Our company has been developing web applications for a long time — whether directly or through outsourcing — throughout its entire existence. For every client, we try to find the optimal solution to their problem so that "everyone comes out well". That means a solution that is effective and right for the client, but also for us, so that we can keep providing our services in the future at the same level of quality. Even though our prices don't come anywhere near the price tags of the big corporate players, clients sometimes decide to go with a different solution — either something off the shelf or a custom build from another supplier. We don't hold grudges, we don't badmouth the competition, and we don't hang up the phone. It's the natural cycle of the business world, and every entrepreneur is responsible for their own business and the decisions they make in it. We have clients we've been working with since the company was founded. A few of them have been with us even longer. We also have a customer who left us and came back a year later. No problem — we're fair, and we have no issue entering the same partnership with serious customers twice, or even more times.
AI as the solution to everything
Recently, however, a new phenomenon has emerged thanks to the massive rise of AI and widely available LL models: agentic programming, where you don't need to know how to code or understand what the language model generated — a model which, in reality, also has "no idea" what it created or why. It works from trained data and mathematical probability, operating exclusively with numbers (not words).
Of course, I'm not going to claim this is a bad thing, or that it doesn't concern our company. Quite the opposite. For our company to stay relevant and competitive, it has to modernise and integrate AI into its everyday operations. The result is more efficient work, faster task completion and, ultimately, lower costs for end customers (for example, we don't bill 4 hours for routine maintenance/updates, but only 2).
There is one catch, though. It still needs to be checked. You always have to verify that the LL model isn't hallucinating, that it doesn't get stuck due to temporary overload of the provider's servers, and that the result actually meets all the requirements. We never deploy anything blindly! And to be able to check something and vouch for the result, we have to understand it. Every change therefore goes through a review process, exactly as if we were checking a colleague's work.
And now to the point
Lately, we have repeatedly encountered the opinion — or the idea — that a (potential) client would like to build their own solution entirely on their own and, in the better case, at least have it checked by someone qualified. Sure, no problem: we can assess it professionally, verify whether the solution follows security principles, whether the whole thing won't collapse within six months due to performance or stability issues, and whether it will actually make the business run more efficiently.
The problem arises when the result isn't as great as it seemed at first glance. Naturally, the shortcomings get written up, "prompted" back to the agent, the agent fixes them, and so on in circles until life is beautiful and colourful again.
But that's not how it actually works.
At our company, we have been developing with the help of agents powered by various LL models for quite a while now. Even so, we have to deal with a range of new challenges. First and foremost is the price, which keeps rising and will definitely continue to rise. So if someone built their own custom CRM system in the evenings 3 months ago and its development and operation cost them, say, €20 back then, today it could easily be €200. And what about in a year or two?
Another unpredictable factor is the constantly changing models and the gradual shutdown of older ones. Even if something runs smoothly and reliably on an old model, that doesn't mean it will do the same on a newer one. We know from experience that every model behaves slightly differently, no matter how detailed the agent instructions are.
Another important fact: a newer model != better results. We have observed this in practice specifically with OpenAI's models. The newer models (GPT 5+) demonstrably have "worse Slovak" than, for example, the older GPT 4o model. This is probably caused by models being optimised to cover more efficiently the tasks and topics they are most often used for, while the less "important" ones fade into the background. We can't see into it in that much detail — it's the policy of the company that owns and develops the model. So an ordinary person who has built an intelligent assistant in Slovak for their business has to deal with it on their own and keep adapting and tuning it (and no longer for those €20).
Our daily bread: laws and directives
Another risk factor is geopolitics, or the security of company data freely travelling across the servers of global corporations. When it comes to source code, I personally have no problem with it — unless it's some unique algorithm that would be worth stealing from a competitive standpoint. But let's be realistic: if something was assembled by an LL model, it won't be anything sensitive. There can be no talk of intellectual property here.
What is sensitive is company data. And the EU, along with the personal data protection act and the cybersecurity act, is strict, and the fines for violating them are often outright ruinous for smaller companies.
For this reason, it's important to realise that even if some "vibe-coded" application passes a security audit, you have to expect that in 99.9% of cases it relies on some external open-source software. The more popular it is, the more attractive it becomes for hunting security vulnerabilities (whether by honest researchers through various bug-bounty programs, or by those who sell the vulnerabilities they find — or exploit them themselves). So regular updates, maintenance and security audits are a necessity. The days when we ran 10-year-old unpatched software on a similarly old (and leaky) operating system are gone for good — and have been for at least 10 years. Cybersecurity is a necessity today.
Back to the obsolete developers
Which raises the question: is anyone really going to manage all of this on their own, alongside running their business, an ever-growing pile of bureaucracy and other obligations? Let everyone answer that for themselves...
That's why, when a (potential) client tells me they'd like something custom-built and want to build it themselves, I don't see it as a problem. It's not my business, nor my responsibility for that decision. But I still can't resist offering at least one piece of good advice: think carefully about whether, in the end, it won't cost you far more than if you had it built and maintained by that "expensive" and "obsolete" developer.