Belitsoft custom software development company notes that the rise of vibe coding is a natural consequence of a broader trend of AI-assisted software development. AI-driven code generation turned out to be one of the most buzzing areas in the press, caused by million-dollar investments in this tech innovation. Reuters informs that code-gen startup valuations have skyrocketed as companies wish to benefit from AI capabilities in software development.
These tools enable users to interpret plain English commands into code, which the media sometimes calls vibe coding. Non-programmers believe that using these tools, even they can create working code just by describing what they want to get, while seasoned software engineers can dramatically speed up routine workflows.
Nowadays, organizations of all sizes are experimenting with LLM-powered development. Many companies have at least started integrating gen AI in their business processes, and software development is a prime area.
Today, the lion’s share of code is generated by LLMs – a quarter of startups may have 95% of LLM-created code. These modern companies rely heavily on AI to release an MVP quickly. Still, this rush often joins hands with technical debt: observers bring out that some founders embracing AI produce low-quality code, as dirty and cheap as they can, without forward-thinking planning.
Ultimately, such startups need someone who can step in and “rescue” these raw AI-coded projects stuck halfway. That is why the niche for companies with senior AI expertise is growing.
Belitsoft and others report being reached out to by startup teams that tried vibe coding in-house but ended up with flawed systems – lacking robust architecture or security functionality. As a result, they need senior developers to audit, refactor, and deploy the outcomes.
This scenario is becoming typical as the hype runs into reality. Companies understand that skipping the architecture phase or missing expert oversight can lead to fragile software that is not able to scale or align with compliance requirements.
In terms of wider market trends, LLM-generated code quality has been continuously improving, which further drives its adoption. Each new model shows a better interpretation of the coding context and fewer flaws. Upgrading from a legacy AI model may double coding velocity because the up-to-date model can deal with prompts more precisely. Already now, some LLMs are top performers in producing reliable code when guided properly.
Belitsoft conducts an ongoing evaluation of these models to apply the best-performing ones for each task. It can be switching to a larger context model for handling big codebases or using code-specific models for particular languages.
The situation resembles a virtuous cycle: as the algorithms get better, more companies wish to benefit from them on real projects, which in turn fuels more venture funding and innovations in this area. Still, the enterprise sector has been slower to fully adopt vibe coding compared with startups.
Enterprises often have obsolete codebases, tough compliance requirements, and formalized development processes. All these factors can act as “crutches” slowing down AI implementation. Also, these companies struggle with a talent gap: many corporate managers and software engineers are still not adept at creating high-quality prompts or adopting AI agents into their working routine. Prompt engineering and AI supervision are becoming a new skill set that businesses must encourage. However, even conservative enterprises have begun dipping their toes into vibe coding, especially for new, autonomous modules or for modernizing legacy applications.
A notable use case is Morgan Stanley, which designed an internal tool to assist in rewriting its obsolete COBOL code – effectively leveraging an LLM to decrypt and regenerate decades-old code faster than any human could. The outcome demonstrates promise, confirming that LLMs can boost even traditionally sluggish enterprise development workflows like legacy refactoring.
Similarly, Microsoft has deeply adopted AI for its development pipeline. The company reported that up to 30% of new code is AI-generated. The part of the staff was laid off, indicating the company expects current teams to complete more tasks with AI assistance.
Observers project that shortly, even large enterprises will regularly integrate vibe coding in handling new components or greenfield projects, while human staff focus on high-level architecture. This is already caught by in-house software engineers. For example, one FAANG developer noted that while AI isn’t yet excellent at higher-level design decisions or bigger picture stuff, it performs well with boilerplate and pumping out well-encapsulated tasks. That is why the day-to-day engineering process is going to look very different after adoption becomes mainstream. We are seeing only the start of this shift.
In 2025, market surveys of CIOs demonstrate a surge in budget allocation for AI developer tools and training, and early enterprise pilots are extending. However, full enterprise adoption will likely be incremental – many companies are proceeding cautiously, implementing AI coding on non-critical projects or in isolated environments until challenges, such as regulatory compliance, data privacy, and staff training, are overcome.
Business Value of Vibe Coding
Under the oversight of senior engineers, integrating vibe coding workflows cuts development time and cost for software projects.
A single experienced developer benefiting from AI assistance can deliver in hours what would take months of traditional efforts. For instance, it is possible to create a full iOS app (with AI functionality and a database) – a project that would typically take a few months and cost over $20,000 – much faster and at a significantly lower budget using AI tools. It is shrinking MVP timeframes and allowing companies to validate ideas much faster.
