Accounts payable: what software and AI can take over, and what still needs a person

Software and AI can already capture invoice data, match it to purchase orders and receipts, schedule payment runs and draft accruals. People still need to approve exceptions, settle disputes, own tax judgements and release money. None of it works until the vendor master and the order and receipt records are clean.

Work that software handles well now

Invoice capture is the most mature area. Optical character recognition paired with a trained model reads supplier invoices, pulls header and line detail, and keys it into the AP system. Accuracy improves sharply when suppliers send structured electronic invoices instead of scanned paper.

Matching follows naturally. Once an invoice is in the system, rules compare it against the purchase order and the goods receipt. Lines that agree within tolerance post without anyone touching them. The same engine can check the pay file against the vendor master before release, flagging a changed bank account or a supplier marked inactive.

Payment runs are largely mechanical. The system selects due items, applies early payment terms, groups by method and produces the bank file. Recurring accruals and their reversals at period start can also be generated automatically, provided the rules behind them are written down and stable. Year end payee reporting, such as information returns to the tax authority, is another good candidate once supplier tax status is held correctly.

Routine supplier questions are newer ground. A portal or chat tool that tells a supplier whether an invoice has been paid, using live system data, removes much of the inbound call volume. It cannot negotiate, and it should not try.

Where a person still has to decide

Approval of payments stays with people. The approver carries accountability for spending, and a model cannot hold delegated authority. Automation can route the request and show the evidence. Someone with authority signs.

Exceptions need judgement. A price variance might be a supplier error or an agreed increase nobody recorded. Working out which means talking to procurement and sometimes to the supplier. AI can suggest a likely cause from history; a human confirms it and decides whether to pay, short pay or hold.

Tax treatment on payables is often harder than it looks. Withholding and reverse charge depend on facts about the supplier and the service that may not appear on the invoice. A system applies the codes it is given. Choosing them belongs to someone who understands the rules in each jurisdiction.

Manual adjustments to the ledger, including non routine accruals at close, should be prepared and approved by people with segregated duties. Software can draft the journal and attach support. Posting it without review weakens the control environment auditors rely on.

What the data has to look like first

The vendor master decides whether automation helps or hurts. Duplicate suppliers, stale bank details and missing tax identifiers will be copied straight into payments at machine speed. Clean it, assign an owner, and lock down who can change bank fields.

Purchase orders must reflect what was actually agreed. If buyers raise orders after the invoice arrives, matching becomes a formality and catches nothing. Receipts matter just as much. Services with no receiving step leave the match weaker than it appears.

Coding needs a settled chart of accounts and cost centre structure. A model trained on years of inconsistent coding will learn the inconsistency. Retention rules should be defined before documents move into a new repository, so records are kept for the required period and can be found during an audit.

Tolerances, approval limits and accrual rules must be explicit. Many teams discover these live only in the heads of experienced clerks.

Questions to ask the people who run it

• Which invoices get opened and fixed by hand each week, and why?
• When the purchase order and invoice disagree, who actually gets the call?
• Are some suppliers always paid early, late or outside the run? Who decided that?
• How do bank detail changes arrive, and how is each one confirmed as genuine?
• Which accruals are calculated in a spreadsheet instead of the system?
• What do approvers check before approving, and what gets skipped when they are busy?
• Which supplier questions take longest to answer, and where does the answer come from?
• At month end, which adjustments get posted without a written request?

The answers usually reveal informal workarounds that the documented process never mentions. Those workarounds are where automation either earns its keep or quietly breaks something.

Where changes tend to go wrong

Teams often automate capture first and celebrate the speed, then find exception queues growing because master data was never fixed. Others remove a manual check without noticing it was the only control over duplicate payments. Before switching anything off, trace what the step protects against and confirm something downstream still covers it.

Keep a person reviewing a sample of automatically posted invoices after go live. Shifts in supplier formats or a reorganised cost centre structure will degrade results gradually, and only a human looking at real transactions will spot it early.

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.