When AI hype breaks down in real products: coding tools fail at scale

The message from so-called AI leaders is consistent and louder every day. Despite a wave of AI-induced layoffs, they keep urging: build faster, leaner, and with fewer engineers. Who needs humans? Well, reality – as ever – is a different story, a veteran software developer tells Cybernews.

We’ve all read or heard those predictions, haven’t we? Alphabet, Google’s parent company, recently proudly announced that as much as 50% of all code produced in the firm is now generated by AI coding agents.

OpenAI’s Sam Altman and Anthropic’s Dario Amodei – who don’t shake hands nowadays – regularly spout similar fantasies, and the CEO of Mistral, based in Europe where tech leaders are typically a lot more realistic, said this week that AI was developing software at the speed of light.

The problem? First of all, the promised day of AI enlightenment has somehow still failed to arrive.

In mid-January 2025, Meta CEO Mark Zuckerberg said that AI will replace mid-level engineers that same year. We’re approaching March in 2026, and that hasn’t happened.

Eleven months ago, Amodei also mused: “In 12 months, we may be in a world where AI is writing essentially all of the code.” Unless this happens in April, that particular vision is also wrong.

Can AI actually handle real-world use?

And errors are plentiful. Amazon Web Services suffered at least two outages due to errors related to the use of its own AI coding tools, and at Google, even though 90% of the company’s software developers use AI for coding, only 24% trust it “a lot.”

AI agents can also leak your private data. Microsoft had to address a security flaw where its Copilot AI mistakenly accessed and summarised confidential emails from users’ draft and sent folders.

“Most corporations are not ready to properly secure and manage AI at workplace, while both employers and employees are rapidly switching to mushrooming AI solutions in the hope of gaining some productivity,” Dr. Ilia Kolochenko, a cybersecurity expert and CEO of ImmuniWeb, told Cybernews.

All these glitches are no accident, explains Konstantin Klyagin, founder of Redwerk, a software development firm that works with teams pushing back against overconfidence driven by executive-level AI claims.

“AI’s getting smarter. There are tools that let software developers be more efficient, that’s undeniable. They also allow you, as a non-techie, to develop a prototype that you want to show to investors or check against the market,” says Klyagin.

“But if you’re trying to take it further and to develop a solution that is scalable, viable, and can withstand and serve multiple users, it often fails because it lacks proper architecture, proper software design, proper industry knowledge, and sometimes also proper user experience.”

AI-assisted products need to survive growth, and this means rewriting brittle code, restoring test coverage, and rebuilding architectures that can handle real-world use,

Konstantin Klyagin.

Speaking with Cybernews, he also cites Garner research predicting that, through 2026, organizations will abandon 60% of AI projects that lack AI-ready data foundations.

“These failures are not about models. They are about execution discipline. AI-assisted products need to survive growth, and this means rewriting brittle code, restoring test coverage, and rebuilding architectures that can handle real-world use,” said Klyagin.

AI agents are more like apprentices

His company Redwerk is adamant that coding skills definitely aren’t optional, even though the AI hype flagbearers like to say so.

“We work with teams that built fast using AI tools and paid the price later. Early velocity masked deeper issues, core systems lacked structure, rests were missing, security assumptions went unchecked,” a Redwerk spokesperson tells us.

"Nothing failed on day one. Everything failed at scale. The moment real users, partners, or auditors enter the picture, teams realize they need more engineering judgment, not less.”

It all began with a concept called vibe coding, which lets you code without knowing how to code. All you need is an idea and being familiar with a large language model like ChatGPT to get started.

Sounds great, but vibe coding seems to be both a blessing and a curse for many developers who are pressured to improve and finish projects at a breakneck speed. Errors are basically fated to occur.

“I like one definition of vibe coding, attributed to Linus Torvalds. He said that in vibe coding, vibe stands for Very Inefficient But Entertaining,” Klyagin laughs.

He explains that AI is not ready yet to make code production-ready by responsibly and autonomously fixing or rewriting it.

“Right now, a tool like Copilot is more like an apprentice. It can speed up your work. But you still need a professional software engineer, an architect, to build a proper solution – as well as industry knowledge, compliance knowledge, and proper environment setup. So we still have our bread and butter,” Klyagin tells Cybernews.

According to the tech entrepreneur, AI in its current state of development still has many flaws with memory, context window, and simply with common sense. And yet it might improve, so Klyagin isn’t willing to predict AI’s path too boldly.

In vibe coding, vibe stands for Very Inefficient But Entertaining,

Konstantin Klyagin.

Instead, he talks about the different trajectories of Russia-made cars and China-manufactured cars. The Russian ones are pretty awful but the cars made in China, especially the EVs, are quickly becoming the envy of the world’s auto industry.

“The question is whether we’re going down the Russian car path, or the Chinese car path. If we’re going the Chinese car path, in a few years or decades we can indeed have an AI that may be making decisions on how to design and build software,” says Klyagin.

Unlock more exclusive Cybernews content on YouTube.