Why your invoices never match your POs (and sane tolerance rules)

Invoices don’t match POs — especially service POs — because matching rules built for goods are applied to work that doesn’t arrive on a pallet. Hours run over estimates, projects bill in slices, currencies move, systems round differently. The fix isn’t stricter matching; it’s sane tolerance rules plus an acceptance step that fits services — so honest invoices flow and real discrepancies stand out.

The practitioner pain here is vivid and specific: invoices “off by a literal penny” that “don’t get paid”, service POs that never match anything, and the workaround economy — “a custom spreadsheet for every single vendor” — that grows around a matching process everyone has quietly stopped trusting. The cost runs both directions: honest suppliers blocked for pennies become suppliers who pre-charge for the hassle, while the process cried wolf so often that real overbilling sails through on exception fatigue.

Why service POs break goods-style matching

Estimated quantity, actual billing. A PO for “~120 consulting hours” meets an invoice for 126.5. Neither party did anything wrong — the PO was a forecast, the invoice is history. A goods-matcher sees a 5% quantity breach and blocks.

Part-billing against a whole. The £60k project bills monthly by progress; every invoice is a partial match against one PO. Unless the system tracks cumulative billed-vs-ordered, each slice looks like a mismatch — or worse, the total quietly overshoots across ten invoices no one summed. (The duplicate’s favourite hiding place, as we’ve covered.)

The penny class. Two systems round a VAT line differently; FX moved between order and invoice; a unit price with four decimals got truncated. This is the “off by a literal penny” blocker — variance with zero information content, blocking payment and burning supplier goodwill.

The missing middle leg. Goods matching has a receiving dock: PO, goods-received note, invoice. Services often have… an email thread. With no acceptance artifact, “did we get what we’re paying for?” has no evidence either way — so AP either blocks (and becomes the bottleneck) or approves (and becomes decorative).

The tolerance rules that work

Write them down; the numbers matter less than their existence:

1. An absolute floor. Auto-approve any variance below £5–£25 (pick per your volumes). This deletes the penny class on day one and costs, by construction, almost nothing.
2. A percentage band on estimated lines. 1–2% on hours, usage and anything the PO admitted was a forecast. The 126.5-hours invoice flows; the 160-hours one stops.
3. Cumulative tracking on part-billed POs. Each slice matches if the running total stays inside the PO plus band. The overshoot gets caught on the invoice that crosses the line, not in a year-end audit.
4. Escalation by size, with context. Small breaches go to the budget owner as a one-click approve with the delta shown; large ones to procurement with the PO, the history and the variance attached. An exception that arrives with its evidence takes a minute; one that requires archaeology takes a week — which is how invoices age.
5. An acceptance artifact for services. Timesheet, milestone sign-off, usage export — agreed at PO time, attached at invoice time. The middle leg of the triangle, sized for services.

Where the agent earns its keep

All of the above is judgment encoded as rules — exactly what should run automatically. Fynex’s AI invoice analysis reads each incoming invoice, finds its PO even through a mangled reference, checks lines against contracted rates and that vendor’s history, applies your tolerance bands (including the cumulative math on part-billed POs), and books the clean majority straight through to scheduled payment. What reaches a human is the honest exception list: the genuine overbilling, the rate that drifted, the running total about to breach — each with the reference it violates attached.

The penny-blocked invoice and the unexamined overcharge are the same failure wearing opposite costumes: matching effort spent where the variance is meaningless and absent where it isn’t. Tolerance rules are how you point the attention at the right one — and once they’re written down, they’re software.