Free AI-generated code might be your most expensive IT decision

The AI hype cycle promised a revolution. Tools would replace developers. Systems would build themselves. Your backlog would disappear overnight. But that's not what happened.

More and more experts are pointing to the same gap: AI enthusiasm rarely matches what organizations actually see in production. AI has real potential. But only when you apply it where it creates genuine business value, and only when you take the risks seriously.

The hype hit hard

The AI boom started in 2022 when ChatGPT went mainstream. Expectations shot up fast. WEBCON research shows 76% of Polish managers believe their teams could work more effectively with AI tools integrated into company systems. That number tells you how much appetite there is for this technology.

At the same time, the developer shortage was already a serious problem. Demand for business applications kept growing. The developers were scarce. AI looked like the answer.

A few years later, the data tells a different story. Gartner estimates that by 2027, up to 50% of organizations that attributed headcount reductions to AI will end up rehiring for the same types of roles. CAST Software research puts the current global tech debt at roughly 61 billion developer workdays' worth of code to fix.

The numbers behind the problem

The METR report found that software developers completed their tasks 19% slower when they used AI tools. CodeRabbit data shows that while pull requests went up 20% year over year with AI assistance, AI-written code generated nearly twice as many bugs. Logic and correctness problems were 75% more common in AI-submitted pull requests.

Here's the core issue. AI tools optimize locally. They solve the task in front of them. Architectural consistency, long-term consequences, your broader software strategy: none of that is in scope. The result is code that compiles and meets the immediate requirement but doesn't fit anything else you've built.

You end up with duplicated logic, inconsistent design patterns, and a layer of code that works, but nobody fully understands why. In IT circles, people call this the "slop layer." It does its job today. What it does next year is harder to predict.

Vibe coding: great for a weekend project, risky for enterprise

In early 2025, Andrej Karpathy, AI researcher and OpenAI co-founder, described a new way of building software he called "vibe coding". The idea: describe what you want in a prompt, let AI write the code, stay focused on the outcome rather than the implementation. According toY Combinator, a quarter of startups in their current cohort now build products with code that is 95% or more AI-generated.

The approach has obvious appeal. But Karpathy himself flagged the core problem: the code quickly escapes the author's full understanding. When the model can't fix a bug, developers work around it, paste in error messages, try different prompts, until the problem disappears. How? Why? That part stays murky.

At the startup scale, you can live with that. At enterprise scale, it becomes a liability.

The debt you will have to pay

The biggest risk from AI-generated code isn't the bugs you catch today. It's the architectural coherence you quietly lose over months. Inconsistent patterns and hard-to-refactor codebases drive up maintenance costs. What looks like a cost-saving shortcut turns into technical debt your organization pays back for years.

Trying to cut development costs through code automation can produce the opposite result. You build a system that runs, costs more to maintain, is harder to change, and increasingly difficult for anyone to understand.

The answer isn't less AI. It's a better-placed AI

The organizations getting this right aren't retreating from AI. They're being deliberate about where they put it.

AI delivers results when it operates inside governed platforms, connected to real business process data, with clear accountability for what it does and how its performance gets measured. AI that works within a structured workflow framework, where every action is tracked and tied to real outcomes, is a fundamentally different proposition from free-floating code generation.

That's the difference between AI that creates risk and AI that creates value: not the model; the architecture around it.

Want to see how AI works when it's embedded where work actually happens? Explore WEBCON AI Enablement and see what governed, process-embedded AI looks like in practice.