What does it mean to run AI 'locally' or 'offline,' and is anyone using this to annotate financials and map them to Drake tax software?

A distinction was drawn between two different meanings of 'local': (1) working with files stored on your local machine while still using Claude in the cloud (Claude reads/accesses local folders, but data still goes to the cloud), versus (2) truly offline use with local models running entirely on your machine, where no data is sent to the cloud at all. One participant clarified they use Claude to generate scripts, but those scripts then run locally without further AI/cloud involvement once written — effectively a one-time use of Claude to build a program (mentioned in context of SOC 2 security concerns) that installs and runs on a local Windows machine. Their specific project goal is to take financials and annotate each line indicating where it should map in Drake tax software, though at the time of the discussion this was still a base capability that needed to be connected to Drake specifically — not yet fully built out.