Magic is inherently unruly.
Magic is when you don't understand how it works.
A magic demo creates a confusing product experience.
Magic is stochastic; sometimes it works great, sometimes it doesn't work at all.
Since you, by definition, lack a mental model of when it works, you can't predict which inputs will work.
For systems that have magic, it's best to have magic containment.
An alternative is to accrete solid layers, each inductively knowable.
If you don't have an inductively knowable system, then users will have "... wait, what?" moments.
LLMs build systems top-down, from a desired end-state.