Why L4, and Why Now?

We are living in hard times, and I'm not going to write around that.

For once I chose the word "hard" instead of "challenging" or "interesting" because I want to acknowledge that this is hard.

It's hard that the organization is changing at a pace that makes everything feel rushed and uncertain, and that our values and culture are being stressed in ways they never were before.

And if somewhere in the last few months you've started to quietly resent AI, you are certainly not alone.

Deep inside, every one of us (of course myself included) has felt some version of the same feeling: Is my craft, my expertise, my career, my identity as a builder, my company, being devalued by a model that can do in hours what used to take me years to learn? That feeling is strong and visceral.

Even if you are excited about the new possibilities, you are unsure about whether you will enjoy this new world the same way you enjoyed the old one.

Everyone is scared to some degree whether they admit it or not.

Now, if anyone had any doubt that the world was changing, the last few months have made it crystal clear: It's changing fast.

Whether we like it or not, we did not get to choose it, and neither did anyone else in our industry.

And, for that matter, disliking it doesn't change the forecast any more than resenting rain keeps you dry.

So, we need to embrace AI not because we like it (which we may, of course), but because it is the path to staying competitive in this new world.

And being competitive in the coming months may be the difference between thriving and becoming irrelevant.

I do not want to sound too Darwinian, but we all know that in times of deep change, the companies that thrive are not the ones that have the smartest people, but the ones that adapt the fastest.

After a lot of conversations around this topic, both inside and outside Critical, it is obvious that the sentence "we are using AI" means ten different things to ten different people. Well, and "We use AI everywhere" usually means "We don't even know what we are talking about", but this is a topic for another day.

We need a way (the simpler, the better) to normalize language. And for that matter I find that borrowing the concept from the self-driving car industry is very useful.

I first read about it in a Dan Shapiro article, and it has been a useful mental model for me ever since.

The levels form a simple ladder, from L0 to L5, that describes how much of the work is done by AI, and how much is done by humans. The higher the level, the more work is done by AI, and the less by humans. The lower the level, the more work is done by humans, and the less by AI.

If we go back to 2025, most (if not all) of Critical was somewhere between L0 and L2.

Now, this raises the question: Where should we be by the end of 2026?

I've discussed this with a lot of people, and the first answer that I always get is: L3. I understand it, because it is the responsible-sounding answer.

People want to be safe, and L3 seems safe. It is the level where AI writes the code, and humans review every line. Regardless of whether it delivers on speed, the human will control every line of code.

And that was my first guess too: "Let's do it one step at a time. Let's first get to L3, and then we can see if we can go further."

However, we are now on an exponential improvement curve.

We cannot imagine what the world will look like in 12 months, so we need to push ahead of the curve, and not just follow it.

If we want to be bold, L4 is what we should be aiming for (by the end of 2026): AI implements, and humans validate the product, not every line of code. That's the first level where output scales past headcount, where AI runs the end-to-end loops while we set direction and own the outcome. And ultimately, that's the way to stay competitive.

And make no mistake, L5 will follow and probably sooner than we think.

And of course, critical still means critical: standards, security and project constraints can't be softened because of AI. Sometimes it means we cannot afford L4 or even L3.

But aiming at L4 will make us focus on how we can get there with confidence, trust and safety. Reimagining how we can have the same guarantees as before with L4.

And folks, don't make the mistake of dismissing this as far-fetched, science fiction, or impossible to pull off. We are already running several L4 projects and initiatives across the teams. And one thing I can promise you: everyone I know who has deeply experimented with L4 development came out of it transformed.

For our entire history, software was expensive to build. And that one fact quietly drew the boundary around what got built at all.

There's a whole class of systems that never existed, not because nobody wanted them, but because they were too complex or too expensive to attempt.

As AI is collapsing the cost of development, the frontier of what's buildable moves outward and systems that were out of reach two years ago will be on the table next year.

So, we should go find the next frontier of complexity in our clients' domains. The system that doesn't exist yet because, until now, nobody could build it.

And this is also exactly why we need to aim for L4.

We can only build a previously impossible system by scaling development beyond our own headcount.

The model makers will keep winning the race for raw model power.

But not one of them can take AI into a critical payments system, a medical device, a railway network, and make it trusted.

This is the problem we are good at solving.

And building those systems doesn't need less of the depth you've spent a career on. It needs more.

Now, let me circle back to the beginning: we are living in hard times.

But at the core, we are problem solvers who want to solve problems that nobody else can solve.

As the complexity frontier moves, we will be solving new types of problems that we cannot imagine today. It will take some time for those problems to present themselves, and for our clients to understand them. We can certainly help.

In the meantime, we prepare and adapt. Let's do it.