In our last issue, I showed you why the next phase of the AI boom could move artificial intelligence out of giant data centers and into individual companies.
But simply bringing AI inside a business doesn’t necessarily make it useful.
For one thing, the AI still has to learn how that business actually works.
Microsoft just showed us how that will happen.
Teaching AI How Work Gets Done
Recently, Microsoft CEO Satya Nadella shared a lengthy post on X outlining the company’s broader AI strategy and linking to a new paper on its MAI family of models.
Microsoft says one of its specialized models can now handle many common Excel tasks about as well as GPT-5.6. Another has improved the quality of code suggestions inside GitHub Copilot.
Those are nice improvements. But I think most people paid attention to the wrong part of the announcement.
You see, what Nadella revealed is that Microsoft isn’t just building smaller AI models.
It’s changing the way AI products are built.
What this means is that the model answering questions inside Excel today doesn’t necessarily have to be the one answering it tomorrow. Microsoft could replace GPT with one of its own MAI models, route a particularly difficult problem to Anthropic or eventually even use a model that hasn’t even been invented yet.
Just as long as the rest of the system stays the same, most customers will never notice the difference.
To me, that’s the real breakthrough here.
Microsoft isn’t treating the AI model as the product anymore. It’s treating it as a replaceable component inside a much larger system.
A company’s proprietary data, software tools, workflows and even the feedback employees provide all remain outside the model, allowing Microsoft to improve the product without rebuilding it every time a better model comes along.
Microsoft refers to this surrounding structure as the AI’s “harness.”
And you can think about it this way.
You’ve just hired a new accountant. They already understand tax law, but they still have to learn how your company works. They need to know where files are stored, who signs off on returns, which clients have unusual requests and how work moves from one department to another.
An AI model has the same problem.
A general-purpose AI model might understand the U.S. tax code, know how to read a balance sheet and recognize thousands of accounting terms. But that doesn’t mean it understands how a particular accounting firm operates.
That’s the knowledge most companies don’t want to leave their own networks.
And that’s what makes Sovereign AI so compelling.
Not only does it allow your data to stay under your own roof, it enables everything an AI learns about your business to stay under your control.
Microsoft is developing what it calls “reinforcement learning environments” to capture that knowledge.
Instead of only training an AI on books, websites and other general information, Microsoft creates a practice environment based on the work the model will actually perform.
Every time the model takes an action, it receives feedback on whether it succeeded. That way, it gradually learns business processes much the same way a new employee learns.
By doing the job instead of just reading a manual.
Microsoft has already used this approach inside Excel. It started with a model built for software coding and then retrained it inside an Excel environment, where it learned how to use spreadsheet tools and complete common knowledge-work tasks.
And that’s not the only example.
At Microsoft’s Build conference, the company said a model tuned for McKinsey’s internal standards achieved the highest win rate of any model the consulting firm tested, beating GPT-5.5 on quality at roughly one-tenth the cost.
Microsoft also says its Excel model can be up to 10X more efficient than the larger general-purpose models it was designed to replace.
That gets to the heart of Microsoft’s strategy.
The most advanced frontier models are incredibly powerful. But businesses don’t always need the world’s smartest AI.
What they need is an AI that understands their business.
Using one of the world’s biggest models for every task would be like hiring a world-famous surgeon to put a Band-Aid on your finger.
They could certainly do the job. But there are much cheaper ways to get the same result.
Here’s My Take
The way AI is evolving, I’m not sure that any single model needs to ultimately win the AI race.
After all, Microsoft didn’t have to build the personal computer to become one of the biggest winners of the PC revolution. It became a tech powerhouse by building the operating system that allowed millions of different computers to run the same software.
Apple took a similar approach with the iPhone. It doesn’t build every app or provide every service. It built the ecosystem that connects all those pieces together.
I believe something similar could happen with AI.
If companies can swap one model for another without losing their data, workflows or everything their AI has learned, then the model itself becomes much less important.
And that could dramatically change where value is created in the AI economy.
Microsoft isn’t the only company preparing for this possibility.
Anthropic has developed a new technology that could make switching between AI models much easier, and it’s already spreading across the industry.
I’ll show you how it works in Friday’s issue.
Regards,
Ian King
Chief Strategist, Banyan Hill Publishing
Editor’s Note: We’d love to hear from you!
If you want to share your thoughts or suggestions about the Daily Disruptor, or if there are any specific topics you’d like us to cover, just send an email to dailydisruptor@banyanhill.com.
Don’t worry, we won’t reveal your full name in the event we publish a response. So feel free to comment away!