relate to a lot of what the current Copilot advocates are saying. Because the same methods were present also for marketing Power Platform earlier.

First of all, bundling the emerging new category of citizen developer tools together with an established corporate function like BI was a brilliant move from Microsoft. By grouping Power BI under the Power Platform umbrella, it gave an instant boost to the total number of active users MS could promote for this low-code application platform. How many of the users were merely building reports or using template flows for email notifications rather than building actual Power Apps? What share of Fortune 500 were customers of Power BI primarily?

It’s not like these numbers are meant to be exactly correct even when presented to public audiences. For example, in July 2022 Satya still said in his Inspire keynote that Power Platform had crossed 20 million monthly active users. Then, only a week later, in the FY22 Q4 earnings call, he said “we now have nearly 25 million monthly active users”. Wow, talk about explosive growth!

Here’s me, 3 years ago explaining the difference on LinkedIn. Just to prove that even when evangelizing such tools due to having co-founded a company “100% focused on Power Platform” 2 years earlier - I still tried to be transparent about the numbers I shared. Now, we did do both apps and BI, so our tagline was justified. But when I was juggling marketing tasks alongside customer projects, my personal focus was 90% on the app side.

Inventing new concepts when creating a new kind of product offering is both a necessity as well as a marketing opportunity. In 2022, Charles Lamanna (then head of MS low-code platforms, before graduating to “Business and Industry Copilot”) said:

Try searching for “citizen professional developers” on the web today and you’ll not get many results back. That’s of course because it isn’t a category anyone started using, even though there were millions of such people out there three years ago already. I can understand how challenging it’s been to find a term that would describe the many types of Makers that create software solutions without being traditional software developers. Yet comparing pro-devs and citizen pro-devs with these kinds of big numbers primarily serves in generating one-off marketing content rather than measuring industry trends in a meaningful way.

Power Apps & Automate were built on the growth hack of seeded Office 365 licensing right from the start, of course. Then, as COVID created a boom in Microsoft Teams usage, another growth hack was launching the “free” Dataverse for Teams edition. Thanks to the UX patterns applied, many users ended up accidentally provisioning new Dataverse databases in the tenant when trying out Power Apps in Teams. Sure, it looked like explosive growth in Power Platform environment count - but it had little to do with actual low-code adoption in organizations.

So, today when the number of “agents” is skyrocketing inside M365 tenants, it’s easy for me to understand what the driving forces behind those figures are. Yes, people are also exploring how they could take AI into better use and gain benefits from Microsoft 365 Copilot that on its own feels like an AI assistant without hands. Are these agents actually taken into regular use? Do they deliver more value than what they consume in time and attention from the workers? That’s not for these numbers to cover. They have been designed for a different purpose altogether.