The math that killed productivity

Fred Brooks formulated his famous insight in the book The Mythical Man-Month in 1975. His logic was simple and relentless: each new person on a team creates new communication channels. Three developers mean three mutual connections. Ten developers already mean 45 connections. With twenty people, we're talking about 190 channels. The more people, the more time is spent on synchronization, meetings, and clarifications — instead of actual programming.

On top of that, there's the onboarding time for newcomers. It takes weeks to months for a new team member to get oriented in the codebase. And during that entire time, they drain time from more experienced colleagues who have to mentor them. The result is that ten programmers often don't deliver even twice the output of one.

An AI colleague that doesn't need standups

AI coding assistants are completely rewriting this equation. They function as non-human teammates who don't need morning standups, don't mix up tasks on Slack, and don't take vacations. The coordination overhead that made large teams inefficient simply doesn't apply to an AI programmer.

Developer investor and analyst Jacob Lauritzen summed it up unequivocally on the 20VC podcast: the main bottleneck in development is no longer writing code itself, but code review and system design. AI tools can now generate functional code in seconds — but someone experienced must check it, fit it into the architecture, and decide whether it makes sense in the context of the entire system.

In practice, this means that five senior developers with AI tools cover the work that previously required fifteen to twenty people. And this isn't some science fiction — companies like Sea Limited report that 87% of their developers use agentic AI tools at least once a week.

Cursor: Half a billion in revenue, but still in the red

If OpenAI and Anthropic build the engines, Cursor builds the car that people actually drive. This AI-native integrated development environment (IDE) has become one of the fastest-growing tools in software history — estimated annual revenue reaches 500 million dollars (roughly 11 billion CZK).

Behind these numbers, however, it's important to note that Cursor is currently operating at a loss. Running cutting-edge AI models and the infrastructure that underpins them is extraordinarily expensive. It's the classic story of tech companies: first build the user base, deal with margins later. Whether this strategy succeeds depends mainly on how fast AI model inference prices drop.

An interesting perspective came from an empirical study examining real Git projects. It found that deploying Cursor brought a significant increase in development speed, but this effect wasn't lasting. The initial productivity jump gradually faded as developers adapted to the new workflow. In other words: AI tools help most during initial adoption, not as a permanent performance multiplier.

What this means for companies — and for Czechia

For Czech tech companies, the lesson is clear. Hiring armies of junior developers is no longer a competitive advantage. Instead, it pays to invest in a smaller, more experienced team and equip them with quality AI tools. Cursor costs $20 per month for the Pro version, GitHub Copilot comes in at $10 per month (or $39 for the Business version), Claude Code by Anthropic is billed based on token consumption. For comparison: the average salary of a senior developer in Czechia ranges around 120,000–180,000 CZK per month.

And it's not just about software development. Lauritzen points out that the ability to design systems is becoming more important than the ability to write code. While companies used to look for programmers who could code fast, today they need people who understand architecture, security, and can recognize when AI has generated nonsense.

Czech companies like Ecomail are already integrating AI into their workflows — they connect email marketing with Claude and ChatGPT and report higher productivity. Similarly, the Czech National Bank is building its own AI center for banking sector oversight. The trend is clear: those who don't learn to work with AI will fall behind.

The risk: over-reliance on AI code

But studies also warn against excessive optimism. AI-generated code still needs human oversight — and the temporary nature of productivity gains suggests that organizations need a thoughtful integration strategy. Simply giving developers access to ChatGPT and expecting miracles isn't enough.

This is particularly true for safety-critical applications — banking systems, healthcare software, or infrastructure. In those domains, human oversight is absolutely essential. As the Anthropic case showed, the Claude Mythos model was able to find over 10,000 critical vulnerabilities, but also raised concerns about the security risks of deploying such powerful AI in the first place.

The economic direction, however, is clear. Model prices are dropping (DeepSeek V4-Pro offers performance comparable to GPT-5.5 at 11x lower cost), tools are becoming more sophisticated, and teams that adapt to the new dynamics first will gain a structural advantage. Brooks probably wouldn't be disappointed — he mainly wanted software to be delivered on time.