I run a local model on my own machine. It is not connected to the internet. It costs me nothing per query. That changes the economics of automation.
The useful thing is not chat. It is structured outputs. I send a blob of messy data. It returns a clean JSON object. That is a real automation primitive. No API key. No rate limits. No data leaving my network.
The second useful thing is classification. I feed it customer emails. It tags them as complaint, refund request or general query. That used to require a rules engine or a remote service. Now it happens locally in under a second.
The third thing is text transformation. Take a long internal memo and produce a one paragraph summary. Or take a raw transcript and generate bullet points. Local models are fast enough for this if you pick the right size.
What is not useful is pretending it replaces everything. I see people trying to use local LLMs as a general reasoning engine. That fails. Small models hallucinate more. They struggle with complex logic. Keep the task narrow. Keep the output format fixed.
Automation with local LLMs works when you treat the model as a function. Input in, output out. No conversation. No personality. Just predictable data work.
That is where the leverage is.
Paul