# #37 What the AI Leaves Behind Low-code platforms have AI assistants now. Describe an app in chat and the assistant assembles it in the visual designer, out of the same components, bindings and event logic a hand on the mouse would have used — their own material says so, and it is the right way to build such an assistant: the platform's building blocks are what the platform's runtime can run. So "an AI builds the app" is now true on both sides of [#35](https://abap2ui5.github.io/docs/advanced/insights/35-low-code-or-abap2ui5.md), and as a demo the two worlds look identical — a prompt goes in, an app comes out. The question that separates them has moved one step later: what is left behind when the assistant is done? On a platform, what is left behind is what was always left behind: a designer artifact in the platform's repository, run by the licensed runtime. The AI changed who operates the designer, not what exists afterwards. Every line of the table in #35 still reads the same — the per-user license, the platform's own versioning and testing, the exit path that ends where the contract does. The assistant makes the artifact cheaper to produce, and no cheaper to own. Here, what is left behind is an ABAP class ([#36](https://abap2ui5.github.io/docs/advanced/insights/36-written-for-agents.md)), and that difference does the work. A class lands in a pull request and shows its diff. It goes through the transport that has always decided what reaches production. It stays in your system when any contract ends, because there is no runtime to license under it. The same split answers the governance question the platforms now lead with: who checks what an AI built? A platform assistant answers with the platform — it plans before it writes, asks before it acts, and the person who prompted approves what they see. Code answers with the controls your shop has run for twenty years: review on a readable diff, [the linter](https://abap2ui5.github.io/docs/advanced/linter.md) judging the view without a system, unit tests, the transport route, and an authorization concept that never asked who wrote the code. Governing AI-written apps is not a new layer to buy. It is the old layer, doing its job on one more author. And because the artifact is text, the claim can be measured instead of demonstrated: every portable sample of the official UI5 demo kit has been ported this way — written by agents, judged by the same gates, held green by CI in [samples-controls](https://github.com/abap2UI5/samples-controls). A number like that exists only where the AI's output is something a gate can read. The demo is the same everywhere now: a prompt becomes an app, in under a minute, to applause. The artifact is what you own, review, transport and pay for from that minute on. Judge the tool by what it leaves behind.