AI Agent Payment Reconciliation: Data, Matching and Exceptions

Define the records you are reconciling

Reconciliation needs more than a list of transactions. An invoice says what a customer owes. A bank receipt says money arrived. An allocation records how much of that receipt belongs to an invoice. If the agent merges those records into one status, the reviewer cannot tell whether a remaining balance is a timing issue, a partial payment or an incorrect match.

Our fictional dataset has six invoices totaling $5,585. Five unique bank receipts total $4,440. Of those receipts, $4,140 is allocated to invoices and $300 remains unapplied. The remaining invoice balance is $1,445. These are synthetic teaching records, not measured results from an agent or a customer.

The table makes the two ledgers visible. Acme and Harbor are fully allocated. Northstar has paid $800 against an $875 invoice, leaving $75 open. Atlas and Cedar have no allocation. The bank feed also repeats receipt B201; the repeated feed row is not another $1,200 receipt.

Check two separate balance equations

The invoice equation is $5,585 minus $4,140 equals $1,445 open. The receipt equation is $4,440 minus $4,140 equals $300 unapplied. Both equations use the same allocated amount, but their totals differ because invoices and bank receipts are different populations.

Do not subtract the $300 unapplied receipt from an arbitrary open invoice just to make the report look finished. The money has arrived, but its customer and invoice allocation need evidence. Nor should the $1,445 open amount be called missing cash: it includes invoices that may not yet have been paid.

The balance display uses those separate denominators. The invoice bars describe invoice value; the receipt bars describe received money. A percentage based on one cannot explain the other. An agent should export the underlying values with its summary so a reviewer can recompute the totals without trusting the generated narrative.

Ask the agent for a proposal with evidence

For Northstar's $800 receipt, the agent should show the invoice reference, currency, customer evidence and proposed allocation. The expected result is a partial allocation and a $75 remaining balance. A proposal to close the invoice requires additional evidence, such as an authorized credit or documented adjustment; it cannot be justified by saying the amounts are close.

Beam describes payment-matching workflows, and Sema4 describes extracting remittance information and matching it to invoices. Those public descriptions help identify relevant vendors. They do not establish how accurately either product would process this dataset or your own records. [1] [2]

When evaluating a product, ask to inspect one ambiguous proposal from source to final record. The reviewer should see which identifiers supported the match, which fields disagreed and whether a person approved the allocation. A polished dashboard that hides the candidate evidence makes that evaluation difficult.

Keep the proposed allocation separate from the confirmed accounting action. An agent may find a likely match while a reviewer decides whether the evidence is sufficient. If the destination system records an allocation, retain its reference and the version of the proposal that led to it.

Handle fees, timing and split receipts explicitly

Suppose a processor settlement is less than the gross customer payments because a documented fee was deducted. That is a different situation from a customer paying less than the invoice amount. The agent needs the settlement detail and the fee evidence before it can explain the difference. It should not invent a fee whenever a receipt is short.

Timing also changes the interpretation. An invoice dated at month end and a bank receipt recorded the next day may belong to different reporting periods. The reviewer needs the observation date, accounting date and the organization's cutoff policy. The model cannot decide the correct period solely from the order in which files were uploaded.

A consolidated receipt can cover several invoices, and an invoice can receive several payments. Represent these as allocation rows rather than forcing a one-to-one match. Check that the total allocated from each receipt never exceeds the available amount, and that allocations plus the remaining invoice balance reproduce the invoice total.

Currency differences require their own evidence. Preserve the original currency, received currency and any documented conversion or charge. If the necessary rate or settlement breakdown is absent, leave a reasoned exception. A similar numeric amount in another currency is not a reliable match.

Give unresolved money an owner

The queue below separates three issues: Northstar's $75 residual, the $300 unapplied receipt and a duplicate bank-feed row. Each needs a different response. Asking a single reviewer to “clear exceptions” without identifying the decision encourages rushed allocations that hide the problem.

For the residual, the receivables owner checks whether the customer intended a partial payment or claims a deduction. For unapplied cash, the cash-application owner investigates the remittance and payer evidence. For the duplicate row, the integration owner checks the feed identity and ensures the record is not counted twice.

Retain resolved exceptions in the evidence history. Removing a row from the active queue should mean that someone recorded a decision, not that the row disappeared. An exported report should let the reviewer reconstruct the reason, owner and resulting allocation or correction.

Measure coverage without concealing mistakes

Separate proposed matches, confirmed matches and unresolved cases in your evaluation. Count incorrect allocations as well as unprocessed records. A tool can increase apparent match coverage by making unsupported decisions, so coverage alone is an incomplete measure of financial correctness.

For a labeled test set, precision means confirmed correct proposed matches divided by all proposed matches in the same evaluated set. Report the denominator and review method. Do not publish a precision percentage if nobody has checked the proposals against expected outcomes. A small teaching dataset illustrates the calculation but does not establish production performance.

Also track unresolved value and age. Ten small exceptions can be less financially significant than one large unapplied receipt, while a low-value item may still reveal a systematic duplicate problem. Keep both record counts and monetary balances available to the team.

Before expanding the pilot, repeat the run with the duplicate row, a missing remittance and a delayed status update. The agent should preserve the same money totals and make uncertainty visible. Use the detailed matching guide for a larger test set and the audit guide to connect each decision to its evidence.