AI Can Investigate Invoice Problems. Should It Approve Payments?

On 26 October Oracle announced new AI experiences and agentic applications across finance and other business functions. The announcement describes a product direction and intended outcomes, not a guarantee that a customer's approval process is suitable for autonomous action.

A Canadian wholesaler could use an assistant to gather a purchase order, receipt and invoice when quantities do not match. Instead of asking a controller to search three systems, the assistant can assemble the discrepancy and propose a question for the buyer or supplier.

Retain the original invoice and a link to each purchase order and receipt. Use deterministic checks for arithmetic, duplicate invoice numbers and agreed tolerances where appropriate; use AI to interpret unstructured correspondence or draft an explanation. Treat extracted values as proposals until verified against the document.

The agent should not change bank details or release payment merely because an email requests it. Preserve existing segregation of duties and use an independently verified process for sensitive supplier changes. Escalate mismatches and missing evidence to an accountable person.

Create a representative sample containing partial deliveries, credit notes, duplicate invoices, currency differences and legitimate price amendments. Compare proposed resolutions with reviewed outcomes, including cases where the right answer is to wait for evidence.

Measure exception age, reviewer minutes, incorrect classifications and rework after approval. Also track whether the system disproportionately struggles with particular suppliers or document formats. A successful pilot reduces investigation effort while maintaining the same or better control quality.