Blog

You built an app with AI. But can you actually take it live?

You describe an app to your AI. Minutes later, you have a working app. Login, database, dashboard. Everything seems ready.

But can you actually put it live?

Because "it works" and "it's ready for production" are two very different things. (Ask anyone in IT. They'll tell you, at length.)

Your app works for one person

Right now your app has one real user: you. It runs on test data, in a tool's account, and nobody is trying to break it.

Production is a different place. Real colleagues sign in. Some of them shouldn't see everything. The data is real, and some of it is personal. And the app has to keep working after you change it next week.

Your AI built it to work. Nobody asked it to build something the company can trust.

What's missing between "works" and "live"

Here are some of the things that usually aren't there yet:

  1. Real sign-in. Company accounts, sessions that end, and access that stops the day someone leaves. Not a login page anyone with the link can get past.
  2. Rules for who sees what. Sales see their own deals. Finance sees the totals. Nobody sees salaries by accident.
  3. Data your company controls. Not in a tool's sandbox, or an account tied to your personal email.
  4. Secrets kept secret. API keys belong in a vault, not in the code.
  5. A safe way to change it. A copy to try changes on, someone who checks them, and a way back when something breaks.
  6. Someone who answers for it. What happens when you're on holiday, or move on?

Every one of those is a question IT will ask. And every one is a place where AI-written code often gets it wrong. In October 2026, Veracode tested 18 of the latest AI models on coding tasks where security matters. Nearly 4 in 10 tasks still produced code with a security flaw. Even the best models failed more than 1 in 4. (And that's Veracode's best result yet.)

Security in and of itself is not just a checkbox. Yes, there are certain things you need to check off to make sure an app is secure. But you also need to understand how to put them together. How to make each part work securely, alone and together, while achieving the functionality you require and the performance metrics you need.

It's easy to make an app secure by locking everything down. It's easy to make it fast and useful by leaving everything open. Getting all three at once is design work. A checklist can't do that for you, and neither can a prompt.

Why isn't it ready to go live?

This is the part that catches people out. The problem isn't one bad line you can find and fix.

It's that your AI wrote everything. The app, yes, but also the sign-in, the access rules and the database setup. All of it is new code, and all of it needs checking. Then next week you ask for a small change, and the AI quietly rewrites some of it again.

So IT is left with three options:

  1. Check everything you build, every time you change it
  2. Hand it to developers to rebuild
  3. Say no.

Most say no. (Honestly, fair enough.)

"Can't I just tell the AI to make it secure?"

You can. Lots of people do. "Make it safe and secure." Some go further and build a whole team of agents: one writes the code, one handles DevOps, one is the QA tester and one is the pen tester.

It feels thorough. It still isn't secure, and here's why.

"But my agents are isolated. I use different models." That's the usual answer. The DevOps agent, the QA agent and the pen tester each have their own role and their own context, and each runs on a different model, so they can't all make the same mistake. That's the theory.

In practice, they still share blind spots. Different models learned from much the same public code. In a study from October 2026, eight different AI models reviewed 265 real code changes that had known security flaws. The best caught 78%. All eight together still missed 23, far more than you'd expect if their mistakes were unrelated.

And a role is only a label. Calling an agent a pen tester doesn't make it one. You wrote the instructions for all of them, and each one is still an AI reading code and giving its best guess. Five guesses that agree are not a check.

"Make it secure" is a wish, not a check. The AI will try. Then it tells you it's done, in the same confident voice whether it is or not. Researchers from Carnegie Mellon and Columbia tested this in 2026: one leading AI coding agent got 57% of 186 real features working, but only 12% were secure. Adding security hints to the request didn't fix it.

The ground keeps moving. Your agents signed off last week's version. Then you ask for a small change, and the AI quietly rewrites some of it again. All of that is new code, and all of it needs checking again, by the same agents with the same blind spots.

You can't run a system at a bank and tell them, "Don't worry. I asked my agents to make it secure." Nobody would sign that off, and they'd be right.

It's the same reason the person who wrote the accounts doesn't get to audit them. AI can't check itself. Something outside it has to do the checking: something that gives the same answer every time, that the AI can't talk its way past, and a person who decides what goes live.

Three ways to get it to production

1. Hand it to developers. They'll rebuild it properly. It works, but you join the queue, and that can take months.

2. Harden it yourself. Work through the list above, one item at a time, and have the result checked by something other than the AI that wrote it. It's possible, but now you're maintaining a security system as well as an app, and it all needs checking again after every change.

3. Build on a foundation that already has those parts. This is why we built Langrove. Sign-in, access rules, the database, secrets and hosting are already built and running, in your company's own AWS account. Your AI writes only the app. Langrove, which isn't an AI, enforces the rest, the same way every time, however the app changes. The parts were designed to work together. Access rules, for example, are applied inside every database query, so every screen, export and report follows the same rules. And nothing goes live until a person reviews the changes and puts them live.

Before you go live

Whichever route you take, make sure you can answer these questions before your app goes to production:

  • Who can sign in, and what happens when they leave?
  • Who can see which data?
  • Where does the data live, and who else can reach it?
  • How are changes checked before they go live, and how do you undo one?
  • What happens if you, or the tool, go away?

So, which app of yours is still waiting to go live? Book a demo and we'll show you how it would run on Langrove, or try it now.

← All posts