2 months of feeding invoice scans to the AI reader at my firm, and the pile on my desk is gone
I work at a small accounting shop in Des Moines and back in March we were drowning in paper. Every client receipt and invoice got retyped by hand, about 200 pages a month, and it ate most of my Fridays. We started running scans through an AI reader in early April, mostly as a test on one client. The first pass pulled the vendor, date, and total right on maybe 60 percent of pages, which was not great. After we fixed how the scans were fed in, two months later it is hitting around 95 percent and I actually finish my Friday stack by lunch. Anyone else using one of these readers, and what trick got you past the last few percent of bad scans?
Man, the thing nobody talks about is what those bad pages do AFTER the reader spits them out. We had a 5 percent fail rate too, but the real mess was that failed scans still got logged somewhere and then quietly got paid twice or got missed completely, and finding those took longer than retyping would have. What fixed it for us was making every low confidence page land in a separate "needs eyes" folder instead of the main batch, so a human has to look at it before anything moves. That cut our cleanup time way more than squeezing out the last few percent ever did. Chase the errors, not the score.