I was all in on RPA scripts until I ran a side by side with a local LLM last Friday

Been doing RPA for like 3 years now, automating invoice processing with UiPath bots. Took me 8 hours to tweak a script when supplier changed their PDF format last month. A buddy told me to try running a small local LLM to just parse the text directly instead of hunting for selector coordinates. Set up Ollama with a 7B model on my work laptop in 20 minutes. Ran the same 50 invoices through both side by side - the LLM caught every field correctly, no layout dependencies. My RPA missed 4 because of a shifted table column. I spent all weekend rewriting stuff into Python scripts using the model. Anyone else ditch traditional automation tools for this?

2 comments

2 Comments

evanr791mo ago

Caught every field correctly, no layout dependencies" - I gotta push back on that a little. I've been messing with local LLMs for data extraction too, but they screw up just as often as RPA scripts, just differently. Last week I had a model decide a $12.50 shipping charge was actually a discount code because the PDF had a faded asterisk next to it. My old RPA script would've just pulled the exact number from the right coordinate every time. Plus, you're trusting a 7B model to not hallucinate a random invoice number when the text is blurry. I've had models straight up invent a PO number that didn't exist. It's a tradeoff, not a clear win.

dylanward1mo ago

Some guy on HN said LLMs can't count whitespace reliably either...