Manage transactions: what to automate, what to keep human, and what the data needs

Software can already capture, match, convert and post most routine cross-border transactions, and AI can draft reconciliations and flag exceptions. People are still needed to approve unusual payments, settle intercompany disputes and sign off eliminations. None of it works until entity codes, counterparty records and exchange rate sources are clean.

Where software carries the load today

Capture is the easiest win. Invoices, receiving reports and payment requests arrive in predictable formats from most suppliers and subsidiaries. Extraction tools read them well, including scanned documents in other languages. The output still needs a validation rule behind it, but the keying work disappears.

Matching follows naturally. A system can compare an invoice against its purchase order and goods receipt, then hold anything that fails tolerance. The same logic applies to expense claims checked against an approved trip or budget line. When the reference data is sound, the match rate is high and the queue of held items becomes the real work.

Foreign exchange conversion belongs entirely to the machine. Rates should be pulled from one agreed source at one agreed cut-off, applied by rule, and stored with the transaction. Manual rate entry is where most translation differences begin.

Posting to the general ledger is mechanical once subledgers are mapped correctly. Recurring journals, accruals that reverse, and depreciation runs on fixed assets can all be scheduled. Payment runs can be built automatically too: items not yet due sit in a holding file, and the system calculates due dates, early payment discounts and late interest from contract terms.

AI adds value at the reconciliation stage. It can propose matches between two sides of an intercompany balance, group unmatched items by likely cause, and write a first draft of the explanation. Treat those drafts as suggestions. They speed up a reviewer; they do not replace one.

Where a person still has to decide

Approval of anything outside policy stays human. That covers payments to new or changed bank accounts, advances, and invoices that only partly match. Fraud tends to hide in exactly these cases.

Intercompany disputes need someone with authority in both entities. When one subsidiary books a charge and the other rejects it, no matching engine can say who is right. A controller has to agree the treatment, and often the transfer pricing team too.

Manual journal vouchers deserve a human preparer and a separate approver. Netting a supplier's receivable against what is owed to them is another judgment call, especially across legal entities or currencies. The decision to accelerate or delay payment for cash reasons also sits with treasury, not with a scheduling rule.

Final sign-off on eliminations and on the reconciliation of ledger to subledger is a control. Auditors will want a named person who reviewed it and can trace a balance back to its source documents.

What has to be true about the data first

Every entity needs a single code used in every system, and every counterparty must map to exactly one of those codes. If a subsidiary appears under three names in the vendor master, intercompany matching will fail quietly.

The chart of accounts must map consistently from local ledgers to the group ledger. Tie-points between subledgers and control accounts should be defined in writing before anyone builds an automated reconciliation.

Purchase orders and receipts have to be recorded before invoices arrive, or the matching engine has nothing to compare against. Where receipts are entered late as a habit, fix the habit first.

One rate table, with clear rules on which rate applies to which transaction type, is non-negotiable. So is a source document attached to every posting, because an AI explanation with no evidence behind it will not survive audit.

Suspense accounts should be near empty and actively owned. Automating into a messy suspense balance only produces mess faster.

Questions to ask the people who run it

What do you do with an invoice that matches on amount but not on quantity? Who actually approves it, and is that the person on the delegation list?

Which exchange rate do you really use when the system rate looks wrong, and where do you get it?

When the other entity disagrees with an intercompany balance, how do you find out, and how long does it sit before someone acts?

Which journals do you post by hand every month that nobody has asked you to stop?

Are there spreadsheets outside the ledger that the close depends on? Who owns them?

What is in suspense right now, and why has each item stayed there?

Which suppliers or subsidiaries cause the most rework, and what is different about how they send documents?

If this step disappeared tomorrow, what would break first?

APQC's Process Classification Framework® (PCF) is an open standard developed by APQC, a nonprofit that promotes benchmarking and best practices worldwide. To download the full PCF or to view definitions and measures, please visit www.apqc.org/pcf.