Peak-Season Invoice Intake
Northwind's accounts payable team receives about 900 supplier invoices each working day in a shared mailbox. Roughly 60% arrive as text-based PDFs, 25% as scanned images, and 15% as invoice details typed only in the email body. Four clerks currently key about 140 invoices each per day; month-end volume triples for three days. The 24-hour processing SLA is already missed during month-end. A pilot bot extracted scanned PDFs at 92% field confidence and processed text PDFs in about 8 seconds per page. The finance manager wants to know what to automate first and how to keep clerks productive when the bot cannot finish an item.
1. Decide which of the three invoice streams should be unattended, attended, or manual, and justify the choice with the given volumes and SLA.
Read the worked answer
Text PDFs (540/day) are structured and high-confidence: unattended. Scanned PDFs (225/day) also unattended, but only above a confidence threshold; below threshold route to human verification, not silent retry. Email-body invoices (135/day) are unstructured and variable: use attended or manual handling, because a brittle parser for 15% of volume risks more rework than it saves.
Capacity math: clerks can key 560/day, while month-end demand is 2,700/day. Automating 765/day leaves 135 email-body items plus exceptions. At month-end the bot still handles about 2,295 and clerks handle roughly 405, which is within 560 if exceptions are controlled. The bottleneck is exception handling, not extraction, so staff the verification queue during month-end rather than trying to automate the last 15% immediately.
2. Design the exception path for a scanned invoice where the total amount extracts at 84% confidence. Cover queue status, human action, retry policy, and audit evidence.
Read the worked answer
Do not auto-retry OCR at 84%: a second pass on the same image rarely adds information and can produce a different wrong value. Mark the queue item as a business exception or Failed with reason 'LowConfidence_Total:84%', and route it to a verification queue. A clerk compares the extracted total with the image, corrects the value, and approves or rejects. The bot should not resubmit the original item automatically; create a new verified item or update the same item after human approval, depending on audit rules.
Preserve the original PDF, page image, extracted JSON, confidence score, bot user, timestamp, and clerk decision. If the same supplier exceeds a 10% low-confidence rate over 50 invoices, flag the template for tuning rather than accepting manual rework indefinitely.