Start deterministic is the advice I would give too, and the reason is debuggability rather than accuracy. A deterministic router that is wrong 10% of the time is wrong in ways you can enumerate and fix; an LLM router wrong 5% of the time is wrong differently each run, so you never accumulate knowledge about the failure.
From the document-AI side there is a free benefit too: a rule-based router produces labelled data as a byproduct, so when you do want a learned router later you have a training set drawn from real traffic rather than guesses.
Worth stating the reverse condition as well. The honest trigger for going non-deterministic is when the rule set stops being enumerable, not when someone wants it to feel smarter.
Start deterministic is the advice I would give too, and the reason is debuggability rather than accuracy. A deterministic router that is wrong 10% of the time is wrong in ways you can enumerate and fix; an LLM router wrong 5% of the time is wrong differently each run, so you never accumulate knowledge about the failure.
From the document-AI side there is a free benefit too: a rule-based router produces labelled data as a byproduct, so when you do want a learned router later you have a training set drawn from real traffic rather than guesses.
Worth stating the reverse condition as well. The honest trigger for going non-deterministic is when the rule set stops being enumerable, not when someone wants it to feel smarter.