I Don't Know What Satya Nadella Is Talking About

I have over 20 years of tech experience, and after hearing Microsoft's CEO speak recently, I'm genuinely wondering if he's been replaced by an AI avatar. Because he said a lot of words without imparting any actual meaning.

Here's the quote from the video that's been making the rounds:

“The approach at least we’re taking is, I think, the notion that business applications exist, that’s probably where they’ll all collapse, right in the agent era, because if you think about it, right, they are essentially CRUD databases with a bunch of business logic. The business logic is all going to these agents, and these agents are going to be multi-repo CRUD, right? So they’re not going to discriminate between what the back end is. They’re going to update multiple databases, and all the logic will be in the AI tier, so to speak. And once the AI tier becomes the place where all the logic is, then people will start replacing the back ends, right?”

First, for my non-technical readers, CRUD is an acronym that stands for Create, Retrieve, Update, Delete. And yes, most software applications have core functionality that allows them to save and load data.

But beyond that, what the hell is he even talking about?

“CRUD Plus Business Logic?”

This is a gross oversimplification of the software that powers our world. Yes, in a way, vehicles are just engines with wheels attached. A symphony is really just air vibrations at different frequencies. Simple right? And when you’re talking to an audience of non-technical investors and executives, that’s the kind of generic statement that is easy to understand, so it must be true.

The "business logic" he's so casually dismissing is literally the entire value proposition of these applications. It's decades of domain expertise, user interface design, performance optimization, and handling edge cases.

There is always a tension in business, the “build vs buy” decision.

When you buy off-the-shelf software, you're aligning with industry standards. This reduces direct costs for the software itself, but there are additional, rarely measured impacts on subsequent expenses, including finding workers who already know your tools, training time, integration with other standard systems, community support, and established best practices. You benefit from the collective wisdom and shared development costs of thousands of companies.

However, when you have a specific way of doing things and can utilize technology as a competitive advantage, then building starts to make more sense. Your unique business processes, proprietary algorithms, and specific workflow optimizations become differentiators that generic software can't provide.

“But", the SASS vendor will say, “you can customize Salesforce and other products!” Yes, you can, and then you can pay to be on the upgrade treadmill, being at the mercy of the provider not to make any changes that break your customizations.

I often advise my clients to adapt their processes to the software tools they purchase, not the other way around.

This is one reason why I'm not panicked about software developer jobs. When software becomes cheaper to produce, it has historically led to increased demand for IT talent. Cheaper production means more experimentation, more custom solutions, and more businesses willing to invest in technology they previously couldn't afford. This leads to an increase in demand for software, which in turn drives up demand for IT professionals.

Yes, some skills move downstream or get eliminated. When I entered the field around the dot-com crash, I was earning premium IT wages to create basic websites. Then, WordPress and other tools emerged, and my skills shifted to more complex, back-end systems. Next, we invented the SPA app and frameworks like Angular and React, and suddenly, I was coding in the front-end again.

The “AI Tier” Fantasy

And then there's this gem: "all the logic will be in the AI tier" and "once the AI tier becomes the place where all the logic is, then people will start replacing the back ends."

So let me get this straight. We're going to take all the carefully crafted, tested, and debugged business logic from our applications and hand it over to large language models that are famously inconsistent, can't do basic math reliably, and occasionally hallucinate entire datasets.

What could possibly go wrong?

Instead of deterministic business rules that behave the same way every time, you get probabilistic responses that might work differently on Tuesday than they did on Monday, depending on how the model is feeling.

In addition, research scientists already struggle with whether these tools can actually reason and how they arrive at the “decisions” they make. There are also examples of LLMs outright lying when challenged, depending on their incentive prompts. As a business, how the hell are you going to audit this?

We’ve seen recent reports of LLM tools dropping production databases and prompt injection attacks compromising data. Putting an LLM in charge of your business logic tier is one of the most irresponsible things I’ve ever heard uttered.

Ok, But Let’s Assume this is Possible

Alright, let’s give him the benefit of the doubt. What does this look like from a real-world design-build-deploy scenario?

We’d probably start with the data. So, you start talking to your LLM about the type of data you want to store, what rules apply to the data, such as whether this field can be empty, whether it accepts negative values, and so forth. And so, with a wave of your hand, you empower the AI agent and it…

Creates a database.

Whoops, we just added software. But apps are supposed to collapse! So, is the agent creating its database from scratch? If it is, how does it handle security? Does it encrypt the data? Can it segment off who should and shouldn’t have access to particular bits of data? If there is no app, how do you audit this? What about disaster recovery? Is it being backed up? Can you restore it? How do you restore it when LLMs are famously terrible at temporal inference?

But whatever, let’s move on. Now you want to see your data. Maybe some nice charts and graphs, table formats, sorting, filtering, and aggregates. So, we ask our trusty Clippy AI and it… creates Tableau? Again, we have no auditability, and we have no deterministic code to review. Is the data correct? Who knows?! Can I share the data with partner organizations? I don’t know, maybe I should ask the AI to invent APIs. What format is the data in? After all, my AI needs to tell my partner’s AI what’s coming in and how to process it.

Wow, that sounds complicated. Maybe I’ll instruct the AI to send the data directly. Oh, I guess the tool should invent documents and email, maybe even file sharing?

And across my company? Does everyone get their agent? Will we have data silos everywhere without standardized formats? How can we consolidate all this data in a way that enables us to make informed, strategic decisions about our company? I suppose we will ask an AI agent to invent the data warehouse?

Again, what…the…fuck…is he even talking about?

You haven’t eliminated applications. But, you have made them shittier, slower, and more expensive to run by forcing everything through an LLM interface.

I’m reasonably certain that the shared dream of all these AI vendors is to ensure that no one can do anything at all without paying for tokens.

The Danger of Empty Buzzwords

What bothers me most about Nadella's comments is that they're meaninglessly vague. "Collapse it all" isn't a strategy, it's Buzzword Bingo designed to generate headlines and hype without saying anything concrete or actionable. "Hey potential customers, just paint whatever you think AI should be on this canvas, and don’t forget to give us money.”

If Mr. Krabs was a real person, he would be shilling AI Agents.

Because “We’re going to make your existing products incrementally better by leveraging the productivity gains that are possible by responsibly integrating AI” doesn’t make for an exciting keynote.

My Prediction

We’re going to see a cycle similar to Blockchain, except that LLMs are more broadly useful. We’ve been in the hype cycle, and we’re starting to see some of the trough of disillusionment. The level of investment being made, given the incremental real-world progress and the lack of evidence of real GDP impact or quality improvements, cannot continue.

The bubble will deflate, though it’s uncertain whether it will burst or bleed slowly. Most of the startups will die. The investors will lose their money, or trick new investors into taking the fall, Ponzi-style.

After that, these tools will be used selectively for what makes sense. They’ll slip into the background, and the real winners will be the companies that create applications that perform their function well without slapping AI all over the marketing brochure.