Could Microsoft Win The War For Enterprise AI?

As OpenAI and Anthropic talk about going public, everyone is fascinated. These two companies with battling CEOs make up the drama of today’s AI industry. And as more information comes out from their public offering, everyone is wondering how they will compete with each other. (Listen to my latest podcast on culture in AI companies.)

Well as the drama continues, I’d like to advance a different thesis.

While the enterprise market is filled with players (Google, Amazon, Nvidia, Oracle, Anthropic, OpenAI) and the consumer market may be dominated by Apple and others, the biggest winner of all could be Microsoft. Let me walk you through my thinking.

There appear to be three parts to the enterprise market for AI.

First is the model itself, and the challenge to decide which applications it serves

Should coding and analytic apps use Claude? Should narrative and document apps use OpenAI? Should analysis and scientific apps use Gemini? Should robotics and motion apps use Grok? Where will Nvidia’s world model fit?

This product-market fit is still young, as AI labs optimize their algorithms and data training to meet different needs. It’s now quite clear that one model won’t do everything.

We also know that model training is far more than compute: it includes the collecting, labeling, and “fixing” of model data, because one “model” isn’t optimized for everything. So if you’re a pharma company and you want a model that understands proteins and advanced genetics, you want a platform trained in that domain.

Today Anthropic set the pace for code generation, which is the root of many things we ask AI to do.

But will OpenAI really focus on healthcare? Will Google focus on biology? Which model will optimize around the physical world, where robots and manufacturing and transport applications reside? Nvidia and possibly Grok?

Business buyers will need multiple models. So an AI that says they do “everything” is just not credible.

Remember also that these vertical, specialized systems get smarter over time. Our AI Galileo has become amazingly intelligent by staying laser focused on HR, jobs, labor market, skills, and management topics, to the degree it now serves as a management consultant for almost any human capital problem.

Second is the surface, or application experience around the AI.

Which vendor delivers the desktop, toolset, integration, and development tools that make it easy to use? These systems are applications, not models. Questions like how much memory it has, how it personalizes itself, the user experience, and how well it interacts with external data and systems are critical.

This layer of software around the model is also called the AI Harness.

We now care more about the user experience than ever. If Siri from Apple became intelligent and easy to use, a billion people would use it within months. Which “model” it uses is only a small part of the experience.

How did Microsoft win over the PC market? By both copying and licensing the graphic interface and relentlessly focusing on the application experience of Excel, PowerPoint, Outlook, and Windows. Lotus 1-2-3 and Multiplan were far earlier to market than Microsoft, but the eventual “fit and finish” of M365 won out. 450 million paying users have voted.

Business developers and IT have the same issue. We now need more:

• Connectivity and context layers to corporate apps (more below)
• Security and data protection tools and layers
• Analytics, reporting, and monitoring tools
• Features to turn on/off various functions and content by user group
• Many forms of UI customization for customer apps
• e-Commerce features to “sell” AI to customers
• Agent management, monitoring, steering, and maintenance.

So even if you do love Claude or Gemini, if you want to use it in your company you need lots of other “things” to make it a complete solution.

And that gets to the issue of how well the “Surface” or “Harness” interacts with SAP, Oracle, Workday, Salesforce, ServiceNow, Quickbooks, Hubspot, or whatever you already own. And that gets me to this:

Third is the ecosystem.

A business wants an AI platform that has lots of apps, integrations, tools, and third party support.

As we built out Galileo we immediately found customers asking “How do we connect Galileo to XYZ?” – and of course we had to deal with it. “I want Galileo to connect to all our policy databases and our leadership model and compliance training, and on and on.” So we built all sorts of solutions for this, none of which existed in the core platform.

In the enterprise space, where much of the AI “profit” will come, vendors need ecosystems of partners who can make money by building on their platform.

I spend an enormous amount of time with HR and IT leaders and all our discussions follow a similar vein.

