Two lists: the ninety that cleared, the six that did not

What is the exact prompt?

Run it monthly, or weekly at higher volume:

• "Gather this month's invoices from two places: bills in Xero, and PDF attachments in the [invoices@] inbox that are not yet in Xero. List those separately as 'received, not entered'. For each invoice, extract vendor, invoice number, PO reference, line items, and total."
• "Match each invoice to its purchase order in the [PO register] sheet. A clean match is: vendor matches, PO number is valid and open, total is within 1% of the PO amount, and no line item's quantity or unit price exceeds the PO's. Everything else is an exception."
• "Output two lists. CLEARED: invoice number, vendor, amount, matched PO (one line each). EXCEPTIONS: the same, plus the exact reason with both numbers shown ('invoice unit price 14.80 vs PO 12.40, line 3'), and classify each as PRICE, QUANTITY, NO-PO, DUPLICATE (same vendor + amount + near date as another invoice), or VENDOR-MISMATCH."
• "Read-only: enter nothing into Xero, approve nothing, pay nothing, email no vendor."

Classified exceptions route themselves

Each class goes to a different person.

PRICE goes to whoever negotiated the contract, QUANTITY to whoever received the goods, NO-PO to the process conversation, DUPLICATE straight to "do not pay this".

Showing both numbers is what makes the list actionable in one sitting. "Price mismatch" sends you back to the PDFs. "14.80 vs 12.40, line 3" is a decision you make from the list, with the source there if you want it.

Why must the agent never post or pay?

Because matching is a control, and the separation of duties auditors expect is exactly this: a companion checks, a person clears, the accounting system pays.

Consequential actions stop at an approval card in any case, and Ask Always in Settings, AI permissions extends the pause to everything, which the security page explains. For financial workflows the cleaner design is a prompt that never requests the action.

The extraction audit, before you trust it

Amounts and line items read from PDF invoices are the fragile step, because layouts vary.

For the first month, re-check every CLEARED invoice over your materiality threshold and every DUPLICATE call by eye. When a vendor misparses consistently, name it in the prompt with the fix: "Acme's invoices show the total including tax at the top; use the 'total excl. tax' line instead."

Then watch the exception rate. A stable 5% means the process works; a rising rate usually means POs are being raised after the purchase. The "received, not entered" list is the quiet bonus, since invoices that arrived and never reached the ledger are how late fees happen. The other direction, money owed to you, is invoice chasing, and the recurring half of the ledger is a subscription audit every finance team runs.

Frequently asked questions