Can AI build a whole app for me?

What this video shows

One line goes into an AI app builder: “Build me a booking app for my salon.”

The phone fills from the top. Home, List, Details: three screens draw themselves in a few seconds each, and the thing already looks like an app. Then the builder reaches Login. The card fills halfway, shakes, and drops back to an outline. It tries again. A small loop arrow starts to turn. Pay does the same. Your data fills, but an open padlock appears on it: the data is in, and it is not safe. The top half of the phone is gold and still. The bottom half is grey and stuck.

Then a second builder joins on the right: a person. A thin gold line connects the person and the AI, and now the lines come from both sides. Login fills and holds. Pay fills and holds. The padlock closes and turns gold. The phone brightens, and one word appears above it: Live.

That is the whole answer. Yes, AI builds most of an app, and fast. The last part, the part you cannot see in a demo, is finished by an engineer working with the AI. That last part is the work I do with my agent team when someone brings me a half-built app, or when I build one from the start.

Nothing shakes in red and nothing says “error” in this film. That is on purpose. Getting stuck at login, payments and data is the normal shape of an AI-built app, not a failure, and the fix is a known list.

What you will learn

• What an AI app builder does well: the screens, the buttons, the look, in minutes.
• Where it gets stuck, in three words: login, payments, your data. The builders’ own documents say so.
• Why the same bug comes back: each fix is a guess, and each failed guess goes into the next prompt.
• What finishes the app: an engineer and the AI, together, with the padlock closed before anything goes live.

The numbers behind the picture

Nothing with a digit is on screen. The numbers belong here, by year.

• 2025: Lovable, the biggest AI app builder, reported 100,000 new projects a day and 25 million projects in its first year. Addy Osmani at Google wrote that non-engineers “get 70% of the way there surprisingly quickly”. Veracode found a security flaw in 45% of AI-written code across 80 tasks and more than 100 models. (Lovable, July and December 2025; Osmani, December 2024; Veracode, July 2025)
• 2026: Lovable: over 200,000 new projects a day by March and 60 million projects by August; 80% of its builders are in non-technical roles. Anthropic: more than 80% of the code merged into its own codebase in May 2026 was written by Claude. Veracode: the security pass rate is 56%, “barely changed from 55%”, while the code now runs correctly almost every time. A university audit of 200 live apps built with AI tools found 91% had at least one vulnerability. (TechCrunch, March 2026; Lovable, June and August 2026; Anthropic, 2026; Veracode, July 2026; Deng, Fan and Meng, arXiv, 2026)
• 2027, my guess: more apps built by people who never wrote code, and more of them stuck at the same three places. Gartner expects 40% of new enterprise production software to be made with these tools by 2028. The security number has not moved in a year, so I expect the last part to stay a person’s job. The article Stuck at 80% has the full numbers. (Gartner, May 2025, via CIO Dive; my guess)

Questions people ask

Can AI build a whole app by itself?

It builds most of one, and fast: the screens, the buttons and the look. The parts a demo does not show, a login that works for strangers, a payment that really charges, data only the right person can read, and the app stores, are where it stops. Lovable’s own documents say its payment page does not work inside the preview and that missing data rules are the most common way app data gets exposed.

Why does it get stuck at login, payments and data?

Because those parts live outside the screen, in other systems, with rules the builder cannot see. Each fix is a guess, and each failed guess goes back into the next prompt, so the same bug comes round again. The builders know this: Lovable’s own advice is to use its automatic fix once or twice, then stop and investigate.

Is an app built by AI safe to launch?

Not by default. Veracode’s 2026 report puts the share of AI-written code that passes security tests at 56%, almost the same as the year before. The open padlock in the video is that number. Closing it is the first job of the engineer who finishes the app.

What does “an engineer and the AI finish it” mean in practice?

An audit of a day or two, then the data rules and secrets, then login and payments tested with real keys, then tests and a review, then the live address and the stores. The AI keeps writing the code; the engineer decides, checks and approves. The article “Stuck at 80%” has the full plan.

Further reading

• Lovable raises $330M to power the age of the builder, Lovable, December 2025. The 100,000 projects a day and 25 million projects.
• The 70% problem: Hard truths about AI-assisted coding, Addy Osmani, December 2024. How far non-engineers get, and the fix-breaks-something cycle.
• When AI builds itself, Anthropic, 2026. More than 80% of merged code written by Claude; direction-setting stays a human role.
• 2026 GenAI Code Security Report, Veracode, July 2026. The 56% pass rate.
• Statement on CVE-2025-48757, Matt Palmer, May 2025. The 170 Lovable apps with open data rules.
• Lovable docs, Stripe integration and Supabase integration. The preview and data-rule lines quoted above.
• Understanding the (In)Security of Vibe-Coded Applications, Deng, Fan and Meng, arXiv, 2026. The 200-app audit.
• App Store Review Guidelines (rule 4.2.6), Apple, updated June 2026.