Major paradigm shift?
Yes. It is not just AI hype. Software is wanted because it is useful. Good software (and good marketing and distribution) has made many extremely profitable companies, because software is not just useful, it has been hard to make. What happens when it becomes significantly easier to make? What happens when it becomes 100x easier?
Are SaaS companies in trouble?
My simple answer is, yes, but I think the truth is a bit more nuanced. I think SaaS companies have to rethink their value proposition, and fast.
Many have considered good software to be a moat that could protect a company's position in the market. But the software itself was never really the moat, it was really the fact that it was very hard to make good software, and it was even harder to then deliver it at scale. If it's really hard, it's generally also very risky and expensive to try. Those things together make a good moat. Maybe I'm simplifying a little, but the point is this: Software is getting easy. Delivering at scale may still be hard, but does that still matter if no one wants what you're delivering?
It's like this: If I could get my matter replicator to make any flavor of pizza I want without the limitation or hassle of going through someone else's limited list of available toppings, why would I bother? Just give me more matter-juice for my replicator please, I'll make my pizza at home...
I'm pretty sure you get it, but just in case, the pizza is software. The matter juice is my tokens for my agent. (Or even better than tokens! Heavily subsidized subscriptions to Claude Code and Codex that let me explore the possibilities of well-harnessed frontier models while the large "matter juice" companies wrestle for dominance...score!)
What about other kinds of software companies?
Yes, them too. But it's not just the companies that make or sell software. This changes things for ALL companies. (Okay, almost all. Like, 99.42% of them.) Why? Because the rules have changed. Because the leverage points have changed. Because when trillions of dollars of empire is trembling atop shifting sands, stuff moves fast.
I thought AI was going to make software companies lots of money?
There's tons of money to be made, but software is now already a commodity for those who have figured out how to effectively leverage AI agents for coding.
I thought AI code was slop?
It can be. But at this point there's no doubt at all.
The age of making millions because software is useful, really hard to make, but nearly infinitely scalable, is over.
If you've been on the inside of this shift over the last couple years, actually coding real things, increasingly leveraging AI (even though you have the skill to do it yourself), you know what I mean. If you're not in that group--if your perception of AI-utility is still shaped by the era of hallucination-prone chats with an LLM--you will likely think I'm being a little hyperbolic. (Note: the pseudo em-dashes are mine. I'm a real human who can't be bothered to remember the key combo to type one right now.) Maybe you think I'm just reading trends and thinking too-optimistically, "we'll get useful AI soon." No. That's not it.
This is not a future prediction anymore, useful AI-powered agentic coding for real software development is already here.
Isn't AI an overblown marketing gimmick that nobody really wants?
I get it. I don't want AI junk features in my accounting software either. Most of what I see people tacking on to their SaaS offering is junk. I can see why many are convinced that this AI trend is just a misguided economic bubble full of promises that have not materialized. I agree--to an extent. But here's the thing: Those AI-add-on use cases to keep your SaaS relevant are mostly a dead end. They're the temporary churn of mass missing-the-point.
LLM's are non-deterministic.
I'm an engineer. I have technical depth enough to have caught blatantly-ridiculous errors from LLM's for years now. But my perspective on this goes further back than that. I sat in a university lab in 2000, having learned neural network math, and coded a C program to train an algorithm to recognize letters on blocks and then used that to program a robot arm to sort blocks with previously-unseen shapes and fonts of letters into the right piles.
Recent "AI" hype aside, neural networks and machine learning techniques proved their utility in real applications a long time ago.
But since LLM's are non-deterministic, that means they're unreliable, right?
Well, humans are non-deterministic too. Statistical probability might be non-deterministic, but it's not useless slop. For humans on an assembly line, we make jigs and fit-one-way components so our non-deterministic friends can get it right with six-sigma levels of confidence. How do you leverage AI agents for real business utility today? There's your free hint for today.
Subscribe for more if you're interested.