Capstone Two: An Extraction Into Fields

In practice

First, the data. The emails contain supplier contacts and bank details, so the work stays in the company's governed tool, and bank account numbers are excluded from the extraction entirely. Second, the prompt. Each field is named and defined: supplier, invoice number, invoice date, due date, amount, currency, and purchase order number. The output is one row per invoice, and the instruction says to write not stated wherever a value is missing, and never to fill one in. Third, the run and the checks. The extraction produces sixty-two rows. The analyst checks the numbers pass first: totals by supplier are compared with the suppliers' own statements, and two do not match. Tracing them, she finds one invoice extracted twice, because it was attached to two emails, and one amount where the model read a subtotal as the total. Then she checks the not-stated cells: eleven are genuinely missing purchase order numbers, which go to the suppliers as questions. One not-stated due date is actually in the email, in an unusual format, and she adds it. Fourth, the record. The spreadsheet notes that it was extracted with AI, which checks were done, and the three corrections.