Microsoft Shows How Not to Introduce AI in Your Business

Since Thursday afternoon I can't use my Outlook. On November 12 Microsoft 365 went down worldwide. Exchange Online, which runs the email, is still down. It's now Friday morning, and Microsoft still hasn't said what caused it. A couple of weeks before that, Microsoft offered up to 60,000 employees money to leave. For more than a year it has also pushed its people to use AI. I'm all in on AI myself. But I think Microsoft shows how not to introduce AI in a business.

For me this is more than annoying. It basically blocks my business, because I don't get my emails. Now and then a few come through. But I'm never sure if I get all of them, or if the emails I send go out. Microsoft now holds delayed emails for 48 hours, so I hope they arrive at some point. I have critical business stuff I need to do today, and I can't do it.

Microsoft posts updates on its status page. At first many Exchange servers failed their health checks and went offline. That put more load on the servers that were left. They restarted servers, but some of them stopped responding again once traffic came back. On Thursday night Microsoft wrote that "our actions taken thus far have not provided the anticipated relief." This morning they say there are early signs that it gets better, but it's not back to normal yet. They will explain the cause later in a report.

The Hacker News thread about the outage has more than 700 comments, and many of them are pretty funny. People joke that Microsoft should ask Copilot to fix it. But Copilot is also down. Someone who says they worked on Exchange at Microsoft until 2024 wrote that the Exchange code "frankly was a behemoth. No one had a full understanding of how it worked E2E."

Microsoft has pushed AI on its own people for a while now. In April 2025 Satya Nadella said that maybe 20 or 30 percent of the code in Microsoft's repos is written by AI. Two months later Julia Liuson wrote to managers that "using AI is no longer optional." She was the head of Microsoft's developer division. She also wrote that AI use should count in performance reviews.

At the same time Microsoft is getting rid of people. In the spring it offered long-time employees money to retire early, and more than 30% of them took it. In July it laid off about 4,800 people. And on October 26 it offered up to 60,000 employees money to leave. That's about a quarter of the company.

I know nothing about what goes on inside Microsoft, and I'm not saying AI caused this outage. But right now the AI haters have a good case, and I find it a bit funny. Microsoft tells everyone they must use AI and lets many experienced people go. Then its email breaks worldwide, and the next morning it still can't say why. From the outside it looks like the people who knew how the system works are gone.

I'm not an AI hater at all. I use AI for all of my software development, and I'm easily ten times as productive, if not more. If you know what you're doing, AI writes very good code. You can write professional enterprise-grade code with it, a lot better than you could without AI and many times faster. For me it's like a miracle.

But this works for developers who want it. They are curious about AI and figure out how it works. Then step by step they change how they build software, and they get great results. I wrote more about this in Why the Same AI Tool Gives Different People Different Results.

In a company it's a lot harder. I don't think it matters much if you have five developers or 50,000. If you tell all of them they must use AI now, many of them will use it wrong. And with AI, if you use it wrong, it can break everything. One commenter on Hacker News wrote that "one bad engineer can now ship a million lines." Checking AI use in performance reviews doesn't fix that, because usage doesn't tell you if the result is any good. And buying everyone an AI license doesn't make a company more productive by itself.

Then there's the incentive problem, which I wrote about in May. Say you are a developer at a big company. With AI you figure out how to do a day of work in one hour. You don't get paid more for that. You probably get more tickets. Or you tell Claude Code to do your ticket and go to the gym. Nobody will notice. If your performance review checks whether you use AI, you even look good. So the easy way for an employee is to let the AI write slop and not care about it. I think that's where most of the AI slop in companies comes from.

If you run a business, you should start with the people who want to use AI. When they get more done with it, they should also get something out of it. And you should make sure the people who know your systems are still there while you change how your software gets built. I wrote down how to go about it in my practical guide to using AI in your business.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com