The problem
What was broken before AI
Company S’s finance process depended heavily on manual receipt review, account-code verification, prohibited-item checks, and exception handling. The paper says the team handled roughly 1,448 to 1,450 cases per month and spent more than 24 hours per month on this receipt-review work. Conventional RPA could handle routine steps but struggled with unstructured documents, new item names, ambiguous classifications, OCR errors, and judgment-heavy exceptions.
What changed
What the use case made possible
The workflow connected OCR/IDP extraction, a policy-based whitelist/blacklist database, LLM-supported exception handling, a finance-officer review screen, and a feedback loop that saved approved human judgments back into the automation system. That meant a new item could be reviewed once, added with synonyms or similar terms, and processed more automatically the next time it appeared.
Why this matters
Why this use case is worth studying
The useful lesson is the shape of the control system. Finance did not hand policy compliance entirely to a model; it used AI to narrow the work, explain a recommendation, and make the human decision reusable. For teams automating regulated or audit-sensitive workflows, that is often the safer architecture: extract structured data, check deterministic policy rules first, ask AI only where ambiguity remains, and preserve the human decision as training data or policy memory.
Use this when
When this pattern applies
Use this when a back-office workflow has structured rules, messy documents, repeated exceptions, and a review team whose decisions can be captured for future automation.
