AI Can Ship Your App. It Still Won't Architect Your Business.

Meta quietly launched a gaming app called Pocket this month that was almost entirely vibe-coded, built through AI prompts rather than a team of engineers writing every line by hand. When a company with Meta's engineering headcount ships something that way, the argument about whether AI-generated software counts as real software is over. It is real. The question that actually matters now is a different one: what still needs a human making the call.

I do not think vibe coding is a fad and I do not think founders should be scared off it. For a landing page, an internal tool, or a fast prototype to show investors, it is often the right call, faster and cheaper than briefing a developer for something that might get thrown away in a month anyway. Where it runs out of road is the moment a business has real structure behind it: more than one brand, a plan to spin something off, a compliance requirement, a growth path that is not a straight line.

We ran into exactly that structure earlier this year, building the website and internal systems for a South African maritime holding group with six distinct subsidiary brands under one company. An AI prompt for "build us a corporate website" would have produced six pages with six different logos. What the business actually needed was something no AI tool has a way of knowing it needs: the ability, if the group later spun one subsidiary out as its own company, to extract that subsidiary's entire site into a standalone platform in under four hours, not a six month rebuild. We built it using domain driven design, treating each subsidiary as an isolated module with its own logic, content and access controls, so the technical structure matched the business's actual shape and where it was headed, the same principle behind structuring a database schema before AI agents need to query it instead of retrofitting one under pressure later.

The part AI tools do not fix is judgment, not speed. Veracode's own testing across the current generation of coding models found that nearly half of AI-generated code still ships with a known security vulnerability when nobody explicitly asks the model to write securely, and according to their Spring 2026 GenAI code security testing, that number has barely moved in two years despite every new model release claiming otherwise. A vibe-coded prototype with a security hole is a Tuesday problem. A vibe-coded system handling customer data, payments, or POPIA-regulated information with the same hole is a different category of problem, and it is not one a faster model fixes on its own.

If you are a founder deciding how to build your product, the question is not whether to use AI tools. Use them, we do too, for scaffolding and for the parts of a build that are genuinely repetitive. The question is where a human has to own the decision: how the system is structured to survive the business changing shape, what happens to customer data, what breaks if you need to split the product in two later. Those are architecture decisions, and a prompt cannot make them for you because the model does not know your business, only your description of it that day. That is the difference between an embedded technical partner scoping the platform before writing a line of code and a tool generating whatever the prompt implies, and it is exactly the kind of custom software development work we get called in for once a vibe-coded MVP hits its ceiling.

Meta can afford to ship something fast and throw it away if it does not work. Most founders cannot, because the product they vibe-code today is the one their customers, their investors, and eventually their compliance officer will hold them to next year. Build fast. Just make sure someone is still making the calls the AI was never going to make for you.