AI coding tools are used for repetitive tasks (drafting basic UI forms or a standard user login). Today, these tasks are carried out almost automatically, freeing software engineers to concentrate on big picture stuff and refinement. Scope that might take days can be completed within hours, freeing more time for addressing complex cases or polishing architecture.
As a result, the total cost of ownership for the client dramatically decreases because far fewer billable hours are needed overall. Business reports meaningful cost savings (cutting a project’s effort by about 5 times) while still delivering a well-tested, secure solution on a shortened schedule.
Consequently, clients get faster delivery and higher satisfaction with the end product, because they see excellent outcomes for a much lower investment.
Instead of reducing the workforce, companies are redeploying their top developers to multiply value. Tech leaders predict that demand is shifting towards AI Engineers with expertise in integrating AI into software development (and it will become the most desired engineering position of the decade).
Risk Management with Vibe Coding
Still admitting that LLMs are powerful tools, it is worth noting that they can be useless – or even risky – in unskilled hands. Profitable vibe coding is very much a senior-level expertise. It requires a skilled developer to enforce coding standards, supervise the model carefully, and take full responsibility for outcomes. Imagine that an LLM is a highly efficient junior software engineer: it can create a lot of code fast but with zero real understanding or charge. That is how an LLM works.
The lead developer treats an LLM as if it were their junior colleague. The engineer provides concise instructions and reviews every draft. The large language model generates scaffolding in minutes, which saves the engineer’s own coding time. The less skilled developers adopt a different approach. They paste the model’s outcomes directly into the codebase and move on. They don’t take a break to review architecture or security concerns. Their performance can seem outstanding – one day of work now delivers a full-week scope. The point is that an LLM can create convincing code that is still flawed, and specialists without sufficient skills often miss the difference. When lead developers open their pull requests, they see bugs that take time to understand. Output volume has grown, but so has the lead developer’s oversight load. Progress now is connected with the engineer’s skill to detect and explain subtle defects the LLM has produced.
Security is another space where expertise is critical. LLMs do not understand by default secure coding principles – they often skip important steps, such as authentication checks, input validation, or proper error management, unless explicitly requested.
A junior specialist might not even realize these nuances. Research shows that naive AI prompts (requesting a feature without a security context) cause every tested LLM to generate outputs vulnerable to multiple common threats. For example, if an unskilled specialist just asks “Design me a login form” and copies the AI-generated code, they might deploy weak code that is sensitive to various attacks. On the other hand, a senior developer will either prompt the LLM with detailed security requirements or check its outputs manually. They are familiar with all the gotchas and make sure that nothing critical is hidden.
Multiple senior-led AI projects adopt a manual control of security: in high-risk functionality (authentication, payments, third-party APIs), the code might even be hand-written or significantly refactored by the senior to ensure safety.
The sharp difference in outcomes caused a saying in these circles: vibe coding isn’t bad – bad engineers using vibe coding are. In other words, the rumor that AI-powered coding = trash code arises mainly from examples where companies let loose AI agents without proper training and control. The real burden is insufficient competencies guiding the AI, not the vibe coding itself. When handled by a skilled hand, an AI tool can deliver clean, reliable code, but in unskilled hands, it can follow poor practices (mass-generated defective code at an incredible pace).
One commenter projected we will be swimming in an ocean of poor code written by unskilled developers leaning on AI, while skilled professionals will be able to build better software using the same tools.
Talking about risk management, forward-thinking teams are establishing rules to keep vibe coding in check. These measures involve code review protocols: any LLM-generated code must be tagged and sent for an extra manual oversight. Also, they include rules for what category of tasks is appropriate for AI agents and security scanning of AI contributions.
Some organizations implement tools that automatically check AI-generated code for known vulnerability scenarios. The consensus is that vibe coding should be used with trust, but with validation. AI agents can draft 90% of a product, but a human must validate that last 10%, and often that last contribution is critical, as an undetected AI failure could be a disaster.
When steered in the right direction, AI coding is an incredibly potent aid, when mishandled, it can speed up bug generation or security breaches.
Thus, forward-looking organizations invest in training employees on how to apply AI tools responsibly, pair newcomers with mentors and senior colleagues, and encourage a culture where using an AI doesn’t mean skipping accountability. The developer – not an AI agent – remains the ultimate owner of product quality.
About the Author:
Dmitry Baraishuk is a partner and Chief Innovation Officer at a software development company Belitsoft (a Noventiq company). He has been leading a department specializing in custom software development for 20 years. The department has hundreds of successful projects in AI software development, healthcare and finance IT consulting, application modernization, cloud migration, data analytics implementation, and more for startups and enterprises in the US, UK, and Canada.
Source: Belitsoft Investigates the Soaring Demand for Vibe Coding in 2025
