Draw the data path.
List what enters the system, where it is stored, what processes it, and what leaves. Include logs, backups, search indexes, and external tools.
Separate the questions.
Where a model runs, what a provider stores, and whether data is used for training are different questions. Answer each one. A local model alone does not make the whole application local.
Test the task.
Run representative examples. Check output quality, response time, and the hardware needed. A model that fits in memory still has to do the work well.
Choose with care.
Some workloads belong on controlled local hardware. Others need an agreed cloud service. Define that boundary before the build, and keep it visible as the system changes.