We built an app with AI and it almost works. Now what?
You are in the most common new rescue category, and it is usually saveable. AI-built systems fail in patterns: security open beside one working login, tests that cannot fail, documentation for things that do not exist. A structured read finds them fast.
The longer answer
In practice.
Why it almost works
AI optimises for looking done: screens render, the happy path happens, the
notes glow. The gaps hide in what nobody asked it to prove: the second user, the
bad input, the refund, the login nobody retired. On one live platform I proved
six such gaps by using them, including logging into someone else's account.
The triage
Secure your accounts first, then a structured read of what is actually there
versus claimed, then a short list: what is sound, what needs repair, what needs
replacing, priced. Often more survives than you fear; occasionally the honest
answer is a rebuild that reuses the lessons rather than the code.
The lesson for round two
The tool was never the problem; the missing specification and adversarial
checking were. Keep building with AI, run it properly this time, per
AI development and
rescues.