Can an AI agent match invoices to purchase orders?

Yes. Matching invoices to purchase orders is one of the clearest finance jobs for an agent. It compares invoice, order and receipt line by line, clears matches within tolerance at depth 3 and routes the rest. It never edits an order or receipt to make a match work. An exception is resolved by a person independent of the order.

What does matching involve?

A two-way match compares the invoice with the purchase order: same supplier, same items, same prices. A three-way match adds the goods receipt, so the business only pays for what it ordered and what actually arrived. For services, the receipt is usually a confirmation from the person who received the service.

The work is easy when lines map one to one. It gets harder when an invoice covers several orders, when a supplier describes items differently from the order, when deliveries are partial, or when freight and surcharges appear that were never on the order. An agent handles these well because it can read descriptions rather than demanding exact codes, and it can explain each pairing it makes.

What conditions decide whether it clears or routes?

The agent clears an invoice for payment at depth 3 only when:

• every invoice line pairs with an order line and a receipt line;
• quantities invoiced do not exceed quantities received;
• prices sit within the tolerance set in policy, for example a small fixed amount per line;
• the supplier on the invoice matches the supplier on the order and the master file;
• the invoice is not a suspected duplicate.

Anything else is an exception. The agent routes it with the differences laid out.

Who does what?

1. Agent, depth 1: reads the invoice and finds candidate orders and receipts.
2. Agent, depth 3: clears full matches within tolerance and marks them ready for payment approval.
3. Agent, depth 2: for exceptions, drafts the resolution options: accept the price difference, request a credit, wait for the remaining receipt, or reject.
4. Buyer or budget holder: decides the exception. This must not be the person who raised the order.
5. AP lead: approves any tolerance change and samples cleared matches.

Separation of duties rests on step 4. If the person who raised an order can also approve an invoice that exceeds it, they can approve payment for something nobody checked. The agent's role is to make that decision visible and evidenced, not to take it.

An illustrative example

Take a hospital's supplies team. An invoice arrives for 500 boxes of gloves at 6.20 a box, plus a fuel surcharge of 45. The purchase order says 500 boxes at 5.90, and the goods receipt shows 450 boxes delivered so far.

The agent pairs the lines and finds three differences: 50 boxes invoiced but not received, a price 0.30 above order, and a surcharge not on the order. The price difference across 500 boxes is above tolerance. It routes the invoice to the procurement officer with a short summary and three options. The officer, who did not raise the order, checks the framework agreement, finds the price rise was not agreed, and asks the supplier for a corrected invoice for 450 boxes at the order price without the surcharge. The agent drafts the request; the officer sends it.

What controls make it safe?

• Orders and receipts are read-only to the agent. It cannot amend either to force a match.
• Tolerances set by people. Changed only by the AP lead, with the change recorded.
• Independent exception owners. The person resolving an exception did not raise the order.
• Sampling. A person reviews a sample of agent-cleared matches every period.
• Every pairing explained. Each match shows which lines paired and why, so an auditor can follow it.

Where do orders and invoices go wrong upstream?

Many matching exceptions start with the order. Orders raised after the invoice arrives, orders with vague descriptions, orders never receipted. Fixing those helps more than tuning tolerances. The rules for approving purchase orders and handling invoices that arrive without a purchase order cover the two most common causes. The wider controls are in AI agents for finance: books you can audit.