What's the difference between using scripts/automation versus putting data into an LLM, and why does it matter for client data sensitivity?

Most of the powerful automation work being built doesn't actually require putting sensitive client data into an LLM query at all. A lot of what's useful in accounting can be done with pure scripting (code that performs a defined task) rather than sending data to a language model. Example given: a deferred revenue month-end reconciliation. Instead of manually tracking each deal, the speaker sets up 12 journal entries upfront when the initial invoice is booked, so reconciliation each month is just checking that each customer's deferred revenue balance nets to zero. This was automated with a script that connects to the QuickBooks API, pulls the GL account, groups by customer, flags any non-zero balances, and sends the results via Slack. No LLM is involved anywhere in this process — even if the underlying invoice contained sensitive information like a Social Security number, it's never sent to Anthropic, OpenAI, or any AI provider, because the script just performs deterministic operations on the data.