Vibe Coding: Fun for Prototypes, a Disaster for Enterprises
Imagine ordering software like you order a pizza: “Uh yeah, I’ll take one ecommerce site, extra secure, hold the SQL injections.” That’s the dream of “vibe coding” — just describe what you want in human language, and let AI generate the code. Startup founders love the pitch. Enterprise CTOs, meanwhile, nod politely while drafting backup staffing budgets for the inevitable meltdown.
James Gosling — the father of Java — has already given his verdict. In short: vibe coding “blows up” as soon as projects get complicated. And in enterprise software, as he bluntly put it, “this stuff has to work every single time.” Nobody wants to hear, “Oops, AI hallucinated your payroll system, but hey, you can vibe code a fix!”
Simon Ritter, deputy CTO at Azul Systems, explains why the skeptics are right. Two problems keep coming up: garbage training data and the impossibility of making English stop being English.
Training an AI Chef at Microwave University
AI assistants are trained on publicly available code — basically mountains of GitHub projects and Stack Overflow threads. But GitHub is mostly abandoned experiments, half-baked student projects, and prototypes that never saw daylight. Stack Overflow, meanwhile, rewards the fastest working answer, not the best solution.
Expecting perfect enterprise code out of that data set is like expecting a gourmet meal from a chef who learned cooking by watching late-night ramen TikToks. Sure, you get something edible, but don’t serve it at a Michelin-star restaurant.
English Is Not a Programming Language (for a Reason)
Even with perfect data, plain English is the problem. It’s delightfully ambiguous.
- “The chicken is ready to eat.” Either you’re pulling a roasted bird from the oven or some live poultry is sharpening its knife.
- “Buy two pints of milk, and if they have eggs, get twelve.” Twelve… what now?
Programming languages exist precisely because natural language is mushy. Java doesn’t guess. if (x > 5) never secretly means “sometimes less than five if you’re feeling generous.” English, on the other hand, is basically a prank on compilers.
Where AI Actually Helps
That said, AI in coding is useful — just not in the “replace all software engineers” way. It shines at:
- Autocompletion: Your IDE suggests the logical next step. It’s autocomplete on steroids, not a genius composer.
- Boilerplate: Need a class to query a database? AI generates one faster than you can say JDBC.
- Refactoring: When a legacy codebase looks like it was written by three raccoons in a trench coat, AI can help clean it up.
- Prototyping: For small personal apps — like a sports tracker or side project — vibe coding is fine. If it crashes, nobody dies.
Enterprise Horror Scenarios
Now, picture AI vibe coding in critical domains:
- Finance: The AI decides rounding customer balances downward is “easier.” Congratulations — you’ve built the world’s first accidental digital pickpocket.
- Healthcare: Your hospital system requests penicillin, but the AI reads it as “pepperoni,” and suddenly patient charts come with a free pizza upgrade.
- Aviation: The autopilot’s natural-language prompt was “maintain altitude,” but hey, altitude zero still technically qualifies. Good luck with that landing.
These may sound extreme, but they drive home Gosling’s point: enterprise-level systems require perfect precision, not vibes.
The Reality for Enterprises
Hospitals, banks, and airlines don’t get to “shrug and try again.” Enterprise systems are tested obsessively, reviewed line by line, and designed to last for decades. When the development team from 2025 has long since retired, future engineers still need to read, understand, and extend that code. AI-generated scripts written by misinterpreted English prompts? That’s a maintenance nightmare — a future archaeologist’s curse.
Even worse, if the AI writes the code and the AI also writes the tests, who’s really checking the homework? Spoiler: someone still has to review the whole thing. Which means you need programmers. Which means the “no more coders” narrative collapses faster than an AI-generated recursion loop.
The Real Future
Vibe coding is not the death of programming. At best, it’s “autocomplete with ambition.” This isn’t a bug; it’s history repeating itself. Just as we moved from punch cards to assembly, from C to Java, and from hand-rolled HTTP servers to frameworks, each leap made developers faster — but never made them obsolete.
And so AI joins the lineage: a tool that cuts busywork but still bows to the immutable law of enterprise software — reliability at all costs. Until someone invents an English compiler that understands “chicken” in context, human programmers will remain firmly in the loop.
Think of AI vibe coding as your overeager intern. Great for getting coffee orders, fine for making mock-ups, but for running the electrical grid or the national banking system? You wouldn’t even trust it with the office snack inventory.