Invoice OCR Pipeline.

PAVA acquires aesthetics clinics and relies on consolidated financials to find profit opportunities across the portfolio. But invoices and reports arrived in mismatched formats, so analysts re-keyed data line by line and month-end close slipped. The whole acquisition thesis depends on seeing across practices, and the data was arriving in a shape that made that impossible.

I added a confidence threshold and a review queue instead of trusting the OCR end to end. Finance data that is 95% right and silently wrong is worse than no automation, because nobody knows which 5% to check. Anything under 90% confidence routes to a human, which is what made the output trustworthy enough to close the books on.

• Catalogued all inbound file types and mapped the required fields: practice, vendor, amount, GL code, date.
• Make.com scenario triggered on file arrival in a dedicated Google Drive folder, with an OCR step that extracts text regardless of format.
• Regex and JSON parsing to structure the fields, then append to a master workbook auto-grouped by practice and vendor.
• Filename conventions and a flag-for-review queue for anything under 90% confidence.
• Documented the scenario, trained a finance associate, and set up weekly error-log summaries.

60%

less manual data entry, about 20 hours to under 8 per close

2 days

faster month-end close

$15K

in duplicate charges found in the first quarter