What it shapes
What the model knows and how it reasons
How that knowledge is expressed in your product
Speed to change
Slow : weeks to months
Fast : hours to days
Cost to change
High
Low
Who owns it
ML / data team
PM, prompt engineers, product team
Examples of decisions
Data selection, fine-tuning, RLHF, task optimisation
System prompts, guardrails, retrieval, temperature, output filters
Right question to ask
"Does the model fundamentally not know how to do this?"
"Is the model capable but behaving wrong in this context?"