“Yes, we want AI tools that are easy, packaged solutions for employees today. But more importantly, we need a platform to build and buy and manage Agentic applications, as we complement and replace the trillions of dollars of systems we have today. And we don’t want to be locked into one vendor when the market is so young and creative.”

So in this era the “engine” is not the issue, it’s all about the “surface.”

The Surface vs. The Model

I’m not sure who started using this word but now we talk about AI “surfaces” not just “models.”

A Surface is the application experience, not the underlying LLM. In other words, the application on top of AI is what matters, not only the AI itself. It’s the combination of the surface and the model creates the experience.

In the corporate world the surface means the tools, speed, user experience, how much history is available, and how the semantic connectivity layer works. If you connect your AI to your HR system or email, you want that connection to bring back valuable data, not just sit there and show a random number.

I tried Claude’s integration with Hubspot. Both companies promoted it so I set it up. I asked our Hubspot system “show me a list of our largest clients and the 10 most recent marketing interactions for each.”

It completely choked. Not only did it fail to find much data, it timed out and failed to complete. The “context layer” which is supposed to make Claude fire the right query for me was just no good. So I dumped it.

This is a problem of the “surface” not “the model.”

How are Anthropic and OpenAI going to do this? They can’t do it alone. They’re hoping that ServiceNow, Microsoft, Accenture, or some other third party does it. And if the company that builds the integration app does a crummy job, their platform looks bad and won’t get used.

Enter Microsoft

Right now, if you looked under the covers of the supposed $30 Billion of revenue (more below) which Anthropic and OpenAI claim, you find that 70% of OpenAI’s revenue comes from consumer subscriptions and 70% of Anthropic’s comes from selling AI compute to other providers.

OpenAI’s $30 Billion could be computed as follows. There are 500-600 million users and if 25% pay $20 per month it totals right around $30 Billion.

Anthropic’s $30 Billion is something like 300-500 huge companies (Meta, Cursor, Microsoft Github, some banks, etc.) generating $500 million or so each, totalling around $30 billion.

So who’s generating revenue from the “surface?” It appears to be Microsoft.

Microsoft claims to have 15 million licensed users of Copilot. If you consider average price of $25 per month, that’s $4.5-5 Billion right there. Now add the fees Microsoft charges for Azure API services, and their AI revenue (growing at 39%) brings them to $25 billion or more.

Microsoft itself projects more than $100 billion in new AI revenue in the coming 3 years and some analysts think it happens faster.

Why Is Microsoft Gaining Share?

A few big things. I just met with Seth Patton, the head of Microsoft Copilot product marketing, and as I explain in this podcast, the “Surface” for Copilot has changed. It’s no longer a “plugin” to each Microsoft product. It’s now an integrated platform that brings together OpenAI models, Anthropic models, and all your corporate information.

How Copilot Has Evolved: From Pieces To A Whole

In the early (2022) days of Copilot we mostly learned about the OpenAI deal and how Microsoft had given up on its internal AI ambitions to license ChatGPT into Bing. That quickly moved to the Microsoft Copilot vision, which initially felt like an intelligent version of Clippy.

Very quickly Microsoft launched Copilot for M365 as well as versions of Copilot within Dynamics, Excel, Github, and other apps. Microsoft’s product teams ambitiously started building more and more “surfaces” on top of ChatGPT.

Then Microsoft launched Copilot Studio, Agent 365, Work IQ, and dozens of other Copilot-fueled “surfaces” to fill out its solution. While all this was cooking along the company built out its M365 Graph Connectors (to bring your Microsoft data into Copilot), Fine Tuning (to optimize your own data into intelligence), and other tools for IT. I was amazed at the speed and pace of new products coming to market, to the point that it felt disjointed in its flurry.

Now, a few years into this, Microsoft has thousands of corporate license deals and they’ve evolved the strategy significantly. Watching customers (and consumers) get confused, Satya Nadella reorganized the Copilot product teams into a single product organization.

