AI lets anyone ship an app in a weekend. Keeping it running is a different job. Why a wave of fragile AI-built apps is about to need rescuing.
Twelve months ago, building an app meant a brief, a budget and a development team. Today it means a Saturday afternoon and a prompt. Marketers are building campaign tools. Agencies are shipping client portals. Founders are launching products they could never have afforded to build a year ago.
I think that is brilliant. I also think a lot of it is going to fall over.
Quantity is up. Quality is not.
Software output has gone through the roof. Features ship weekly that used to ship quarterly. But look at the things that actually matter, like revenue, retention and how happy users are, and they are not climbing at anything like the same rate. We are producing more software, not better software.
That gap is where the trouble sits. An app that looks finished and an app that is finished are different things, and AI is extraordinarily good at the first one.
Why the demo works and launch day does not
AI writes code that works for one person on one laptop. What it rarely writes are the parts nobody asked for:
- Security. Keys left in the browser where anyone can copy them, databases with their access rules switched off.
- Scale. Queries that are fine for ten users and time out at a thousand.
- Foundations. No staging site, no backups, no monitoring, no way to roll back a bad deploy.
- Economics. AI features that cost cents per call, which looks like nothing until a busy customer makes thousands of calls a day.
We were recently brought in to keep an AI-built platform alive through a global product launch. It had been put together quickly by a talented person who was not an engineer, and it worked beautifully at small scale. At launch it was falling over every few hours, and every fix broke something else. Nobody had a map of how the pieces fitted together, because nobody had ever drawn one.
That is not a failure of the person who built it. It is what happens when a prototype is asked to do a product's job.
The refactor wave
My prediction: in the next eighteen months to two years, a lot of businesses are going to face a serious refactor. Not because they used AI, but because they shipped what AI produced without anyone owning the architecture underneath.
The tech debt being created right now is enormous, and most of it is invisible until traffic, a security incident or an investor's due diligence finds it.
The businesses that come out ahead will not be the ones that refused to use AI. Waiting for perfect, hand-written code is a fast way to become irrelevant. They will be the ones that used AI for speed and put engineering judgement around it.
Longevity is the dividing line
The useful question is not "was this built with AI?" It is "how long does this need to last?"
A six-week campaign tool can be held together with tape. It needs to survive the campaign, and then it can be retired. A product you expect to run for five years is different. It needs foundations: a data model someone understands, security that has been checked, a safe way to deploy, and an owner.
Be honest about which one you are building. Most of the pain I see comes from something built like the first and used like the second.
How to build with AI without building a liability
- Let AI write, let people decide. Use it to draft code, scripts and features at speed. Keep a person responsible for scope, architecture and sign-off.
- Review more, not less. An AI-generated change can run to thousands of lines. That deserves more scrutiny than a hand-written one, not a quicker glance.
- Automate the checks. Tests, security scans and quality gates that run before anything goes live. If someone has to remember, it will be forgotten on the busiest day.
- Get a second pair of eyes before launch. An hour with an experienced engineer before real users arrive is cheaper than a week of firefighting after.
- Plan for success. Ask what happens at ten times today's users. If the answer is "I don't know", find out before your launch finds out for you.
Summing up
Vibe coding is not the problem. It is one of the best things to happen to marketers and founders in years. The problem is treating a prototype as a product. Build fast, by all means. Just make sure someone has their hands on the foundations before you invite the world in.
If your AI-built app is already wobbling, this is what we do.