Because it was trained on what was published, and what decides a real operation is rarely published. Next year's model will be better at everything public and will still know nothing about your operation.
Does feeding the model our own documents solve it?
It solves the written part, which is the smaller one. What decides the operation — why the exception was opened, which supplier actually delivers — isn't in any document to feed it.
Why does the model sound confident even when it’s wrong?
Because a plausible answer and a correct one look identical in a generative model. That's why every claim in the layer carries its source: to be checked, not trusted.
What changes once a layer exists?
The answer starts citing where it came from. Error stops being invisible and becomes traceable — which is different from not existing.
With every generation, the model gets better at what any competitor also has access to. What it will never have, unless someone asks, is your Tuesday.