Now, with the Copilot product led by one team (instead of Copilots being developed in each product group), Microsoft can build an integrated Surface for corporate vs. consumer and its AI engineering group can focus on their own models. Jacob Andreou, ex-Snap, is leading Copilot growth.

As Microsoft states, “Ryan Roslansky, Perry Clarke, and Charles Lamanna will lead M365 apps and the Copilot platform. Together, Jacob, Ryan, Charles, Perry, and Mustafa make up the Copilot LT and over the next few weeks they’ll work to align the teams.”

Ryan Roslansky leads LinkedIn, Perry Clarke leads Copilot Core, and Charles Lamanna leads Agents and Apps. This leadership team can focus on overall Agent enablement and corporate user value, not just functionality within one app.

While this has been a little tricky to get here, the company can operate more like Nvidia where all engineering layers are integrated around one strategy.

This effectively:

• Elevates Copilot to a horizontal product platform
• Removes fragmentation across business units
• Creates a single P&L–like accountability structure.

I hate to use the word “win,” but I expect Microsoft’s share to grow, and here’s why. (Read financial analysis here.)

First, the corporate market wants an integrated toolset that includes desktop apps, development tools, IT management of agents, and connectivity to legacy systems. While ServiceNow, OKTA, and others play in this ecosystem, Microsoft can build this out with partners. The WorkIQ strategy and the massive effort in Agent365 and Copilot Studio is clear.

Second, the application development world (which is enormous) is waiting for a more integrated set of tools. So every ERP, financial, productivity, analytics, or other vendor now has a set of APIs to build into Copilot-land. It’s not always easy to figure out which API to use (connect to Teams? Graph? WorkIQ? Fabric?), the options are becoming clear.

Third, the end-user world of PC buyers, users, and IT helpdesks can see how all these AI apps could come together into the Microsoft desktop. The current Copilot experience is getting better, and I have to bet the company will put its top UI designers into this area. (Right now it looks a tiny bit “Frankensteinish” with colors and icons, but you can easily see this getting beautified over time.)

Fourth, Microsoft’s partner network will accelerate. Now that the Work IQ APIs are coming out, more and more corporate cloud vendors (who are all worried about being displaced by agents) will look for opportunities to plug in.

Microsoft’s Value-Add

One final important point. Now that Copilot is an open platform, there is major value-add to come.

First is deep research across the MS Graph and more. The button called “Researcher” above (I’ve used it) can look through your MS Graph, for example, and do deep research on your calendar or other details to give you advice, council, or other context help. It’s not as fast as I’d like, but as it’s expanded with memory and context it brings enormous value to individuals and leaders.

Second is the intelligent router. One of the new MS Agents lets you compare a query against models and show you which works better. So before you spend $20 on tokens to run something it will help you optimize where to go, and over time it will decompose your AI task and send different parts to different agents.

Third is agentic interface to Excel, PowerPoint, Word, and other MS apps. In the new Copilot you can interact with a complex document and ask questions, change tables, run reports, create graphs in the Copilot and see the document change. This extends the “in-app” Copilot across Microsoft apps, and it will only get smarter over time.

Fourth is the intelligent context layer in Work IQ. As the API comes out (soon) companies will be able to import and build “context” into Copilot, and that extends Copilot into true Agentic HR, Agentic Finance, Agentic Sales, and more.

As I discuss in the recent podcast and article about semantic layers, this opens the door to in-depth plugins from your corporate apps, which can then be intelligently used or “Agentified” into the Copilot framework.

We now have Galileo (our AI for leadership, management, HR, training, etc.) integrated through the Graph connector and also as a fine-tuned model so all employees can access it. This expanded API lets us add even more use-cases for Galileo, making it a world class management and HR advisor for any employee.

Read this in-depth analysis which compares business models, financials, and strategy between Microsoft, Anthropic, and OpenAI.

Listen to my new podcast explaining this topic.

Get Galileo for AI access to all our research and data.

Join us at Irresistible 2026 for much more.