Q&A
we receive about 100 vendor invoices per day via email. theyre all in different formats - some are PDFs, some are in the email body, some are forwarded chains. extracting the invoice number, amount, vendor name, and due date manually takes our AP team hours
has anyone used a tinyagent to parse these emails and extract the structured data into a table? the AI understanding part seems perfect for this since the data is there but never in the same format
thinking: email forwarding triggers a workflow > agent reads the email/attachment > extracts fields > writes to invoice table
would love to hear if anyones done something like this and how accurate it is
weve done something very similar for parsing customer order confirmations. the agent approach works well for this because invoice formats vary so much that rule-based extraction would need hundreds of rules
our accuracy is about 92% for invoice number and amount, slightly lower for due dates because those are formatted differently everywhere. we have a human review step for anything the agent flags as low confidence
the key is giving the agent clear instructions about what fields to extract and what format you want them in. we also include a few example invoices in the system prompt
92% accuracy is better than i expected honestly. whats the plan for the 8% that get parsed wrong? manual correction?
yep. the workflow adds a "confidence" column to the table. anything below 80% confidence gets flagged and a team member reviews it. still way less work than manually processing all 100 invoices a day
the confidence score approach is really smart. we use the same pattern for our lead qualification agent. anything the AI isnt sure about gets routed to a human. best of both worlds
Community