LLM driven vs Runtime driven steps in a workflow — where do you draw the line?

I'm going through the course and practicing by designing a workflow for an automated booking system.

My client wants "everything handled by AI" but technically most of the steps are deterministic — input validation, database lookups, business rule checks. These don't need LLM reasoning, just programmatic execution.

Only one step genuinely needs the model — picking the best option based on user preferences and contextual data. Everything else is just sequential runtime operations.

My question is — in a workflow like this, should tool calls always be emitted by the LLM in sequence even for deterministic steps? Or is the right approach to let runtime handle what it can deterministically and only invoke the LLM where genuine judgment is needed?

Feels like routing every step through the LLM just to satisfy "AI powered" is over-engineering and adds unnecessary latency and cost — but wanted to sanity check with people who've built or are practicing similar workflows.

Anyone dealt with this before?

1 comment

LLM driven vs Runtime driven steps in a workflow — where do you draw the line?